Add dated historical interpolation weights and matching runtime

2409_variants/v7_dup_14_continue_2_step655000/README.md

@@ -0,0 +1,34 @@
+# 2409 historical interpolation variant — step 655000
+
+Historical interpolation checkpoint. Its relationship to any final TANGO or ACInterp paper model is unconfirmed.
+
+The date prefix records source material from 2024-09; it does not reconstruct the training start date. The original run is `v7_dup_14_continue_2`, step `655000`. This historical variant is separate from the existing `finetune_weighted/checkpoint-8000` release.
+
+## Verification status
+
+The downloaded source file passed a full SHA256 check. A strict CPU comparison against the meta-constructed model passed for all 1,974 keys with no missing or unexpected keys and no shape mismatch. The seed-guider convolution has shape `[320, 5, 3, 3]`.
+
+**GPU forward execution, rendered-video evaluation and fresh-Hub-download verification are pending.** The CPU check did not run a forward pass.
+
+| File | Bytes | SHA256 |
+| --- | ---: | --- |
+| `pytorch_model.bin` | 8699157476 | `beae0aadd1fc049da2f776ecef51a356727eccd18fa5cb5481738ae3adfbabe1` |
+
+See the [asset manifest](asset_manifest.json) and [standalone runtime instructions](../../2610_variant_runtime/README.md). The runtime is frozen and identified by its file hashes; its upstream base commit alone does not contain the new variant support.
+
+## Configured inference command
+
+From `2610_variant_runtime/`, after installing the pinned environment, downloading the base assets and creating your aligned input JSON:
+
+```sh
+WANDB_MODE=disabled .venv/bin/python diffusion/inference.py --config diffusion/configs/2409_v7_dup_14_continue_2_step655000.yaml --checkpoint ../2409_variants/v7_dup_14_continue_2_step655000/pytorch_model.bin --json inputs/input.json --output-dir outputs
+```
+
+Configuration SHA256: `100c01244d99bc9bd263a9695e6cc95685b0acdebc17681741d709e3116ba5b4`.  
+Seed: 2024; layout: 3 + 8 + 3 frames; image size: 512 with black padding; DDIM steps: 20; guidance: 3.5.
+
+The standard SD VAE is used. No custom temporal decoder is required by this packaged profile. This is a full custom PyTorch state dictionary, not a standalone Diffusers pipeline. No quality, best-checkpoint or paper-benchmark equivalence claim is made.
+
+## Retention and license
+
+Optimizer, scheduler, mixed-precision scaler and random-number-generator state are excluded, so exact training continuation is not supported. Weights retain the repository's CreativeML OpenRAIL-M terms; upstream components retain their respective licenses. The runtime source's Apache-2.0 and third-party source notices are separate from model licensing. See the repository [model license](../../LICENSE) and runtime [notices](../../2610_variant_runtime/NOTICE).

2409_variants/v7_dup_14_continue_2_step655000/asset_manifest.json

@@ -0,0 +1,155 @@
+{
+  "format_version": 1,
+  "status": "source_sha256_and_cpu_key_shape_verified_gpu_render_pending",
+  "variant": {
+    "variant_id": "v7_dup_14_continue_2_step655000",
+    "relative_directory": "2409_variants/v7_dup_14_continue_2_step655000",
+    "source_material_month": "2024-09",
+    "date_basis": "Earliest confirmed source-material month; not an inferred training-start date.",
+    "original_run": "v7_dup_14_continue_2",
+    "checkpoint_step": 655000,
+    "scope": "Historical interpolation checkpoint. Its relationship to any final TANGO or ACInterp paper model is unconfirmed.",
+    "assets": [
+      {
+        "path": "2409_variants/v7_dup_14_continue_2_step655000/pytorch_model.bin",
+        "filename": "pytorch_model.bin",
+        "bytes": 8699157476,
+        "sha256": "beae0aadd1fc049da2f776ecef51a356727eccd18fa5cb5481738ae3adfbabe1",
+        "format": "PyTorch full model state dictionary",
+        "sha256_evidence": "Full SHA256 of the downloaded source file matches the source-storage SHA256; independent fresh-Hub-download verification is pending."
+      }
+    ],
+    "architecture": {
+      "code_family": "Custom pose-guided diffusion interpolation",
+      "seed_guider_channels_expected_from_source": 5,
+      "seed_guider_tensor_shape_verified": [
+        320,
+        5,
+        3,
+        3
+      ],
+      "context_frames_per_end": 3,
+      "generated_frames": 8,
+      "total_frames": 14,
+      "image_size": 512,
+      "verification_status": "CPU strict key/shape verified; GPU forward pending",
+      "seed_mask_channels": "4 latent + 1 known-frame mask"
+    },
+    "inference": {
+      "profile": "standard_pose",
+      "vae": "stabilityai/sd-vae-ft-mse",
+      "custom_temporal_decoder_required": false,
+      "code_repo": "https://github.com/H-Liu1997/Video-Motion-Graphs",
+      "code_revision": null,
+      "config_path": "2610_variant_runtime/diffusion/configs/2409_v7_dup_14_continue_2_step655000.yaml",
+      "config_sha256": "100c01244d99bc9bd263a9695e6cc95685b0acdebc17681741d709e3116ba5b4",
+      "environment_lock_path": "2610_variant_runtime/requirements.lock.txt",
+      "environment_lock_sha256": "dfdd012e92ece6da107024040eded8ef55e75b9ab51ba0a8e95b19f0f7621f20",
+      "commands": [
+        "WANDB_MODE=disabled .venv/bin/python diffusion/inference.py --config diffusion/configs/2409_v7_dup_14_continue_2_step655000.yaml --checkpoint ../2409_variants/v7_dup_14_continue_2_step655000/pytorch_model.bin --json inputs/input.json --output-dir outputs"
+      ],
+      "status": "frozen command; CPU key/shape passed; GPU execution pending",
+      "code_source_base_commit": "331fcff579d1bc60e54e568b9f524aa812f64c80",
+      "code_source_contains_frozen_uncommitted_changes": true,
+      "runtime_root": "2610_variant_runtime",
+      "runtime_manifest": "2610_runtime_asset_manifest.json",
+      "runtime_bundle_sha256": "2b1b5a1b19f8be6f97a74a640a085b728ee568d8953f8dbc7a3b5054ac5d8e95",
+      "command_working_directory": "2610_variant_runtime",
+      "seed": 2024,
+      "guidance_scale": 3.5,
+      "denoising_steps": 20
+    },
+    "validation": {
+      "source_file_sha256": "verified",
+      "source_sha256": "beae0aadd1fc049da2f776ecef51a356727eccd18fa5cb5481738ae3adfbabe1",
+      "source_bytes_verified": true,
+      "hub_commit": null,
+      "hub_lfs_sha256": "pending",
+      "fresh_download_sha256": "pending",
+      "cpu_key_shape_check": {
+        "status": "passed",
+        "method": "CPU mmap state dict against exact meta-constructed model; shape/key check without tensor copies or forward",
+        "key_count": 1974,
+        "missing_keys": [],
+        "unexpected_keys": [],
+        "seed_guider_weight_shape": [
+          320,
+          5,
+          3,
+          3
+        ],
+        "forward_executed": false
+      },
+      "strict_gpu_load": "pending",
+      "real_input_render": "pending",
+      "input_assets": [],
+      "output_video": null,
+      "hardware": null,
+      "environment": null,
+      "seed": 2024,
+      "metrics": null,
+      "receipt": null
+    },
+    "excluded_training_state": [
+      "optimizer",
+      "scheduler",
+      "mixed-precision scaler",
+      "random-number generator state"
+    ],
+    "training_resume_supported": false,
+    "license": "creativeml-openrail-m",
+    "license_note": "Retain the applicable Stable Diffusion / CreativeML OpenRAIL-M model terms and upstream component terms. Project source-code licensing is separate."
+  },
+  "base_components": [
+    {
+      "path": "weights/image_encoder/pytorch_model.bin",
+      "group": "base",
+      "url": "https://huggingface.co/lambda/sd-image-variations-diffusers/resolve/42bc0ee1726b141d49f519a6ea02ccfbf073db2e/image_encoder/pytorch_model.bin",
+      "bytes": 1215993967,
+      "sha256": "89d2aa29b5fdf64f3ad4f45fb4227ea98bc45156bbae673b85be1af7783dbabb"
+    },
+    {
+      "path": "weights/image_encoder/config.json",
+      "group": "base",
+      "url": "https://huggingface.co/lambda/sd-image-variations-diffusers/resolve/42bc0ee1726b141d49f519a6ea02ccfbf073db2e/image_encoder/config.json",
+      "bytes": 703,
+      "sha256": "ba41139cc9dad159163a53e2aa847356b627f53fef988ad461659bd2d587d886"
+    },
+    {
+      "path": "weights/stable-diffusion-v1-5/unet/diffusion_pytorch_model.bin",
+      "group": "base",
+      "url": "https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/451f4fe16113bff5a5d2269ed5ad43b0592e9a14/unet/diffusion_pytorch_model.bin",
+      "bytes": 3438354725,
+      "sha256": "c7da0e21ba7ea50637bee26e81c220844defdf01aafca02b2c42ecdadb813de4"
+    },
+    {
+      "path": "weights/stable-diffusion-v1-5/unet/config.json",
+      "group": "base",
+      "url": "https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/451f4fe16113bff5a5d2269ed5ad43b0592e9a14/unet/config.json",
+      "bytes": 743,
+      "sha256": "78f474de6bab3d893868f37be97b636ae65c0df3073ed3256ca458ff599b5f96"
+    },
+    {
+      "path": "weights/sd-vae-ft-mse/diffusion_pytorch_model.bin",
+      "group": "base",
+      "url": "https://huggingface.co/stabilityai/sd-vae-ft-mse/resolve/31f26fdeee1355a5c34592e401dd41e45d25a493/diffusion_pytorch_model.bin",
+      "bytes": 334707217,
+      "sha256": "1b4889b6b1d4ce7ae320a02dedaeff1780ad77d415ea0d744b476155c6377ddc"
+    },
+    {
+      "path": "weights/sd-vae-ft-mse/config.json",
+      "group": "base",
+      "url": "https://huggingface.co/stabilityai/sd-vae-ft-mse/resolve/31f26fdeee1355a5c34592e401dd41e45d25a493/config.json",
+      "bytes": 547,
+      "sha256": "92d3dfb746fca211a2c9e019e285f8597412211728dce3c5bcf4eda0f2d62e7e"
+    },
+    {
+      "path": "weights/animatediff/mm_sd_v15_v2.ckpt",
+      "group": "base",
+      "url": "https://huggingface.co/guoyww/animatediff/resolve/fdfe36afa161e51b3e9c24022b0e368d59e7345e/mm_sd_v15_v2.ckpt",
+      "bytes": 1817888431,
+      "sha256": "69ed0f5fef82b110aca51bcab73b21104242bc65d6ab4b8b2a2a94d31cad1bf0"
+    }
+  ],
+  "base_components_note": "Seven pinned standard-VAE base components have been downloaded and SHA256 verified. GPU execution with the new variants is pending."
+}

2409_variants/v7_dup_14_continue_2_step655000/pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:beae0aadd1fc049da2f776ecef51a356727eccd18fa5cb5481738ae3adfbabe1
+size 8699157476

2411_variants/v7_dup_14_s2_12frames_512_step30000/README.md

@@ -0,0 +1,34 @@
+# 2411 historical 12-frame 512 variant — step 30000
+
+Historical checkpoint referenced by the 12-frame 512 inference configuration; no best-checkpoint or paper-final claim.
+
+The date prefix records source material from 2024-11; it does not reconstruct the training start date. The original run is `v7_dup_14_s2_12frames_512`, step `30000`. This historical variant is separate from the existing `finetune_weighted/checkpoint-8000` release.
+
+## Verification status
+
+The downloaded source file passed a full SHA256 check. A strict CPU comparison against the meta-constructed model passed for all 1,974 keys with no missing or unexpected keys and no shape mismatch. The seed-guider convolution has shape `[320, 5, 3, 3]`.
+
+**GPU forward execution, rendered-video evaluation and fresh-Hub-download verification are pending.** The CPU check did not run a forward pass.
+
+| File | Bytes | SHA256 |
+| --- | ---: | --- |
+| `pytorch_model.bin` | 8699157476 | `98bc64d7c931a7e414e9e9b45709420477d8625055043e945ad01d121f2302fb` |
+
+See the [asset manifest](asset_manifest.json) and [standalone runtime instructions](../../2610_variant_runtime/README.md). The runtime is frozen and identified by its file hashes; its upstream base commit alone does not contain the new variant support.
+
+## Configured inference command
+
+From `2610_variant_runtime/`, after installing the pinned environment, downloading the base assets and creating your aligned input JSON:
+
+```sh
+WANDB_MODE=disabled .venv/bin/python diffusion/inference.py --config diffusion/configs/2411_v7_dup_14_s2_12frames_512_step30000.yaml --checkpoint ../2411_variants/v7_dup_14_s2_12frames_512_step30000/pytorch_model.bin --json inputs/input.json --output-dir outputs
+```
+
+Configuration SHA256: `7258d8124521c81d2b2724115f0e55d2ece513f8715a0f8a3d378c6537096ef1`.  
+Seed: 2024; layout: 1 + 12 + 1 frames; image size: 512 with black padding; DDIM steps: 20; guidance: 3.5.
+
+The standard SD VAE is used. No custom temporal decoder is required by this packaged profile. This is a full custom PyTorch state dictionary, not a standalone Diffusers pipeline. No quality, best-checkpoint or paper-benchmark equivalence claim is made.
+
+## Retention and license
+
+Optimizer, scheduler, mixed-precision scaler and random-number-generator state are excluded, so exact training continuation is not supported. Weights retain the repository's CreativeML OpenRAIL-M terms; upstream components retain their respective licenses. The runtime source's Apache-2.0 and third-party source notices are separate from model licensing. See the repository [model license](../../LICENSE) and runtime [notices](../../2610_variant_runtime/NOTICE).

2411_variants/v7_dup_14_s2_12frames_512_step30000/asset_manifest.json

@@ -0,0 +1,155 @@
+{
+  "format_version": 1,
+  "status": "source_sha256_and_cpu_key_shape_verified_gpu_render_pending",
+  "variant": {
+    "variant_id": "v7_dup_14_s2_12frames_512_step30000",
+    "relative_directory": "2411_variants/v7_dup_14_s2_12frames_512_step30000",
+    "source_material_month": "2024-11",
+    "date_basis": "Earliest confirmed source-material month; not an inferred training-start date.",
+    "original_run": "v7_dup_14_s2_12frames_512",
+    "checkpoint_step": 30000,
+    "scope": "Historical checkpoint referenced by the 12-frame 512 inference configuration; no best-checkpoint or paper-final claim.",
+    "assets": [
+      {
+        "path": "2411_variants/v7_dup_14_s2_12frames_512_step30000/pytorch_model.bin",
+        "filename": "pytorch_model.bin",
+        "bytes": 8699157476,
+        "sha256": "98bc64d7c931a7e414e9e9b45709420477d8625055043e945ad01d121f2302fb",
+        "format": "PyTorch full model state dictionary",
+        "sha256_evidence": "Full SHA256 of the downloaded source file matches the source-storage SHA256; independent fresh-Hub-download verification is pending."
+      }
+    ],
+    "architecture": {
+      "code_family": "Custom pose-guided diffusion interpolation",
+      "seed_guider_channels_expected_from_source": 5,
+      "seed_guider_tensor_shape_verified": [
+        320,
+        5,
+        3,
+        3
+      ],
+      "context_frames_per_end": 1,
+      "generated_frames": 12,
+      "total_frames": 14,
+      "image_size": 512,
+      "verification_status": "CPU strict key/shape verified; GPU forward pending",
+      "seed_mask_channels": "4 latent + 1 known-frame mask"
+    },
+    "inference": {
+      "profile": "standard_pose",
+      "vae": "stabilityai/sd-vae-ft-mse",
+      "custom_temporal_decoder_required": false,
+      "code_repo": "https://github.com/H-Liu1997/Video-Motion-Graphs",
+      "code_revision": null,
+      "config_path": "2610_variant_runtime/diffusion/configs/2411_v7_dup_14_s2_12frames_512_step30000.yaml",
+      "config_sha256": "7258d8124521c81d2b2724115f0e55d2ece513f8715a0f8a3d378c6537096ef1",
+      "environment_lock_path": "2610_variant_runtime/requirements.lock.txt",
+      "environment_lock_sha256": "dfdd012e92ece6da107024040eded8ef55e75b9ab51ba0a8e95b19f0f7621f20",
+      "commands": [
+        "WANDB_MODE=disabled .venv/bin/python diffusion/inference.py --config diffusion/configs/2411_v7_dup_14_s2_12frames_512_step30000.yaml --checkpoint ../2411_variants/v7_dup_14_s2_12frames_512_step30000/pytorch_model.bin --json inputs/input.json --output-dir outputs"
+      ],
+      "status": "frozen command; CPU key/shape passed; GPU execution pending",
+      "code_source_base_commit": "331fcff579d1bc60e54e568b9f524aa812f64c80",
+      "code_source_contains_frozen_uncommitted_changes": true,
+      "runtime_root": "2610_variant_runtime",
+      "runtime_manifest": "2610_runtime_asset_manifest.json",
+      "runtime_bundle_sha256": "2b1b5a1b19f8be6f97a74a640a085b728ee568d8953f8dbc7a3b5054ac5d8e95",
+      "command_working_directory": "2610_variant_runtime",
+      "seed": 2024,
+      "guidance_scale": 3.5,
+      "denoising_steps": 20
+    },
+    "validation": {
+      "source_file_sha256": "verified",
+      "source_sha256": "98bc64d7c931a7e414e9e9b45709420477d8625055043e945ad01d121f2302fb",
+      "source_bytes_verified": true,
+      "hub_commit": null,
+      "hub_lfs_sha256": "pending",
+      "fresh_download_sha256": "pending",
+      "cpu_key_shape_check": {
+        "status": "passed",
+        "method": "CPU mmap state dict against exact meta-constructed model; shape/key check without tensor copies or forward",
+        "key_count": 1974,
+        "missing_keys": [],
+        "unexpected_keys": [],
+        "seed_guider_weight_shape": [
+          320,
+          5,
+          3,
+          3
+        ],
+        "forward_executed": false
+      },
+      "strict_gpu_load": "pending",
+      "real_input_render": "pending",
+      "input_assets": [],
+      "output_video": null,
+      "hardware": null,
+      "environment": null,
+      "seed": 2024,
+      "metrics": null,
+      "receipt": null
+    },
+    "excluded_training_state": [
+      "optimizer",
+      "scheduler",
+      "mixed-precision scaler",
+      "random-number generator state"
+    ],
+    "training_resume_supported": false,
+    "license": "creativeml-openrail-m",
+    "license_note": "Retain the applicable Stable Diffusion / CreativeML OpenRAIL-M model terms and upstream component terms. Project source-code licensing is separate."
+  },
+  "base_components": [
+    {
+      "path": "weights/image_encoder/pytorch_model.bin",
+      "group": "base",
+      "url": "https://huggingface.co/lambda/sd-image-variations-diffusers/resolve/42bc0ee1726b141d49f519a6ea02ccfbf073db2e/image_encoder/pytorch_model.bin",
+      "bytes": 1215993967,
+      "sha256": "89d2aa29b5fdf64f3ad4f45fb4227ea98bc45156bbae673b85be1af7783dbabb"
+    },
+    {
+      "path": "weights/image_encoder/config.json",
+      "group": "base",
+      "url": "https://huggingface.co/lambda/sd-image-variations-diffusers/resolve/42bc0ee1726b141d49f519a6ea02ccfbf073db2e/image_encoder/config.json",
+      "bytes": 703,
+      "sha256": "ba41139cc9dad159163a53e2aa847356b627f53fef988ad461659bd2d587d886"
+    },
+    {
+      "path": "weights/stable-diffusion-v1-5/unet/diffusion_pytorch_model.bin",
+      "group": "base",
+      "url": "https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/451f4fe16113bff5a5d2269ed5ad43b0592e9a14/unet/diffusion_pytorch_model.bin",
+      "bytes": 3438354725,
+      "sha256": "c7da0e21ba7ea50637bee26e81c220844defdf01aafca02b2c42ecdadb813de4"
+    },
+    {
+      "path": "weights/stable-diffusion-v1-5/unet/config.json",
+      "group": "base",
+      "url": "https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/451f4fe16113bff5a5d2269ed5ad43b0592e9a14/unet/config.json",
+      "bytes": 743,
+      "sha256": "78f474de6bab3d893868f37be97b636ae65c0df3073ed3256ca458ff599b5f96"
+    },
+    {
+      "path": "weights/sd-vae-ft-mse/diffusion_pytorch_model.bin",
+      "group": "base",
+      "url": "https://huggingface.co/stabilityai/sd-vae-ft-mse/resolve/31f26fdeee1355a5c34592e401dd41e45d25a493/diffusion_pytorch_model.bin",
+      "bytes": 334707217,
+      "sha256": "1b4889b6b1d4ce7ae320a02dedaeff1780ad77d415ea0d744b476155c6377ddc"
+    },
+    {
+      "path": "weights/sd-vae-ft-mse/config.json",
+      "group": "base",
+      "url": "https://huggingface.co/stabilityai/sd-vae-ft-mse/resolve/31f26fdeee1355a5c34592e401dd41e45d25a493/config.json",
+      "bytes": 547,
+      "sha256": "92d3dfb746fca211a2c9e019e285f8597412211728dce3c5bcf4eda0f2d62e7e"
+    },
+    {
+      "path": "weights/animatediff/mm_sd_v15_v2.ckpt",
+      "group": "base",
+      "url": "https://huggingface.co/guoyww/animatediff/resolve/fdfe36afa161e51b3e9c24022b0e368d59e7345e/mm_sd_v15_v2.ckpt",
+      "bytes": 1817888431,
+      "sha256": "69ed0f5fef82b110aca51bcab73b21104242bc65d6ab4b8b2a2a94d31cad1bf0"
+    }
+  ],
+  "base_components_note": "Seven pinned standard-VAE base components have been downloaded and SHA256 verified. GPU execution with the new variants is pending."
+}

2411_variants/v7_dup_14_s2_12frames_512_step30000/pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98bc64d7c931a7e414e9e9b45709420477d8625055043e945ad01d121f2302fb
+size 8699157476

2411_variants/v7_dup_14_s2_8frames_512_step9000/README.md

@@ -0,0 +1,34 @@
+# 2411 historical 8-frame 512 variant — step 9000
+
+Historical checkpoint referenced by the 8-frame 512 inference configuration; no best-checkpoint or paper-final claim.
+
+The date prefix records source material from 2024-11; it does not reconstruct the training start date. The original run is `v7_dup_14_s2_8frames_512`, step `9000`. This historical variant is separate from the existing `finetune_weighted/checkpoint-8000` release.
+
+## Verification status
+
+The downloaded source file passed a full SHA256 check. A strict CPU comparison against the meta-constructed model passed for all 1,974 keys with no missing or unexpected keys and no shape mismatch. The seed-guider convolution has shape `[320, 5, 3, 3]`.
+
+**GPU forward execution, rendered-video evaluation and fresh-Hub-download verification are pending.** The CPU check did not run a forward pass.
+
+| File | Bytes | SHA256 |
+| --- | ---: | --- |
+| `pytorch_model.bin` | 8699157476 | `6a1288711eaf2792f3581b3e175fb79bb8988e5b192a61178de322c587e810b8` |
+
+See the [asset manifest](asset_manifest.json) and [standalone runtime instructions](../../2610_variant_runtime/README.md). The runtime is frozen and identified by its file hashes; its upstream base commit alone does not contain the new variant support.
+
+## Configured inference command
+
+From `2610_variant_runtime/`, after installing the pinned environment, downloading the base assets and creating your aligned input JSON:
+
+```sh
+WANDB_MODE=disabled .venv/bin/python diffusion/inference.py --config diffusion/configs/2411_v7_dup_14_s2_8frames_512_step9000.yaml --checkpoint ../2411_variants/v7_dup_14_s2_8frames_512_step9000/pytorch_model.bin --json inputs/input.json --output-dir outputs
+```
+
+Configuration SHA256: `37cf67ec1a2b9c8020927ca849fae7a8da7b63762d9c9010a59570b0619a7aa3`.  
+Seed: 2024; layout: 3 + 8 + 3 frames; image size: 512 with black padding; DDIM steps: 20; guidance: 1.0.
+
+The standard SD VAE is used. No custom temporal decoder is required by this packaged profile. This is a full custom PyTorch state dictionary, not a standalone Diffusers pipeline. No quality, best-checkpoint or paper-benchmark equivalence claim is made.
+
+## Retention and license
+
+Optimizer, scheduler, mixed-precision scaler and random-number-generator state are excluded, so exact training continuation is not supported. Weights retain the repository's CreativeML OpenRAIL-M terms; upstream components retain their respective licenses. The runtime source's Apache-2.0 and third-party source notices are separate from model licensing. See the repository [model license](../../LICENSE) and runtime [notices](../../2610_variant_runtime/NOTICE).

2411_variants/v7_dup_14_s2_8frames_512_step9000/asset_manifest.json

@@ -0,0 +1,155 @@
+{
+  "format_version": 1,
+  "status": "source_sha256_and_cpu_key_shape_verified_gpu_render_pending",
+  "variant": {
+    "variant_id": "v7_dup_14_s2_8frames_512_step9000",
+    "relative_directory": "2411_variants/v7_dup_14_s2_8frames_512_step9000",
+    "source_material_month": "2024-11",
+    "date_basis": "Earliest confirmed source-material month; not an inferred training-start date.",
+    "original_run": "v7_dup_14_s2_8frames_512",
+    "checkpoint_step": 9000,
+    "scope": "Historical checkpoint referenced by the 8-frame 512 inference configuration; no best-checkpoint or paper-final claim.",
+    "assets": [
+      {
+        "path": "2411_variants/v7_dup_14_s2_8frames_512_step9000/pytorch_model.bin",
+        "filename": "pytorch_model.bin",
+        "bytes": 8699157476,
+        "sha256": "6a1288711eaf2792f3581b3e175fb79bb8988e5b192a61178de322c587e810b8",
+        "format": "PyTorch full model state dictionary",
+        "sha256_evidence": "Full SHA256 of the downloaded source file matches the source-storage SHA256; independent fresh-Hub-download verification is pending."
+      }
+    ],
+    "architecture": {
+      "code_family": "Custom pose-guided diffusion interpolation",
+      "seed_guider_channels_expected_from_source": 5,
+      "seed_guider_tensor_shape_verified": [
+        320,
+        5,
+        3,
+        3
+      ],
+      "context_frames_per_end": 3,
+      "generated_frames": 8,
+      "total_frames": 14,
+      "image_size": 512,
+      "verification_status": "CPU strict key/shape verified; GPU forward pending",
+      "seed_mask_channels": "4 latent + 1 known-frame mask"
+    },
+    "inference": {
+      "profile": "standard_pose",
+      "vae": "stabilityai/sd-vae-ft-mse",
+      "custom_temporal_decoder_required": false,
+      "code_repo": "https://github.com/H-Liu1997/Video-Motion-Graphs",
+      "code_revision": null,
+      "config_path": "2610_variant_runtime/diffusion/configs/2411_v7_dup_14_s2_8frames_512_step9000.yaml",
+      "config_sha256": "37cf67ec1a2b9c8020927ca849fae7a8da7b63762d9c9010a59570b0619a7aa3",
+      "environment_lock_path": "2610_variant_runtime/requirements.lock.txt",
+      "environment_lock_sha256": "dfdd012e92ece6da107024040eded8ef55e75b9ab51ba0a8e95b19f0f7621f20",
+      "commands": [
+        "WANDB_MODE=disabled .venv/bin/python diffusion/inference.py --config diffusion/configs/2411_v7_dup_14_s2_8frames_512_step9000.yaml --checkpoint ../2411_variants/v7_dup_14_s2_8frames_512_step9000/pytorch_model.bin --json inputs/input.json --output-dir outputs"
+      ],
+      "status": "frozen command; CPU key/shape passed; GPU execution pending",
+      "code_source_base_commit": "331fcff579d1bc60e54e568b9f524aa812f64c80",
+      "code_source_contains_frozen_uncommitted_changes": true,
+      "runtime_root": "2610_variant_runtime",
+      "runtime_manifest": "2610_runtime_asset_manifest.json",
+      "runtime_bundle_sha256": "2b1b5a1b19f8be6f97a74a640a085b728ee568d8953f8dbc7a3b5054ac5d8e95",
+      "command_working_directory": "2610_variant_runtime",
+      "seed": 2024,
+      "guidance_scale": 1.0,
+      "denoising_steps": 20
+    },
+    "validation": {
+      "source_file_sha256": "verified",
+      "source_sha256": "6a1288711eaf2792f3581b3e175fb79bb8988e5b192a61178de322c587e810b8",
+      "source_bytes_verified": true,
+      "hub_commit": null,
+      "hub_lfs_sha256": "pending",
+      "fresh_download_sha256": "pending",
+      "cpu_key_shape_check": {
+        "status": "passed",
+        "method": "CPU mmap state dict against exact meta-constructed model; shape/key check without tensor copies or forward",
+        "key_count": 1974,
+        "missing_keys": [],
+        "unexpected_keys": [],
+        "seed_guider_weight_shape": [
+          320,
+          5,
+          3,
+          3
+        ],
+        "forward_executed": false
+      },
+      "strict_gpu_load": "pending",
+      "real_input_render": "pending",
+      "input_assets": [],
+      "output_video": null,
+      "hardware": null,
+      "environment": null,
+      "seed": 2024,
+      "metrics": null,
+      "receipt": null
+    },
+    "excluded_training_state": [
+      "optimizer",
+      "scheduler",
+      "mixed-precision scaler",
+      "random-number generator state"
+    ],
+    "training_resume_supported": false,
+    "license": "creativeml-openrail-m",
+    "license_note": "Retain the applicable Stable Diffusion / CreativeML OpenRAIL-M model terms and upstream component terms. Project source-code licensing is separate."
+  },
+  "base_components": [
+    {
+      "path": "weights/image_encoder/pytorch_model.bin",
+      "group": "base",
+      "url": "https://huggingface.co/lambda/sd-image-variations-diffusers/resolve/42bc0ee1726b141d49f519a6ea02ccfbf073db2e/image_encoder/pytorch_model.bin",
+      "bytes": 1215993967,
+      "sha256": "89d2aa29b5fdf64f3ad4f45fb4227ea98bc45156bbae673b85be1af7783dbabb"
+    },
+    {
+      "path": "weights/image_encoder/config.json",
+      "group": "base",
+      "url": "https://huggingface.co/lambda/sd-image-variations-diffusers/resolve/42bc0ee1726b141d49f519a6ea02ccfbf073db2e/image_encoder/config.json",
+      "bytes": 703,
+      "sha256": "ba41139cc9dad159163a53e2aa847356b627f53fef988ad461659bd2d587d886"
+    },
+    {
+      "path": "weights/stable-diffusion-v1-5/unet/diffusion_pytorch_model.bin",
+      "group": "base",
+      "url": "https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/451f4fe16113bff5a5d2269ed5ad43b0592e9a14/unet/diffusion_pytorch_model.bin",
+      "bytes": 3438354725,
+      "sha256": "c7da0e21ba7ea50637bee26e81c220844defdf01aafca02b2c42ecdadb813de4"
+    },
+    {
+      "path": "weights/stable-diffusion-v1-5/unet/config.json",
+      "group": "base",
+      "url": "https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/451f4fe16113bff5a5d2269ed5ad43b0592e9a14/unet/config.json",
+      "bytes": 743,
+      "sha256": "78f474de6bab3d893868f37be97b636ae65c0df3073ed3256ca458ff599b5f96"
+    },
+    {
+      "path": "weights/sd-vae-ft-mse/diffusion_pytorch_model.bin",
+      "group": "base",
+      "url": "https://huggingface.co/stabilityai/sd-vae-ft-mse/resolve/31f26fdeee1355a5c34592e401dd41e45d25a493/diffusion_pytorch_model.bin",
+      "bytes": 334707217,
+      "sha256": "1b4889b6b1d4ce7ae320a02dedaeff1780ad77d415ea0d744b476155c6377ddc"
+    },
+    {
+      "path": "weights/sd-vae-ft-mse/config.json",
+      "group": "base",
+      "url": "https://huggingface.co/stabilityai/sd-vae-ft-mse/resolve/31f26fdeee1355a5c34592e401dd41e45d25a493/config.json",
+      "bytes": 547,
+      "sha256": "92d3dfb746fca211a2c9e019e285f8597412211728dce3c5bcf4eda0f2d62e7e"
+    },
+    {
+      "path": "weights/animatediff/mm_sd_v15_v2.ckpt",
+      "group": "base",
+      "url": "https://huggingface.co/guoyww/animatediff/resolve/fdfe36afa161e51b3e9c24022b0e368d59e7345e/mm_sd_v15_v2.ckpt",
+      "bytes": 1817888431,
+      "sha256": "69ed0f5fef82b110aca51bcab73b21104242bc65d6ab4b8b2a2a94d31cad1bf0"
+    }
+  ],
+  "base_components_note": "Seven pinned standard-VAE base components have been downloaded and SHA256 verified. GPU execution with the new variants is pending."
+}

2411_variants/v7_dup_14_s2_8frames_512_step9000/pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a1288711eaf2792f3581b3e175fb79bb8988e5b192a61178de322c587e810b8
+size 8699157476

2610_runtime_asset_manifest.json

@@ -0,0 +1,200 @@
+{
+  "format_version": 1,
+  "root": "2610_variant_runtime",
+  "status": "source_frozen_cpu_key_shape_verified_gpu_render_pending",
+  "source_base_commit": "331fcff579d1bc60e54e568b9f524aa812f64c80",
+  "source_contains_frozen_uncommitted_changes": true,
+  "source_freeze_sha256": "b8ab0585b0b827d4b19da49d6064793ed0f796d978707b5be3f749aec763efbf",
+  "source_file_count": 32,
+  "generated_document_and_base_manifest_count": 2,
+  "total_file_count": 34,
+  "total_bytes": 420063,
+  "bundle_sha256": "2b1b5a1b19f8be6f97a74a640a085b728ee568d8953f8dbc7a3b5054ac5d8e95",
+  "bundle_sha256_format": "SHA256 of sorted UTF-8 lines: path + TAB + decimal bytes + TAB + sha256 + LF",
+  "files": [
+    {
+      "path": "2610_variant_runtime/LICENSE",
+      "bytes": 11357,
+      "sha256": "f9e2070c247517b1ddf65f7b11b393484a18da91a958fd18a97bd0f241c3125c"
+    },
+    {
+      "path": "2610_variant_runtime/NOTICE",
+      "bytes": 1517,
+      "sha256": "f77eb47163db686b37493c42071822b69823c3997a2be1f17eb53d6b6879d3a4"
+    },
+    {
+      "path": "2610_variant_runtime/README.md",
+      "bytes": 7049,
+      "sha256": "059a5ded8a4e914fc3d727783efa02c80e61fd92bd8a9240d5d4e99d3e9da4f9"
+    },
+    {
+      "path": "2610_variant_runtime/asset_manifest.json",
+      "bytes": 2484,
+      "sha256": "1d72168a5e9ed5bf3e8dc7878526646b9e8ed7091a0c0834f36a93bab9bdd05b"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/configs/2409_v7_dup_14_continue_2_step655000.yaml",
+      "bytes": 1493,
+      "sha256": "100c01244d99bc9bd263a9695e6cc95685b0acdebc17681741d709e3116ba5b4"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/configs/2411_v7_dup_14_s2_12frames_512_step30000.yaml",
+      "bytes": 1502,
+      "sha256": "7258d8124521c81d2b2724115f0e55d2ece513f8715a0f8a3d378c6537096ef1"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/configs/2411_v7_dup_14_s2_8frames_512_step9000.yaml",
+      "bytes": 1497,
+      "sha256": "37cf67ec1a2b9c8020927ca849fae7a8da7b63762d9c9010a59570b0619a7aa3"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/inference.py",
+      "bytes": 18709,
+      "sha256": "09ab1c04ad2c551a5ba5a941207bb96a27107c60a814815fde83dd12f1d2a122"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/models/attention.py",
+      "bytes": 21014,
+      "sha256": "966a677b946c7461343a70f166fad7f6db71b4b89fdf59805913fc65cb1fa388"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/models/conv_in.py",
+      "bytes": 745,
+      "sha256": "6e5938de720d0b58cf70f70e859fcc6aada7ffad7add7a75b97571820447c901"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/models/model_vfi.py",
+      "bytes": 32321,
+      "sha256": "1d82eb6527b48e8a7f91dc399e18a0b22dd09588b651833abc768a57bd2ba04a"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/models/motion_module.py",
+      "bytes": 13140,
+      "sha256": "b7b31fdc11731a14eb8335f658b0b731c137cdacdb4ac7523046fbcee34bd69b"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/models/mutual_self_attention.py",
+      "bytes": 14561,
+      "sha256": "d062e505eea0bbd925b971e4871ce4c0dfc9a6bb99873f5f394411d6b7c338d2"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/models/pose_guider.py",
+      "bytes": 1674,
+      "sha256": "f10fc38a83dda911a252033a01d6b11d19b6becf6ccb80e8b0fe252563ec7528"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/models/resnet.py",
+      "bytes": 8153,
+      "sha256": "186c7efd14d1b6284807b7ab3561973af90b6ffc8e149b73787034840de93e31"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/models/transformer_2d.py",
+      "bytes": 18933,
+      "sha256": "36fe3665376cc07eb5feeae43a3c50c5c64913c02412e2d536954460f103f2bb"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/models/transformer_3d.py",
+      "bytes": 5897,
+      "sha256": "68cd30d1e85122bc7621ef13d3a9a165dc3e6bcbcb97ef5a9aa2500fb6826b35"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/models/unet_2d_blocks.py",
+      "bytes": 41508,
+      "sha256": "da56135d0d1b6e56fa87e636f83c731fbe809de42430dfecccebf27a42434eb2"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/models/unet_2d_condition.py",
+      "bytes": 63680,
+      "sha256": "7aec0a8de40cd3d8ea38c051aebb480a245212ccdb942d1b1db9a43d5347225f"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/models/unet_3d.py",
+      "bytes": 27957,
+      "sha256": "d3e3948b2b29cd3ecbf6c5c3d5affcd1d8987923c85658a5c187c7583585b38c"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/models/unet_3d_blocks.py",
+      "bytes": 31017,
+      "sha256": "274c5eb419db218a16920cf85aa0e2d2f33f0079a3be55da98fb7c2b2fd2a7da"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/pipelines/__init__.py",
+      "bytes": 1,
+      "sha256": "01ba4719c80b6fe911b091a7c05124b64eeece964e09c058ef8f9805daca546b"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/pipelines/pipeline_vfi2video_share_vae.py",
+      "bytes": 19217,
+      "sha256": "fa608e5abdc75e6393bba8aa405e3fe3d1f8b4c0f656e0ae0c5aa0b654a2f355"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/utils/__init__.py",
+      "bytes": 1,
+      "sha256": "01ba4719c80b6fe911b091a7c05124b64eeece964e09c058ef8f9805daca546b"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/utils/util.py",
+      "bytes": 5124,
+      "sha256": "654e0001976be1b0bbfcb8cb5c4def29fb5dabe55d7c6896a02603745d721296"
+    },
+    {
+      "path": "2610_variant_runtime/diffusion/utils/video_utils.py",
+      "bytes": 2859,
+      "sha256": "0ddb8a7cac82b24295c4c0f5a18f529df82e44c0d4a0bb5ea8acf0d128d37baf"
+    },
+    {
+      "path": "2610_variant_runtime/licenses/AnimateDiff.txt",
+      "bytes": 11357,
+      "sha256": "c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4"
+    },
+    {
+      "path": "2610_variant_runtime/licenses/DWPose.txt",
+      "bytes": 11994,
+      "sha256": "3abf0a61cbaa441957d4c5820be5185b910ed207abf80f60f92c9533b2e16e47"
+    },
+    {
+      "path": "2610_variant_runtime/licenses/OpenRAIL-M.txt",
+      "bytes": 14385,
+      "sha256": "be351ebe7ac01bcdbb018639aadcfd38f136b7dc3f2a3d4d3a24db51d1b210ef"
+    },
+    {
+      "path": "2610_variant_runtime/licenses/PixArt-alpha.txt",
+      "bytes": 11365,
+      "sha256": "6582247f37ea5556681f6224e86db09ec7272c99091e7284bb5c9cf83d6edca1"
+    },
+    {
+      "path": "2610_variant_runtime/licenses/diffusers.txt",
+      "bytes": 11357,
+      "sha256": "f9e2070c247517b1ddf65f7b11b393484a18da91a958fd18a97bd0f241c3125c"
+    },
+    {
+      "path": "2610_variant_runtime/licenses/magic-animate.txt",
+      "bytes": 1519,
+      "sha256": "6c76998d8b6a8c903d741de4f0c41fd7971d08746aebd2f39e9a9f0fb9b662e5"
+    },
+    {
+      "path": "2610_variant_runtime/requirements.lock.txt",
+      "bytes": 2766,
+      "sha256": "dfdd012e92ece6da107024040eded8ef55e75b9ab51ba0a8e95b19f0f7621f20"
+    },
+    {
+      "path": "2610_variant_runtime/scripts/download_assets.py",
+      "bytes": 1910,
+      "sha256": "50657ca0d2257d9fdd85c766ffd00ee4b5b0185e0a9a583e27ebd633c41ad488"
+    }
+  ],
+  "excludes": [
+    "model weights inside runtime",
+    "private inputs",
+    "training data",
+    "logs",
+    "tests",
+    "optimizer/RNG state",
+    "custom temporal decoder",
+    "git metadata",
+    "virtual environments"
+  ],
+  "hub_commit": null,
+  "fresh_hub_download_verification": "pending",
+  "gpu_render": "pending"
+}

2610_variant_asset_manifest.json

@@ -0,0 +1,395 @@
+{
+  "format_version": 1,
+  "status": "source_sha256_and_cpu_key_shape_verified_gpu_render_pending",
+  "prepared_at": "2026-10-17",
+  "repo_id": "H-Liu1997/Video-Motion-Graphs",
+  "existing_release_preserved": {
+    "path": "finetune_weighted/checkpoint-8000/pytorch_model.bin",
+    "sha256": "9cc0f0354c5ec4b33085196ba5786f845e0901fd1dc7159e57be14d972c371af",
+    "revision": "44ad3acb918e062f5f0c11ccc2e63cb0f3965e63"
+  },
+  "new_variant_count": 3,
+  "new_model_bytes": 26097472428,
+  "variants": [
+    {
+      "variant_id": "v7_dup_14_continue_2_step655000",
+      "relative_directory": "2409_variants/v7_dup_14_continue_2_step655000",
+      "source_material_month": "2024-09",
+      "date_basis": "Earliest confirmed source-material month; not an inferred training-start date.",
+      "original_run": "v7_dup_14_continue_2",
+      "checkpoint_step": 655000,
+      "scope": "Historical interpolation checkpoint. Its relationship to any final TANGO or ACInterp paper model is unconfirmed.",
+      "assets": [
+        {
+          "path": "2409_variants/v7_dup_14_continue_2_step655000/pytorch_model.bin",
+          "filename": "pytorch_model.bin",
+          "bytes": 8699157476,
+          "sha256": "beae0aadd1fc049da2f776ecef51a356727eccd18fa5cb5481738ae3adfbabe1",
+          "format": "PyTorch full model state dictionary",
+          "sha256_evidence": "Full SHA256 of the downloaded source file matches the source-storage SHA256; independent fresh-Hub-download verification is pending."
+        }
+      ],
+      "architecture": {
+        "code_family": "Custom pose-guided diffusion interpolation",
+        "seed_guider_channels_expected_from_source": 5,
+        "seed_guider_tensor_shape_verified": [
+          320,
+          5,
+          3,
+          3
+        ],
+        "context_frames_per_end": 3,
+        "generated_frames": 8,
+        "total_frames": 14,
+        "image_size": 512,
+        "verification_status": "CPU strict key/shape verified; GPU forward pending",
+        "seed_mask_channels": "4 latent + 1 known-frame mask"
+      },
+      "inference": {
+        "profile": "standard_pose",
+        "vae": "stabilityai/sd-vae-ft-mse",
+        "custom_temporal_decoder_required": false,
+        "code_repo": "https://github.com/H-Liu1997/Video-Motion-Graphs",
+        "code_revision": null,
+        "config_path": "2610_variant_runtime/diffusion/configs/2409_v7_dup_14_continue_2_step655000.yaml",
+        "config_sha256": "100c01244d99bc9bd263a9695e6cc95685b0acdebc17681741d709e3116ba5b4",
+        "environment_lock_path": "2610_variant_runtime/requirements.lock.txt",
+        "environment_lock_sha256": "dfdd012e92ece6da107024040eded8ef55e75b9ab51ba0a8e95b19f0f7621f20",
+        "commands": [
+          "WANDB_MODE=disabled .venv/bin/python diffusion/inference.py --config diffusion/configs/2409_v7_dup_14_continue_2_step655000.yaml --checkpoint ../2409_variants/v7_dup_14_continue_2_step655000/pytorch_model.bin --json inputs/input.json --output-dir outputs"
+        ],
+        "status": "frozen command; CPU key/shape passed; GPU execution pending",
+        "code_source_base_commit": "331fcff579d1bc60e54e568b9f524aa812f64c80",
+        "code_source_contains_frozen_uncommitted_changes": true,
+        "runtime_root": "2610_variant_runtime",
+        "runtime_manifest": "2610_runtime_asset_manifest.json",
+        "runtime_bundle_sha256": "2b1b5a1b19f8be6f97a74a640a085b728ee568d8953f8dbc7a3b5054ac5d8e95",
+        "command_working_directory": "2610_variant_runtime",
+        "seed": 2024,
+        "guidance_scale": 3.5,
+        "denoising_steps": 20
+      },
+      "validation": {
+        "source_file_sha256": "verified",
+        "source_sha256": "beae0aadd1fc049da2f776ecef51a356727eccd18fa5cb5481738ae3adfbabe1",
+        "source_bytes_verified": true,
+        "hub_commit": null,
+        "hub_lfs_sha256": "pending",
+        "fresh_download_sha256": "pending",
+        "cpu_key_shape_check": {
+          "status": "passed",
+          "method": "CPU mmap state dict against exact meta-constructed model; shape/key check without tensor copies or forward",
+          "key_count": 1974,
+          "missing_keys": [],
+          "unexpected_keys": [],
+          "seed_guider_weight_shape": [
+            320,
+            5,
+            3,
+            3
+          ],
+          "forward_executed": false
+        },
+        "strict_gpu_load": "pending",
+        "real_input_render": "pending",
+        "input_assets": [],
+        "output_video": null,
+        "hardware": null,
+        "environment": null,
+        "seed": 2024,
+        "metrics": null,
+        "receipt": null
+      },
+      "excluded_training_state": [
+        "optimizer",
+        "scheduler",
+        "mixed-precision scaler",
+        "random-number generator state"
+      ],
+      "training_resume_supported": false,
+      "license": "creativeml-openrail-m",
+      "license_note": "Retain the applicable Stable Diffusion / CreativeML OpenRAIL-M model terms and upstream component terms. Project source-code licensing is separate."
+    },
+    {
+      "variant_id": "v7_dup_14_s2_8frames_512_step9000",
+      "relative_directory": "2411_variants/v7_dup_14_s2_8frames_512_step9000",
+      "source_material_month": "2024-11",
+      "date_basis": "Earliest confirmed source-material month; not an inferred training-start date.",
+      "original_run": "v7_dup_14_s2_8frames_512",
+      "checkpoint_step": 9000,
+      "scope": "Historical checkpoint referenced by the 8-frame 512 inference configuration; no best-checkpoint or paper-final claim.",
+      "assets": [
+        {
+          "path": "2411_variants/v7_dup_14_s2_8frames_512_step9000/pytorch_model.bin",
+          "filename": "pytorch_model.bin",
+          "bytes": 8699157476,
+          "sha256": "6a1288711eaf2792f3581b3e175fb79bb8988e5b192a61178de322c587e810b8",
+          "format": "PyTorch full model state dictionary",
+          "sha256_evidence": "Full SHA256 of the downloaded source file matches the source-storage SHA256; independent fresh-Hub-download verification is pending."
+        }
+      ],
+      "architecture": {
+        "code_family": "Custom pose-guided diffusion interpolation",
+        "seed_guider_channels_expected_from_source": 5,
+        "seed_guider_tensor_shape_verified": [
+          320,
+          5,
+          3,
+          3
+        ],
+        "context_frames_per_end": 3,
+        "generated_frames": 8,
+        "total_frames": 14,
+        "image_size": 512,
+        "verification_status": "CPU strict key/shape verified; GPU forward pending",
+        "seed_mask_channels": "4 latent + 1 known-frame mask"
+      },
+      "inference": {
+        "profile": "standard_pose",
+        "vae": "stabilityai/sd-vae-ft-mse",
+        "custom_temporal_decoder_required": false,
+        "code_repo": "https://github.com/H-Liu1997/Video-Motion-Graphs",
+        "code_revision": null,
+        "config_path": "2610_variant_runtime/diffusion/configs/2411_v7_dup_14_s2_8frames_512_step9000.yaml",
+        "config_sha256": "37cf67ec1a2b9c8020927ca849fae7a8da7b63762d9c9010a59570b0619a7aa3",
+        "environment_lock_path": "2610_variant_runtime/requirements.lock.txt",
+        "environment_lock_sha256": "dfdd012e92ece6da107024040eded8ef55e75b9ab51ba0a8e95b19f0f7621f20",
+        "commands": [
+          "WANDB_MODE=disabled .venv/bin/python diffusion/inference.py --config diffusion/configs/2411_v7_dup_14_s2_8frames_512_step9000.yaml --checkpoint ../2411_variants/v7_dup_14_s2_8frames_512_step9000/pytorch_model.bin --json inputs/input.json --output-dir outputs"
+        ],
+        "status": "frozen command; CPU key/shape passed; GPU execution pending",
+        "code_source_base_commit": "331fcff579d1bc60e54e568b9f524aa812f64c80",
+        "code_source_contains_frozen_uncommitted_changes": true,
+        "runtime_root": "2610_variant_runtime",
+        "runtime_manifest": "2610_runtime_asset_manifest.json",
+        "runtime_bundle_sha256": "2b1b5a1b19f8be6f97a74a640a085b728ee568d8953f8dbc7a3b5054ac5d8e95",
+        "command_working_directory": "2610_variant_runtime",
+        "seed": 2024,
+        "guidance_scale": 1.0,
+        "denoising_steps": 20
+      },
+      "validation": {
+        "source_file_sha256": "verified",
+        "source_sha256": "6a1288711eaf2792f3581b3e175fb79bb8988e5b192a61178de322c587e810b8",
+        "source_bytes_verified": true,
+        "hub_commit": null,
+        "hub_lfs_sha256": "pending",
+        "fresh_download_sha256": "pending",
+        "cpu_key_shape_check": {
+          "status": "passed",
+          "method": "CPU mmap state dict against exact meta-constructed model; shape/key check without tensor copies or forward",
+          "key_count": 1974,
+          "missing_keys": [],
+          "unexpected_keys": [],
+          "seed_guider_weight_shape": [
+            320,
+            5,
+            3,
+            3
+          ],
+          "forward_executed": false
+        },
+        "strict_gpu_load": "pending",
+        "real_input_render": "pending",
+        "input_assets": [],
+        "output_video": null,
+        "hardware": null,
+        "environment": null,
+        "seed": 2024,
+        "metrics": null,
+        "receipt": null
+      },
+      "excluded_training_state": [
+        "optimizer",
+        "scheduler",
+        "mixed-precision scaler",
+        "random-number generator state"
+      ],
+      "training_resume_supported": false,
+      "license": "creativeml-openrail-m",
+      "license_note": "Retain the applicable Stable Diffusion / CreativeML OpenRAIL-M model terms and upstream component terms. Project source-code licensing is separate."
+    },
+    {
+      "variant_id": "v7_dup_14_s2_12frames_512_step30000",
+      "relative_directory": "2411_variants/v7_dup_14_s2_12frames_512_step30000",
+      "source_material_month": "2024-11",
+      "date_basis": "Earliest confirmed source-material month; not an inferred training-start date.",
+      "original_run": "v7_dup_14_s2_12frames_512",
+      "checkpoint_step": 30000,
+      "scope": "Historical checkpoint referenced by the 12-frame 512 inference configuration; no best-checkpoint or paper-final claim.",
+      "assets": [
+        {
+          "path": "2411_variants/v7_dup_14_s2_12frames_512_step30000/pytorch_model.bin",
+          "filename": "pytorch_model.bin",
+          "bytes": 8699157476,
+          "sha256": "98bc64d7c931a7e414e9e9b45709420477d8625055043e945ad01d121f2302fb",
+          "format": "PyTorch full model state dictionary",
+          "sha256_evidence": "Full SHA256 of the downloaded source file matches the source-storage SHA256; independent fresh-Hub-download verification is pending."
+        }
+      ],
+      "architecture": {
+        "code_family": "Custom pose-guided diffusion interpolation",
+        "seed_guider_channels_expected_from_source": 5,
+        "seed_guider_tensor_shape_verified": [
+          320,
+          5,
+          3,
+          3
+        ],
+        "context_frames_per_end": 1,
+        "generated_frames": 12,
+        "total_frames": 14,
+        "image_size": 512,
+        "verification_status": "CPU strict key/shape verified; GPU forward pending",
+        "seed_mask_channels": "4 latent + 1 known-frame mask"
+      },
+      "inference": {
+        "profile": "standard_pose",
+        "vae": "stabilityai/sd-vae-ft-mse",
+        "custom_temporal_decoder_required": false,
+        "code_repo": "https://github.com/H-Liu1997/Video-Motion-Graphs",
+        "code_revision": null,
+        "config_path": "2610_variant_runtime/diffusion/configs/2411_v7_dup_14_s2_12frames_512_step30000.yaml",
+        "config_sha256": "7258d8124521c81d2b2724115f0e55d2ece513f8715a0f8a3d378c6537096ef1",
+        "environment_lock_path": "2610_variant_runtime/requirements.lock.txt",
+        "environment_lock_sha256": "dfdd012e92ece6da107024040eded8ef55e75b9ab51ba0a8e95b19f0f7621f20",
+        "commands": [
+          "WANDB_MODE=disabled .venv/bin/python diffusion/inference.py --config diffusion/configs/2411_v7_dup_14_s2_12frames_512_step30000.yaml --checkpoint ../2411_variants/v7_dup_14_s2_12frames_512_step30000/pytorch_model.bin --json inputs/input.json --output-dir outputs"
+        ],
+        "status": "frozen command; CPU key/shape passed; GPU execution pending",
+        "code_source_base_commit": "331fcff579d1bc60e54e568b9f524aa812f64c80",
+        "code_source_contains_frozen_uncommitted_changes": true,
+        "runtime_root": "2610_variant_runtime",
+        "runtime_manifest": "2610_runtime_asset_manifest.json",
+        "runtime_bundle_sha256": "2b1b5a1b19f8be6f97a74a640a085b728ee568d8953f8dbc7a3b5054ac5d8e95",
+        "command_working_directory": "2610_variant_runtime",
+        "seed": 2024,
+        "guidance_scale": 3.5,
+        "denoising_steps": 20
+      },
+      "validation": {
+        "source_file_sha256": "verified",
+        "source_sha256": "98bc64d7c931a7e414e9e9b45709420477d8625055043e945ad01d121f2302fb",
+        "source_bytes_verified": true,
+        "hub_commit": null,
+        "hub_lfs_sha256": "pending",
+        "fresh_download_sha256": "pending",
+        "cpu_key_shape_check": {
+          "status": "passed",
+          "method": "CPU mmap state dict against exact meta-constructed model; shape/key check without tensor copies or forward",
+          "key_count": 1974,
+          "missing_keys": [],
+          "unexpected_keys": [],
+          "seed_guider_weight_shape": [
+            320,
+            5,
+            3,
+            3
+          ],
+          "forward_executed": false
+        },
+        "strict_gpu_load": "pending",
+        "real_input_render": "pending",
+        "input_assets": [],
+        "output_video": null,
+        "hardware": null,
+        "environment": null,
+        "seed": 2024,
+        "metrics": null,
+        "receipt": null
+      },
+      "excluded_training_state": [
+        "optimizer",
+        "scheduler",
+        "mixed-precision scaler",
+        "random-number generator state"
+      ],
+      "training_resume_supported": false,
+      "license": "creativeml-openrail-m",
+      "license_note": "Retain the applicable Stable Diffusion / CreativeML OpenRAIL-M model terms and upstream component terms. Project source-code licensing is separate."
+    }
+  ],
+  "base_components": [
+    {
+      "path": "weights/image_encoder/pytorch_model.bin",
+      "group": "base",
+      "url": "https://huggingface.co/lambda/sd-image-variations-diffusers/resolve/42bc0ee1726b141d49f519a6ea02ccfbf073db2e/image_encoder/pytorch_model.bin",
+      "bytes": 1215993967,
+      "sha256": "89d2aa29b5fdf64f3ad4f45fb4227ea98bc45156bbae673b85be1af7783dbabb"
+    },
+    {
+      "path": "weights/image_encoder/config.json",
+      "group": "base",
+      "url": "https://huggingface.co/lambda/sd-image-variations-diffusers/resolve/42bc0ee1726b141d49f519a6ea02ccfbf073db2e/image_encoder/config.json",
+      "bytes": 703,
+      "sha256": "ba41139cc9dad159163a53e2aa847356b627f53fef988ad461659bd2d587d886"
+    },
+    {
+      "path": "weights/stable-diffusion-v1-5/unet/diffusion_pytorch_model.bin",
+      "group": "base",
+      "url": "https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/451f4fe16113bff5a5d2269ed5ad43b0592e9a14/unet/diffusion_pytorch_model.bin",
+      "bytes": 3438354725,
+      "sha256": "c7da0e21ba7ea50637bee26e81c220844defdf01aafca02b2c42ecdadb813de4"
+    },
+    {
+      "path": "weights/stable-diffusion-v1-5/unet/config.json",
+      "group": "base",
+      "url": "https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/451f4fe16113bff5a5d2269ed5ad43b0592e9a14/unet/config.json",
+      "bytes": 743,
+      "sha256": "78f474de6bab3d893868f37be97b636ae65c0df3073ed3256ca458ff599b5f96"
+    },
+    {
+      "path": "weights/sd-vae-ft-mse/diffusion_pytorch_model.bin",
+      "group": "base",
+      "url": "https://huggingface.co/stabilityai/sd-vae-ft-mse/resolve/31f26fdeee1355a5c34592e401dd41e45d25a493/diffusion_pytorch_model.bin",
+      "bytes": 334707217,
+      "sha256": "1b4889b6b1d4ce7ae320a02dedaeff1780ad77d415ea0d744b476155c6377ddc"
+    },
+    {
+      "path": "weights/sd-vae-ft-mse/config.json",
+      "group": "base",
+      "url": "https://huggingface.co/stabilityai/sd-vae-ft-mse/resolve/31f26fdeee1355a5c34592e401dd41e45d25a493/config.json",
+      "bytes": 547,
+      "sha256": "92d3dfb746fca211a2c9e019e285f8597412211728dce3c5bcf4eda0f2d62e7e"
+    },
+    {
+      "path": "weights/animatediff/mm_sd_v15_v2.ckpt",
+      "group": "base",
+      "url": "https://huggingface.co/guoyww/animatediff/resolve/fdfe36afa161e51b3e9c24022b0e368d59e7345e/mm_sd_v15_v2.ckpt",
+      "bytes": 1817888431,
+      "sha256": "69ed0f5fef82b110aca51bcab73b21104242bc65d6ab4b8b2a2a94d31cad1bf0"
+    }
+  ],
+  "base_components_note": "Seven pinned standard-VAE base files totaling 6,806,946,333 bytes have passed full SHA256 checks; GPU inference remains pending.",
+  "runtime": {
+    "root": "2610_variant_runtime",
+    "source_base_commit": "331fcff579d1bc60e54e568b9f524aa812f64c80",
+    "source_contains_frozen_uncommitted_changes": true,
+    "bundle_sha256": "2b1b5a1b19f8be6f97a74a640a085b728ee568d8953f8dbc7a3b5054ac5d8e95",
+    "total_file_count": 34,
+    "total_bytes": 420063,
+    "manifest": "2610_runtime_asset_manifest.json"
+  },
+  "upstream_license_metadata": [
+    {
+      "repo": "stabilityai/sd-vae-ft-mse",
+      "revision": "31f26fdeee1355a5c34592e401dd41e45d25a493",
+      "license": "mit"
+    },
+    {
+      "repo": "guoyww/animatediff",
+      "revision": "fdfe36afa161e51b3e9c24022b0e368d59e7345e",
+      "license": "apache-2.0"
+    },
+    {
+      "repo": "lambda/sd-image-variations-diffusers",
+      "revision": "42bc0ee1726b141d49f519a6ea02ccfbf073db2e",
+      "license": "creativeml-openrail-m"
+    },
+    {
+      "repo": "stable-diffusion-v1-5/stable-diffusion-v1-5",
+      "revision": "451f4fe16113bff5a5d2269ed5ad43b0592e9a14",
+      "license": "creativeml-openrail-m"
+    }
+  ]
+}

2610_variant_runtime/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, Any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

2610_variant_runtime/NOTICE

@@ -0,0 +1,30 @@
+Video Motion Graphs
+Copyright 2026 Haiyang Liu and contributors.
+
+Original project code is licensed under Apache License 2.0. The following
+adapted components retain their original notices and licenses:
+
+Hugging Face Diffusers: diffusion/models/{attention,resnet,transformer_2d,
+  unet_2d_blocks,unet_2d_condition,unet_3d_blocks}.py and diffusion/utils/util.py.
+  Apache-2.0; see licenses/diffusers.txt.
+AnimateDiff: diffusion/models/{motion_module,unet_3d}.py.
+  Apache-2.0; see licenses/AnimateDiff.txt.
+MagicAnimate: diffusion/models/mutual_self_attention.py.
+  BSD-3-Clause; see licenses/magic-animate.txt.
+DWPose: vmg/dataloaders/tools/dwpose/.
+  Apache-2.0; see licenses/DWPose.txt.
+PixArt-alpha: attention implementation cited in diffusion/models/attention.py.
+  Apache-2.0; see licenses/PixArt-alpha.txt.
+
+These implementations were adapted for weighted-pose video interpolation,
+reference attention, motion-graph construction and local video preprocessing.
+Original attribution headers remain in the corresponding source files.
+
+Downloaded model weights retain their own terms. Stable Diffusion 1.5,
+SD image variations and the derived VMG checkpoint use CreativeML OpenRAIL-M
+(licenses/OpenRAIL-M.txt). SD-VAE-ft-mse declares MIT and AnimateDiff model
+weights declare Apache-2.0 in their upstream model cards. The project source
+license does not replace third-party model or example-media terms.
+
+The five-channel seed-mask inference path is adapted from Haiyang Liu's
+November 2024 HMInterp research code.

2610_variant_runtime/README.md

@@ -0,0 +1,115 @@
+# 2610 standalone historical-variant runtime
+
+This bundle runs the three dated five-channel interpolation checkpoints in this model repository with the standard SD VAE. Model SHA256 checks and strict CPU key/shape checks have passed. **GPU forward execution, rendered-video evaluation and fresh-Hub-download verification are pending.** CPU checks do not establish inference quality or a successful GPU run.
+
+The code is a frozen snapshot based on upstream commit `331fcff579d1bc60e54e568b9f524aa812f64c80` with the included variant support. It is identified by the per-file hashes in [runtime manifest](../2610_runtime_asset_manifest.json), not by claiming that the upstream commit already contains these changes. Only the standard-VAE profile is included. The original four-channel `finetune_weighted/checkpoint-8000` release is separate.
+
+## Download
+
+Install `uv`, then run this in a new directory. It resolves one Hub commit and downloads all three variants and the runtime from that same revision. The transfer tools run outside the model environment.
+
+```sh
+uv run --no-project --with huggingface_hub==2.2.0 python - <<'PY'
+from huggingface_hub import HfApi, snapshot_download
+repo = "H-Liu1997/Video-Motion-Graphs"
+revision = HfApi(token=False).model_info(repo).sha
+snapshot_download(
+    repo_id=repo,
+    revision=revision,
+    token=False,
+    local_dir="vmg-variant-release",
+    allow_patterns=[
+        "2610_variant_runtime/*",
+        "2610_runtime_asset_manifest.json",
+        "2610_variant_asset_manifest.json",
+        "2409_variants/v7_dup_14_continue_2_step655000/*",
+        "2411_variants/v7_dup_14_s2_8frames_512_step9000/*",
+        "2411_variants/v7_dup_14_s2_12frames_512_step30000/*",
+    ],
+)
+print("Downloaded revision:", revision)
+PY
+```
+
+Record the printed commit for reproduction. Before inference, verify the downloaded runtime and model files against their manifests:
+
+```sh
+cd vmg-variant-release
+python3 - <<'PY'
+import hashlib, json
+from pathlib import Path
+root = Path.cwd()
+runtime = json.loads((root / "2610_runtime_asset_manifest.json").read_text())
+variants = json.loads((root / "2610_variant_asset_manifest.json").read_text())
+files = list(runtime["files"])
+files += [a for v in variants["variants"] for a in v["assets"]]
+for item in files:
+    path = root / item["path"]
+    assert path.stat().st_size == item["bytes"], item["path"]
+    digest = hashlib.sha256()
+    with path.open("rb") as stream:
+        for chunk in iter(lambda: stream.read(8 * 1024 * 1024), b""):
+            digest.update(chunk)
+    assert digest.hexdigest() == item["sha256"], item["path"]
+print("Verified files:", len(files))
+PY
+```
+
+## Install and fetch base components
+
+The required platform is Linux x86_64 with an NVIDIA CUDA-capable GPU. The model environment pins Python 3.10, PyTorch 2.2.2 / CUDA 12.1 and Diffusers 0.24.0. Variant GPU memory requirements and inference execution remain unverified.
+
+```sh
+cd 2610_variant_runtime
+uv venv --python 3.10 .venv
+uv pip install --python .venv/bin/python --no-deps --index-strategy unsafe-best-match --extra-index-url https://download.pytorch.org/whl/cu121 -r requirements.lock.txt
+.venv/bin/python scripts/download_assets.py --group base
+.venv/bin/python scripts/download_assets.py --group base --verify-only
+```
+
+The seven base files total 6,806,946,333 bytes. Their public URLs and SHA256 hashes are pinned in `asset_manifest.json`. The downloader places them in this runtime's `weights/` directory. Keep the inference environment's pinned `huggingface-hub==0.25.2`; use a separate tool environment for modern Hub transfers.
+
+## Supply aligned input
+
+Provide your own RGB video and its aligned weighted-pose heatmap video with the same frame rate and dimensions. Each needs at least 14 frames from `start_idx`. Ordinary skeleton-only pose renderings are not interchangeable with the weighted heatmaps expected by this profile. This bundle contains no private validation videos, preprocessing models, training data or logs.
+
+Save `inputs/input.json` with the following structure, replacing only the two input paths:
+
+```json
+[
+  {
+    "video_path": "/path/to/rgb.mp4",
+    "kps_path": "/path/to/weighted_heatmap.mp4",
+    "mode": "test",
+    "start_idx": 0
+  }
+]
+```
+
+## Run a selected variant
+
+Run from `2610_variant_runtime/`. The explicit `--checkpoint` argument selects the dated model file; the configuration's historical default weight path is not the Hub layout.
+
+```sh
+WANDB_MODE=disabled .venv/bin/python diffusion/inference.py --config diffusion/configs/2409_v7_dup_14_continue_2_step655000.yaml --checkpoint ../2409_variants/v7_dup_14_continue_2_step655000/pytorch_model.bin --json inputs/input.json --output-dir outputs
+
+WANDB_MODE=disabled .venv/bin/python diffusion/inference.py --config diffusion/configs/2411_v7_dup_14_s2_8frames_512_step9000.yaml --checkpoint ../2411_variants/v7_dup_14_s2_8frames_512_step9000/pytorch_model.bin --json inputs/input.json --output-dir outputs
+
+WANDB_MODE=disabled .venv/bin/python diffusion/inference.py --config diffusion/configs/2411_v7_dup_14_s2_12frames_512_step30000.yaml --checkpoint ../2411_variants/v7_dup_14_s2_12frames_512_step30000/pytorch_model.bin --json inputs/input.json --output-dir outputs
+```
+
+These are the configured commands, not a claim that a GPU render has already completed. Outputs are separated by the configuration's experiment name. Local inference receipts can contain your input paths; review them before sharing.
+
+| Variant | Seed | Context + Generated + Context | Size | DDIM Steps | Guidance |
+| --- | ---: | --- | --- | ---: | ---: |
+| 2409 step 655000 | 2024 | 3 + 8 + 3 | 512 square with black padding | 20 | 3.5 |
+| 2411 8-frame step 9000 | 2024 | 3 + 8 + 3 | 512 square with black padding | 20 | 1.0 |
+| 2411 12-frame step 30000 | 2024 | 1 + 12 + 1 | 512 square with black padding | 20 | 3.5 |
+
+All three checkpoints have 1,974 keys and a seed-guider convolution with shape `[320, 5, 3, 3]`: four latent channels plus one known-frame mask. A strict CPU check against the meta-constructed model found no missing or unexpected keys and no shape mismatch. No forward pass was performed by that check.
+
+## License and provenance
+
+Original project runtime code is Apache-2.0, with adapted components retaining the notices in [NOTICE](NOTICE) and `licenses/`, including the MagicAnimate BSD-3-Clause terms. This source-code license does not relicense model weights or user inputs. The variant model weights retain the repository's CreativeML OpenRAIL-M terms. Upstream SD 1.5 and image-variation components declare CreativeML OpenRAIL-M, SD-VAE-ft-mse declares MIT, and AnimateDiff weights declare Apache-2.0 at the pinned revisions. Preserve these upstream terms.
+
+The `2409` and `2411` model-directory prefixes describe source material from September and November 2024. The `2610` runtime prefix describes this October 2026 packaging. No best-checkpoint, final-TANGO/ACInterp or paper-benchmark equivalence claim is made. Training optimizer, scheduler, scaler and RNG state are not included.

2610_variant_runtime/asset_manifest.json

@@ -0,0 +1,54 @@
+{
+  "format_version": 1,
+  "assets": [
+    {
+      "path": "weights/image_encoder/pytorch_model.bin",
+      "group": "base",
+      "url": "https://huggingface.co/lambda/sd-image-variations-diffusers/resolve/42bc0ee1726b141d49f519a6ea02ccfbf073db2e/image_encoder/pytorch_model.bin",
+      "bytes": 1215993967,
+      "sha256": "89d2aa29b5fdf64f3ad4f45fb4227ea98bc45156bbae673b85be1af7783dbabb"
+    },
+    {
+      "path": "weights/image_encoder/config.json",
+      "group": "base",
+      "url": "https://huggingface.co/lambda/sd-image-variations-diffusers/resolve/42bc0ee1726b141d49f519a6ea02ccfbf073db2e/image_encoder/config.json",
+      "bytes": 703,
+      "sha256": "ba41139cc9dad159163a53e2aa847356b627f53fef988ad461659bd2d587d886"
+    },
+    {
+      "path": "weights/stable-diffusion-v1-5/unet/diffusion_pytorch_model.bin",
+      "group": "base",
+      "url": "https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/451f4fe16113bff5a5d2269ed5ad43b0592e9a14/unet/diffusion_pytorch_model.bin",
+      "bytes": 3438354725,
+      "sha256": "c7da0e21ba7ea50637bee26e81c220844defdf01aafca02b2c42ecdadb813de4"
+    },
+    {
+      "path": "weights/stable-diffusion-v1-5/unet/config.json",
+      "group": "base",
+      "url": "https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/451f4fe16113bff5a5d2269ed5ad43b0592e9a14/unet/config.json",
+      "bytes": 743,
+      "sha256": "78f474de6bab3d893868f37be97b636ae65c0df3073ed3256ca458ff599b5f96"
+    },
+    {
+      "path": "weights/sd-vae-ft-mse/diffusion_pytorch_model.bin",
+      "group": "base",
+      "url": "https://huggingface.co/stabilityai/sd-vae-ft-mse/resolve/31f26fdeee1355a5c34592e401dd41e45d25a493/diffusion_pytorch_model.bin",
+      "bytes": 334707217,
+      "sha256": "1b4889b6b1d4ce7ae320a02dedaeff1780ad77d415ea0d744b476155c6377ddc"
+    },
+    {
+      "path": "weights/sd-vae-ft-mse/config.json",
+      "group": "base",
+      "url": "https://huggingface.co/stabilityai/sd-vae-ft-mse/resolve/31f26fdeee1355a5c34592e401dd41e45d25a493/config.json",
+      "bytes": 547,
+      "sha256": "92d3dfb746fca211a2c9e019e285f8597412211728dce3c5bcf4eda0f2d62e7e"
+    },
+    {
+      "path": "weights/animatediff/mm_sd_v15_v2.ckpt",
+      "group": "base",
+      "url": "https://huggingface.co/guoyww/animatediff/resolve/fdfe36afa161e51b3e9c24022b0e368d59e7345e/mm_sd_v15_v2.ckpt",
+      "bytes": 1817888431,
+      "sha256": "69ed0f5fef82b110aca51bcab73b21104242bc65d6ab4b8b2a2a94d31cad1bf0"
+    }
+  ]
+}

2610_variant_runtime/diffusion/configs/2409_v7_dup_14_continue_2_step655000.yaml

@@ -0,0 +1,55 @@
+wandb_project: video-motion-graphs
+exp_name: 2409_v7_dup_14_continue_2_step655000
+wandb_log_dir: outputs/wandb
+output_dir: outputs/diffusion
+seed: 2024
+base_model_path: weights/stable-diffusion-v1-5
+vae_model_path: weights/sd-vae-ft-mse
+image_encoder_path: weights/image_encoder
+mm_path: weights/animatediff/mm_sd_v15_v2.ckpt
+checkpoint_path: weights/variants/2409_v7_dup_14_continue_2_step655000/pytorch_model.bin
+seed_input_channels: 5
+data:
+  meta_paths:
+  - cache/Sequence_01_6/inference.json
+  image_size: 512
+  sample_frames: 14
+  seed_frames: 3
+  generated_frames: 8
+validation:
+  guidance_scale: 3.5
+  denoising_steps: 20
+solver:
+  gradient_accumulation_steps: 1
+  mixed_precision: fp16
+  enable_xformers_memory_efficient_attention: true
+noise_scheduler_kwargs:
+  num_train_timesteps: 1000
+  beta_start: 0.00085
+  beta_end: 0.012
+  beta_schedule: linear
+  steps_offset: 1
+  clip_sample: false
+unet_additional_kwargs:
+  use_inflated_groupnorm: true
+  unet_use_cross_frame_attention: false
+  unet_use_temporal_attention: false
+  use_motion_module: true
+  motion_module_resolutions:
+  - 1
+  - 2
+  - 4
+  - 8
+  motion_module_mid_block: true
+  motion_module_decoder_only: false
+  motion_module_type: Vanilla
+  motion_module_kwargs:
+    num_attention_heads: 8
+    num_transformer_block: 1
+    attention_block_types:
+    - Temporal_Self
+    - Temporal_Self
+    temporal_position_encoding: true
+    temporal_position_encoding_max_len: 32
+    temporal_attention_dim_div: 1
+enable_zero_snr: true

2610_variant_runtime/diffusion/configs/2411_v7_dup_14_s2_12frames_512_step30000.yaml

@@ -0,0 +1,55 @@
+wandb_project: video-motion-graphs
+exp_name: 2411_v7_dup_14_s2_12frames_512_step30000
+wandb_log_dir: outputs/wandb
+output_dir: outputs/diffusion
+seed: 2024
+base_model_path: weights/stable-diffusion-v1-5
+vae_model_path: weights/sd-vae-ft-mse
+image_encoder_path: weights/image_encoder
+mm_path: weights/animatediff/mm_sd_v15_v2.ckpt
+checkpoint_path: weights/variants/2411_v7_dup_14_s2_12frames_512_step30000/pytorch_model.bin
+seed_input_channels: 5
+data:
+  meta_paths:
+  - cache/Sequence_01_6/inference.json
+  image_size: 512
+  sample_frames: 14
+  seed_frames: 1
+  generated_frames: 12
+validation:
+  guidance_scale: 3.5
+  denoising_steps: 20
+solver:
+  gradient_accumulation_steps: 1
+  mixed_precision: fp16
+  enable_xformers_memory_efficient_attention: true
+noise_scheduler_kwargs:
+  num_train_timesteps: 1000
+  beta_start: 0.00085
+  beta_end: 0.012
+  beta_schedule: linear
+  steps_offset: 1
+  clip_sample: false
+unet_additional_kwargs:
+  use_inflated_groupnorm: true
+  unet_use_cross_frame_attention: false
+  unet_use_temporal_attention: false
+  use_motion_module: true
+  motion_module_resolutions:
+  - 1
+  - 2
+  - 4
+  - 8
+  motion_module_mid_block: true
+  motion_module_decoder_only: false
+  motion_module_type: Vanilla
+  motion_module_kwargs:
+    num_attention_heads: 8
+    num_transformer_block: 1
+    attention_block_types:
+    - Temporal_Self
+    - Temporal_Self
+    temporal_position_encoding: true
+    temporal_position_encoding_max_len: 32
+    temporal_attention_dim_div: 1
+enable_zero_snr: true

2610_variant_runtime/diffusion/configs/2411_v7_dup_14_s2_8frames_512_step9000.yaml

@@ -0,0 +1,55 @@
+wandb_project: video-motion-graphs
+exp_name: 2411_v7_dup_14_s2_8frames_512_step9000
+wandb_log_dir: outputs/wandb
+output_dir: outputs/diffusion
+seed: 2024
+base_model_path: weights/stable-diffusion-v1-5
+vae_model_path: weights/sd-vae-ft-mse
+image_encoder_path: weights/image_encoder
+mm_path: weights/animatediff/mm_sd_v15_v2.ckpt
+checkpoint_path: weights/variants/2411_v7_dup_14_s2_8frames_512_step9000/pytorch_model.bin
+seed_input_channels: 5
+data:
+  meta_paths:
+  - cache/Sequence_01_6/inference.json
+  image_size: 512
+  sample_frames: 14
+  seed_frames: 3
+  generated_frames: 8
+validation:
+  guidance_scale: 1.0
+  denoising_steps: 20
+solver:
+  gradient_accumulation_steps: 1
+  mixed_precision: fp16
+  enable_xformers_memory_efficient_attention: true
+noise_scheduler_kwargs:
+  num_train_timesteps: 1000
+  beta_start: 0.00085
+  beta_end: 0.012
+  beta_schedule: linear
+  steps_offset: 1
+  clip_sample: false
+unet_additional_kwargs:
+  use_inflated_groupnorm: true
+  unet_use_cross_frame_attention: false
+  unet_use_temporal_attention: false
+  use_motion_module: true
+  motion_module_resolutions:
+  - 1
+  - 2
+  - 4
+  - 8
+  motion_module_mid_block: true
+  motion_module_decoder_only: false
+  motion_module_type: Vanilla
+  motion_module_kwargs:
+    num_attention_heads: 8
+    num_transformer_block: 1
+    attention_block_types:
+    - Temporal_Self
+    - Temporal_Self
+    temporal_position_encoding: true
+    temporal_position_encoding_max_len: 32
+    temporal_attention_dim_div: 1
+enable_zero_snr: true

2610_variant_runtime/diffusion/inference.py

@@ -0,0 +1,468 @@
+import argparse
+import logging
+import os
+import warnings
+from pathlib import Path
+import json
+
+import wandb
+from datetime import datetime
+import diffusers
+import numpy as np
+import torch
+import torch.utils.checkpoint
+import transformers
+from accelerate import Accelerator
+from accelerate.logging import get_logger
+from accelerate.utils import DistributedDataParallelKwargs
+from diffusers import AutoencoderKL, DDIMScheduler
+from diffusers.utils import check_min_version
+from diffusers.utils.import_utils import is_xformers_available
+from omegaconf import OmegaConf
+from PIL import Image, ImageOps
+from transformers import CLIPVisionModelWithProjection
+from decord import VideoReader
+
+from pipelines.pipeline_vfi2video_share_vae import Pose2VideoPipeline
+from models.model_vfi import VFIShareVAEModel
+from models.pose_guider import PoseGuider
+from models.conv_in import ConvIn
+
+from models.unet_2d_condition import UNet2DConditionModel
+from models.unet_3d import UNet3DConditionModel
+from models.mutual_self_attention import ReferenceAttentionControl
+
+from utils.util import seed_everything
+from utils.video_utils import concat_pil
+
+warnings.filterwarnings("ignore")
+check_min_version("0.10.0.dev0")
+logger = get_logger(__name__, log_level="INFO")
+
+def save_videos_from_pil(pil_images, path, fps=8, bitrate=None):
+    import av
+
+    save_fmt = Path(path).suffix
+    os.makedirs(os.path.dirname(path), exist_ok=True)
+    width, height = pil_images[0].size
+
+    if save_fmt == ".mp4":
+        codec = "libx264"
+        container = av.open(path, "w")
+        stream = container.add_stream(codec, rate=fps, )
+
+        stream.width = width
+        stream.height = height
+        if bitrate is not None:
+            stream.bit_rate = bitrate
+
+        for pil_image in pil_images:
+            # pil_image = Image.fromarray(image_arr).convert("RGB")
+            av_frame = av.VideoFrame.from_image(pil_image)
+            container.mux(stream.encode(av_frame))
+        container.mux(stream.encode())
+        container.close()
+
+    elif save_fmt == ".gif":
+        pil_images[0].save(
+            fp=path,
+            format="GIF",
+            append_images=pil_images[1:],
+            save_all=True,
+            duration=(1 / fps * 1000),
+            loop=0,
+        )
+    else:
+        raise ValueError("Unsupported file type. Use .mp4 or .gif.")
+
+
+def pad_to_square(image, size):
+    width, height = image.size
+    ratio = size / max(width, height)
+    width, height = int(width * ratio), int(height * ratio)
+    image = image.resize((width, height))
+    left, top = (size - width) // 2, (size - height) // 2
+    return ImageOps.expand(image, (left, top, size - width - left, size - height - top), fill=0)
+
+
+def log_validation(
+        cfg,
+        vae,
+        image_enc,
+        model,
+        scheduler,
+        accelerator,
+        width,
+        height,
+        seed=42,
+        dtype=torch.float32,
+    ):
+    logger.info("Running validation ...")
+    unwrap_model = accelerator.unwrap_model(model).eval()
+    denoising_unet = unwrap_model.denoising_unet
+    reference_unet = unwrap_model.reference_unet
+    pose_guider = unwrap_model.pose_guider
+    seed_guider = unwrap_model.seed_guider
+    generator = torch.manual_seed(seed)
+    vae = vae.to(dtype=dtype)
+    image_enc = image_enc.to(dtype=dtype)
+
+    pipeline = Pose2VideoPipeline(
+        vae=vae,
+        image_encoder=image_enc,
+        reference_unet=reference_unet,
+        denoising_unet=denoising_unet,
+        pose_guider=pose_guider,
+        seed_guider=seed_guider,
+        scheduler=scheduler,
+    )
+    pipeline = pipeline.to(accelerator.device)
+
+    vid_meta = []
+    for data_meta_path in cfg.data.meta_paths:
+        vid_meta.extend(json.load(open(data_meta_path, "r")))
+    all_test_cases = [item for item in vid_meta if item.get("mode") == "test"]
+    current_device_id = accelerator.process_index
+    pil_images = []
+
+    # if cfg.wandb_project == "debug":
+    #     all_test_cases = all_test_cases[:2] + all_test_cases[-2:]
+    # else:
+    #     all_test_cases = all_test_cases[:8] + all_test_cases[-8:]
+
+    with torch.no_grad():
+        for idx, test_case in enumerate(all_test_cases):
+            # if idx > 1: break
+            if idx % accelerator.num_processes == current_device_id:
+                video_path = test_case["video_path"]
+                kps_pth = test_case["kps_path"]
+                rgb_video = VideoReader(video_path)
+                kps_video = VideoReader(kps_pth)
+
+                # if idx >= len(all_test_cases) // 2:
+                #     # for dance video we pick from 8s
+                #     ref_id = 120
+                # else:
+                ref_id = test_case.get("start_idx", 3 - cfg.data.seed_frames) + cfg.data.seed_frames
+
+                start_g = ref_id - cfg.data.seed_frames
+                end_s = ref_id + cfg.data.generated_frames
+                end_g = ref_id + cfg.data.generated_frames + cfg.data.seed_frames
+
+                ref_img = rgb_video[ref_id-1]
+                ref_img_pil = Image.fromarray(ref_img.asnumpy())
+                w, h = ref_img_pil.size
+
+                tgt_vidpil_lst = []
+                tgt_guid_vid_list = []
+                pose_guid_vid_list = []
+                for index in range(start_g, ref_id):
+                    img = rgb_video[index]
+                    pil_img = Image.fromarray(img.asnumpy())
+                    tgt_guid_pil = pil_img.copy()
+                    pose_guid_pil = Image.new(pil_img.mode, pil_img.size, 0)
+                    pose_guid_vid_list.append(pose_guid_pil)
+                    tgt_vidpil_lst.append(pil_img)
+                    tgt_guid_vid_list.append(tgt_guid_pil)
+
+                for index in range(ref_id, end_s):
+                    if index >= ref_id + cfg.data.generated_frames // 2 :
+                        img_dup = rgb_video[end_s]
+                        pil_img_dup = Image.fromarray(img_dup.asnumpy())
+                        tgt_guid_vid_list.append(pil_img_dup)
+                    else:
+                        img_dup = rgb_video[ref_id-1]
+                        pil_img_dup = Image.fromarray(img_dup.asnumpy())
+                        tgt_guid_vid_list.append(pil_img_dup)
+                    pose_guid_pil = Image.fromarray(kps_video[index].asnumpy())
+                    pose_guid_vid_list.append(pose_guid_pil)
+
+                    img = rgb_video[index]
+                    pil_img = Image.fromarray(img.asnumpy())
+                    tgt_vidpil_lst.append(pil_img)
+
+                for index in range(end_s, end_g):
+                    img = rgb_video[index]
+                    pil_img = Image.fromarray(img.asnumpy())
+                    tgt_guid_pil = pil_img.copy()
+                    pose_guid_pil = Image.new(pil_img.mode, pil_img.size, 0)
+                    pose_guid_vid_list.append(pose_guid_pil)
+                    tgt_vidpil_lst.append(pil_img)
+                    tgt_guid_vid_list.append(tgt_guid_pil)
+
+                seed_mask = None
+                if cfg.get("seed_input_channels", 4) == 5:
+                    w = h = cfg.data.image_size
+                    ref_img_pil = pad_to_square(ref_img_pil, w)
+                    tgt_guid_vid_list = [pad_to_square(image, w) for image in tgt_guid_vid_list]
+                    pose_guid_vid_list = [pad_to_square(image, w) for image in pose_guid_vid_list]
+                    tgt_vidpil_lst = [pad_to_square(image, w) for image in tgt_vidpil_lst]
+                    seed_mask = torch.zeros(
+                        (1, 1, cfg.data.sample_frames, h // 8, w // 8), device=accelerator.device
+                    )
+                    seed_mask[:, :, :cfg.data.seed_frames] = 1
+                    seed_mask[:, :, -cfg.data.seed_frames:] = 1
+
+                ori_w = w
+                ori_h = h
+
+                # generator = torch.manual_seed(seed)
+                if cfg.get("seed_input_channels", 4) == 5:
+                    generator = torch.Generator(device=accelerator.device).manual_seed(seed)
+                else:
+                    generator = torch.manual_seed(seed)
+                video_tensor = pipeline(
+                    ref_img_pil,
+                    pose_guid_vid_list,
+                    tgt_guid_vid_list,
+                    w,
+                    h,
+                    cfg.data.sample_frames,
+                    cfg.validation.denoising_steps,
+                    cfg.validation.guidance_scale,
+                    generator=generator,
+                    seed_mask=seed_mask,
+                ).videos
+                video_tensor = video_tensor[0, ...].permute(1, 2, 3, 0).cpu().numpy()
+
+                video_pil_lst = []
+                video_pil_lst_original_size = []
+                gap_w = (w - ori_w) // 2
+                gap_h = (h - ori_h) // 2
+                for frame_idx, image_tensor in enumerate(video_tensor):
+                    result_img_pil = Image.fromarray((image_tensor * 255).astype(np.uint8))
+                    result_img_pil_2 = result_img_pil.copy()
+                    result_img_np = np.array(result_img_pil_2)
+                    result_img_np_ori = result_img_np[gap_h:gap_h+ori_h, gap_w:gap_w+ori_w]
+                    result_img_pil_ori = Image.fromarray(result_img_np_ori)
+                    # print(result_img_pil_ori.size)
+                    if frame_idx >= cfg.data.seed_frames and frame_idx < cfg.data.seed_frames+cfg.data.generated_frames:
+                        result_pil_lst = [pose_guid_vid_list[frame_idx], tgt_guid_vid_list[frame_idx], result_img_pil, tgt_vidpil_lst[frame_idx], result_img_pil]
+                    else:
+                        result_pil_lst = [pose_guid_vid_list[frame_idx], tgt_guid_vid_list[frame_idx], result_img_pil, tgt_vidpil_lst[frame_idx], tgt_vidpil_lst[frame_idx]]
+                    concated_pil = concat_pil(result_pil_lst)
+                    video_pil_lst_original_size.append(result_img_pil_ori)
+                    video_pil_lst.append(concated_pil)
+                pil_images.append({"name": f"{test_case['video_path'][-14:-4]}", "video": video_pil_lst})
+                # clip = ImageSequenceClip(video_pil_lst_original_size, fps=24)
+                # clip.write_videofile(video_path.replace("_rgb_", "_blend_res_"), codec='libx264', bitrate="4000k")
+                output_path = video_path.replace("_rgb_", "_blend_res_")
+                if cfg.get("seed_input_channels", 4) == 5 or output_path == video_path:
+                    output_path = str(Path(cfg.output_dir) / cfg.exp_name / f"{Path(video_path).stem}_blend_res.mp4")
+                save_videos_from_pil(video_pil_lst_original_size, output_path, bitrate=4000000)
+
+    unwrap_model.train()
+    vae = vae.to(dtype=torch.float16)
+    image_enc = image_enc.to(dtype=torch.float16)
+
+    del pipeline
+    torch.cuda.empty_cache()
+    return pil_images
+
+def main(cfg):
+    os.environ.setdefault("WANDB_MODE", "disabled")
+    kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
+    accelerator = Accelerator(
+        gradient_accumulation_steps=cfg.solver.gradient_accumulation_steps,
+        mixed_precision=cfg.solver.mixed_precision,
+        log_with="wandb",
+        project_dir=cfg.wandb_log_dir,
+        kwargs_handlers=[kwargs],
+    )
+
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    logger.info(accelerator.state, main_process_only=True)
+
+    if accelerator.is_local_main_process:
+        transformers.utils.logging.set_verbosity_warning()
+        diffusers.utils.logging.set_verbosity_info()
+    else:
+        transformers.utils.logging.set_verbosity_error()
+        diffusers.utils.logging.set_verbosity_error()
+
+    if cfg.seed is not None:
+        seed_everything(cfg.seed)
+
+    sched_kwargs = OmegaConf.to_container(cfg.noise_scheduler_kwargs)
+    if cfg.enable_zero_snr:
+        sched_kwargs.update(
+            rescale_betas_zero_snr=True,
+            timestep_spacing="trailing",
+            prediction_type="v_prediction",
+        )
+    val_noise_scheduler = DDIMScheduler(**sched_kwargs)
+    sched_kwargs.update({"beta_schedule": "scaled_linear"})
+
+    vae = AutoencoderKL.from_pretrained(cfg.vae_model_path).to("cuda", dtype=torch.float32)
+    reference_unet = UNet2DConditionModel.from_pretrained(
+        cfg.base_model_path,
+        subfolder="unet",
+    ).to(device="cuda")
+    denoising_unet = UNet3DConditionModel.from_pretrained_2d(
+        cfg.base_model_path,
+        cfg.mm_path,
+        subfolder="unet",
+        unet_additional_kwargs=OmegaConf.to_container(
+            cfg.unet_additional_kwargs
+        ),
+    ).to(device="cuda")
+    image_enc = CLIPVisionModelWithProjection.from_pretrained(
+        cfg.image_encoder_path,
+    ).to(dtype=torch.float32, device="cuda")
+    pose_guider = PoseGuider(
+        conditioning_embedding_channels=320, block_out_channels=(16, 32, 96, 256),
+        conditioning_channels=3,
+    ).to(device="cuda")
+    seed_guider = ConvIn(
+        conditioning_embedding_channels=320, conditioning_channels=cfg.get("seed_input_channels", 4)
+    )
+
+    vae.requires_grad_(False)
+    image_enc.requires_grad_(False)
+    denoising_unet.requires_grad_(False)
+    pose_guider.requires_grad_(False)
+    seed_guider.requires_grad_(False)
+
+    reference_control_writer = ReferenceAttentionControl(
+        reference_unet,
+        do_classifier_free_guidance=False,
+        mode="write",
+        fusion_blocks="full",
+    )
+    reference_control_reader = ReferenceAttentionControl(
+        denoising_unet,
+        do_classifier_free_guidance=False,
+        mode="read",
+        fusion_blocks="full",
+    )
+
+    model = VFIShareVAEModel(
+        reference_unet,
+        denoising_unet,
+        seed_guider,
+        pose_guider,
+        reference_control_writer,
+        reference_control_reader,
+    )
+
+    checkpoint_path = cfg.get("checkpoint_path") or os.path.join(
+        cfg.stage1_ckpt_dir, f"checkpoint-{cfg.stage1_ckpt_step}", "pytorch_model.bin"
+    )
+    old_ckpt = torch.load(checkpoint_path, map_location="cpu", weights_only=True)
+    load_result = model.load_state_dict(old_ckpt, strict=True)
+    checkpoint_info = {
+        "path": str(checkpoint_path), "strict": True,
+        "missing_keys": load_result.missing_keys, "unexpected_keys": load_result.unexpected_keys,
+        "seed_guider_weight_shape": list(old_ckpt["seed_guider.conv_in.weight"].shape),
+    }
+    del old_ckpt
+    logger.info(f"Strictly loaded checkpoint {checkpoint_path}")
+
+    model = accelerator.prepare(
+        model
+    )
+
+    if cfg.solver.enable_xformers_memory_efficient_attention:
+        if is_xformers_available():
+            reference_unet.enable_xformers_memory_efficient_attention()
+            denoising_unet.enable_xformers_memory_efficient_attention()
+        else:
+            raise ValueError(
+                "xformers is not available. Make sure it is installed correctly"
+            )
+
+    if accelerator.is_main_process:
+        run_time = datetime.now().strftime("%Y%m%d-%H%M")
+        accelerator.init_trackers(
+            cfg.wandb_project,
+            init_kwargs={"wandb": {"name": cfg.exp_name + "_" + run_time, "dir":cfg.wandb_log_dir}},
+        )
+        config_dict = OmegaConf.to_container(cfg)
+        wandb.config.update(config_dict)
+
+    accelerator.wait_for_everyone()
+    sample_dicts = log_validation(
+        cfg=cfg,
+        vae=vae,
+        image_enc=image_enc,
+        model=model,
+        scheduler=val_noise_scheduler,
+        accelerator=accelerator,
+        width=cfg.data.image_size,
+        height=cfg.data.image_size,
+        seed=cfg.seed,
+    )
+    global_step = 0
+    image_save_dir = f"{cfg.wandb_log_dir}/{cfg.exp_name}/{global_step:06d}/"
+    if not os.path.exists(image_save_dir):
+        os.makedirs(image_save_dir, exist_ok=True)
+
+    for sample_id, sample_dict in enumerate(sample_dicts):
+        sample_name = sample_dict["name"]
+        img = sample_dict["video"]
+        if img[0].size[0] >= img[0].size[1]:
+            img_path = f"{image_save_dir}{global_step:06d}-{sample_name}-higher.mp4"
+        else:
+            img_path = f"{image_save_dir}{global_step:06d}-{sample_name}-wider.mp4"
+        save_videos_from_pil(img, img_path, fps=6)
+
+    accelerator.wait_for_everyone()
+    if accelerator.is_main_process:
+        images_to_log_higher = []
+        images_to_log_wider = []
+        for img_path in os.listdir(image_save_dir):
+            if img_path.endswith("-higher.mp4"):
+                wandb_img = wandb.Video(image_save_dir+img_path, caption=f"{global_step:06d}-{img_path}")
+                images_to_log_higher.append(wandb_img)
+            elif img_path.endswith("-wider.mp4"):
+                wandb_img = wandb.Video(image_save_dir+img_path, caption=f"{global_step:06d}-{img_path}")
+                images_to_log_wider.append(wandb_img)
+        wandb.log({
+                    "images_higher": images_to_log_higher,
+                    "images_wider": images_to_log_wider,
+                    "iteration": global_step
+                })
+    accelerator.wait_for_everyone()
+    accelerator.end_training()
+    if accelerator.is_main_process:
+        receipt_dir = Path(cfg.output_dir) / cfg.exp_name
+        receipt_dir.mkdir(parents=True, exist_ok=True)
+        receipt = {
+            "status": "complete", "checkpoint": checkpoint_info,
+            "seed": cfg.seed, "data": OmegaConf.to_container(cfg.data),
+            "validation": OmegaConf.to_container(cfg.validation),
+            "vae_type": "AutoencoderKL", "vae_model_path": cfg.vae_model_path,
+            "input_jsons": list(cfg.data.meta_paths), "samples": len(sample_dicts),
+        }
+        (receipt_dir / "inference_receipt.json").write_text(json.dumps(receipt, indent=2) + "\n")
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--config", type=str, default="diffusion/configs/inference.yaml")
+    parser.add_argument("--json", type=str, default="")
+    parser.add_argument("--checkpoint", type=str)
+    parser.add_argument("--output-dir", type=str)
+    parser.add_argument("--guidance-scale", type=float)
+    args = parser.parse_args()
+
+    if args.config[-5:] == ".yaml":
+        config = OmegaConf.load(args.config)
+    else:
+        raise ValueError("Do not support this format config file")
+
+    if args.json:
+        config.data.meta_paths = [args.json]
+
+    if args.checkpoint:
+        config.checkpoint_path = args.checkpoint
+    if args.output_dir:
+        config.output_dir = args.output_dir
+        config.wandb_log_dir = str(Path(args.output_dir) / "logs")
+    if args.guidance_scale is not None:
+        config.validation.guidance_scale = args.guidance_scale
+
+    main(config)

2610_variant_runtime/diffusion/models/attention.py

@@ -0,0 +1,526 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention.py
+
+from typing import Any, Dict, Optional
+
+import torch
+from diffusers.models.attention import AdaLayerNorm, Attention, FeedForward
+from diffusers.models.embeddings import SinusoidalPositionalEmbedding
+from einops import rearrange
+
+from torch import nn
+import math
+# import torch.onnx.operators
+# from fairseq import utils
+
+# class ScaledSinusoidalPositionalEmbedding(SinusoidalPositionalEmbedding):
+#     def __init__(self, embedding_dim, padding_idx, init_size=1024, base_fps=30):
+#         super().__init__(embedding_dim, padding_idx, init_size)
+#         self.base_fps = base_fps
+
+#     @staticmethod
+#     def get_embedding(
+#         num_embeddings: int,
+#         embedding_dim: int,
+#         padding_idx: Optional[int] = None,
+#         current_fps: Optional[float] = None,
+#         base_fps: float = 30.0
+#     ):
+#         if current_fps is None:
+#             scale_factor = 1.0
+#         else:
+#             scale_factor = current_fps / base_fps
+
+#         half_dim = embedding_dim // 2
+#         emb = math.log(10000) / (half_dim - 1)
+#         emb = torch.exp(torch.arange(half_dim, dtype=torch.float) * -emb)
+
+#         emb = torch.arange(num_embeddings, dtype=torch.float).unsqueeze(1) * emb.unsqueeze(0)
+#         emb = torch.cat([torch.sin(emb * scale_factor), torch.cos(emb * scale_factor)], dim=1)
+
+#         if embedding_dim % 2 == 1:
+#             emb = torch.cat([emb, torch.zeros(num_embeddings, 1)], dim=1)
+
+#         if padding_idx is not None:
+#             emb[padding_idx, :] = 0
+
+#         return emb
+
+#     def forward(
+#         self,
+#         input,
+#         incremental_state: Optional[Any] = None,
+#         timestep: Optional[Tensor] = None,
+#         positions: Optional[Any] = None,
+#     ):
+#         """Input is expected to be of size [bsz x seqlen]."""
+#         bspair = torch.onnx.operators.shape_as_tensor(input)
+#         bsz, seq_len = bspair[0], bspair[1]
+#         max_pos = self.padding_idx + 1 + seq_len
+#         self.weights = self.get_embedding(
+#                 max_pos, self.embedding_dim, self.padding_idx, current_fps = current_fps,
+#             )
+#         self.weights = self.weights.to(self._float_tensor)
+
+#         if incremental_state is not None:
+#             # positions is the same for every token when decoding a single step
+#             pos = timestep.view(-1)[0] + 1 if timestep is not None else seq_len
+#             if self.onnx_trace:
+#                 return (
+#                     self.weights.index_select(index=self.padding_idx + pos, dim=0)
+#                     .unsqueeze(1)
+#                     .repeat(bsz, 1, 1)
+#                 )
+#             return self.weights[self.padding_idx + pos, :].expand(bsz, 1, -1)
+
+#         positions = utils.make_positions(
+#             input, self.padding_idx, onnx_trace=self.onnx_trace
+#         )
+#         if self.onnx_trace:
+#             flat_embeddings = self.weights.detach().index_select(0, positions.view(-1))
+#             embedding_shape = torch.cat(
+#                 (bsz.view(1), seq_len.view(1), torch.tensor([-1], dtype=torch.long))
+#             )
+#             embeddings = torch.onnx.operators.reshape_from_tensor_shape(
+#                 flat_embeddings, embedding_shape
+#             )
+#             return embeddings
+#         return (
+#             self.weights.index_select(0, positions.view(-1))
+#             .view(bsz, seq_len, -1)
+#             .detach()
+#         )
+
+class BasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm (:
+            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        only_cross_attention (`bool`, *optional*):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, *optional*):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        upcast_attention (`bool`, *optional*):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
+        final_dropout (`bool` *optional*, defaults to False):
+            Whether to apply a final dropout after the last feed-forward layer.
+        attention_type (`str`, *optional*, defaults to `"default"`):
+            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
+        positional_embeddings (`str`, *optional*, defaults to `None`):
+            The type of positional embeddings to apply to.
+        num_positional_embeddings (`int`, *optional*, defaults to `None`):
+            The maximum number of positional embeddings to apply.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        attention_type: str = "default",
+        positional_embeddings: Optional[str] = None,
+        num_positional_embeddings: Optional[int] = None,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+
+        self.use_ada_layer_norm_zero = (
+            num_embeds_ada_norm is not None
+        ) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (
+            num_embeds_ada_norm is not None
+        ) and norm_type == "ada_norm"
+        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
+        self.use_layer_norm = norm_type == "layer_norm"
+
+        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+            )
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(
+                dim, max_seq_length=num_positional_embeddings
+            )
+        else:
+            self.pos_embed = None
+
+        # print(self.pos_embed, positional_embeddings)
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        if self.use_ada_layer_norm:
+            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        elif self.use_ada_layer_norm_zero:
+            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+        else:
+            self.norm1 = nn.LayerNorm(
+                dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
+            )
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+            upcast_attention=upcast_attention,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(
+                    dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
+                )
+            )
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=(
+                    cross_attention_dim if not double_self_attention else None
+                ),
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        if not self.use_ada_layer_norm_single:
+            self.norm3 = nn.LayerNorm(
+                dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
+            )
+
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+        )
+
+        # 4. Fuser
+        if attention_type == "gated" or attention_type == "gated-text-image":
+            self.fuser = GatedSelfAttentionDense(
+                dim, cross_attention_dim, num_attention_heads, attention_head_dim
+            )
+
+        # 5. Scale-shift for PixArt-Alpha.
+        if self.use_ada_layer_norm_single:
+            self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+    ) -> torch.FloatTensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        if self.use_ada_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states, timestep)
+        elif self.use_ada_layer_norm_zero:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+        elif self.use_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states)
+        elif self.use_ada_layer_norm_single:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
+            ).chunk(6, dim=1)
+            norm_hidden_states = self.norm1(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+            norm_hidden_states = norm_hidden_states.squeeze(1)
+        else:
+            raise ValueError("Incorrect norm used")
+
+        if self.pos_embed is not None:
+            norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+        # 1. Retrieve lora scale.
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+
+        # 2. Prepare GLIGEN inputs
+        cross_attention_kwargs = (
+            cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+        )
+        gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=(
+                encoder_hidden_states if self.only_cross_attention else None
+            ),
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+        if self.use_ada_layer_norm_zero:
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        elif self.use_ada_layer_norm_single:
+            attn_output = gate_msa * attn_output
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 2.5 GLIGEN Control
+        if gligen_kwargs is not None:
+            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            if self.use_ada_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states, timestep)
+            elif self.use_ada_layer_norm_zero or self.use_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states)
+            elif self.use_ada_layer_norm_single:
+                # For PixArt norm2 isn't applied here:
+                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+                norm_hidden_states = hidden_states
+            else:
+                raise ValueError("Incorrect norm")
+
+            if self.pos_embed is not None and self.use_ada_layer_norm_single is False:
+                norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+            attn_output = self.attn2(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                **cross_attention_kwargs,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        if not self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm3(hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            norm_hidden_states = (
+                norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+            )
+
+        if self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm2(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        ff_output = self.ff(norm_hidden_states, scale=lora_scale)
+
+        if self.use_ada_layer_norm_zero:
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+        elif self.use_ada_layer_norm_single:
+            ff_output = gate_mlp * ff_output
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
+class TemporalBasicTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+        self.use_ada_layer_norm = num_embeds_ada_norm is not None
+        self.unet_use_cross_frame_attention = unet_use_cross_frame_attention
+        self.unet_use_temporal_attention = unet_use_temporal_attention
+
+        # SC-Attn
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            upcast_attention=upcast_attention,
+        )
+        self.norm1 = (
+            AdaLayerNorm(dim, num_embeds_ada_norm)
+            if self.use_ada_layer_norm
+            else nn.LayerNorm(dim)
+        )
+
+        # Cross-Attn
+        if cross_attention_dim is not None:
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )
+        else:
+            self.attn2 = None
+
+        if cross_attention_dim is not None:
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim)
+            )
+        else:
+            self.norm2 = None
+
+        # Feed-forward
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
+        self.norm3 = nn.LayerNorm(dim)
+        self.use_ada_layer_norm_zero = False
+
+        # Temp-Attn
+        assert unet_use_temporal_attention is not None
+        if unet_use_temporal_attention:
+            self.attn_temp = Attention(
+                query_dim=dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )
+            nn.init.zeros_(self.attn_temp.to_out[0].weight.data)
+            self.norm_temp = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim)
+            )
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        timestep=None,
+        attention_mask=None,
+        video_length=None,
+    ):
+        norm_hidden_states = (
+            self.norm1(hidden_states, timestep)
+            if self.use_ada_layer_norm
+            else self.norm1(hidden_states)
+        )
+        print(hidden_states.shape, video_length, "tem layer")
+        if self.unet_use_cross_frame_attention:
+            hidden_states = (
+                self.attn1(
+                    norm_hidden_states,
+                    attention_mask=attention_mask,
+                    video_length=video_length,
+                )
+                + hidden_states
+            )
+        else:
+            hidden_states = (
+                self.attn1(norm_hidden_states, attention_mask=attention_mask)
+                + hidden_states
+            )
+
+        if self.attn2 is not None:
+            # Cross-Attention
+            norm_hidden_states = (
+                self.norm2(hidden_states, timestep)
+                if self.use_ada_layer_norm
+                else self.norm2(hidden_states)
+            )
+            hidden_states = (
+                self.attn2(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                )
+                + hidden_states
+            )
+
+        # Feed-forward
+        hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
+
+        # Temporal-Attention
+        if self.unet_use_temporal_attention:
+            d = hidden_states.shape[1]
+            hidden_states = rearrange(
+                hidden_states, "(b f) d c -> (b d) f c", f=video_length
+            )
+            norm_hidden_states = (
+                self.norm_temp(hidden_states, timestep)
+                if self.use_ada_layer_norm
+                else self.norm_temp(hidden_states)
+            )
+            hidden_states = self.attn_temp(norm_hidden_states) + hidden_states
+            hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)
+
+        return hidden_states

2610_variant_runtime/diffusion/models/conv_in.py

@@ -0,0 +1,26 @@
+from typing import Tuple
+
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.nn.init as init
+from diffusers.models.modeling_utils import ModelMixin
+
+from models.motion_module import zero_module
+from models.resnet import InflatedConv3d
+
+
+class ConvIn(ModelMixin):
+    def __init__(
+        self,
+        conditioning_embedding_channels: int,
+        conditioning_channels: int = 3,
+        # block_out_channels: Tuple[int] = (16, 32, 64, 128),
+    ):
+        super().__init__()
+        self.conv_in = InflatedConv3d(
+            conditioning_channels, conditioning_embedding_channels, kernel_size=3, padding=1
+        )
+
+    def forward(self, conditioning):
+        embedding = self.conv_in(conditioning)
+        return embedding

2610_variant_runtime/diffusion/models/model_vfi.py

@@ -0,0 +1,987 @@
+import torch
+import torch.nn as nn
+from models.unet_2d_condition import UNet2DConditionModel
+from models.unet_3d import UNet3DConditionModel
+from models.pose_guider import PoseGuider
+
+class VFIPoseModel(nn.Module):
+    def __init__(
+        self,
+        reference_unet: UNet2DConditionModel,
+        denoising_unet: UNet3DConditionModel,
+        seed_guider: PoseGuider,
+        pose_guider: PoseGuider,
+        reference_control_writer,
+        reference_control_reader,
+    ):
+        super().__init__()
+        self.reference_unet = reference_unet
+        self.denoising_unet = denoising_unet
+        self.pose_guider = pose_guider
+        self.seed_guider = seed_guider
+        self.reference_control_writer = reference_control_writer
+        self.reference_control_reader = reference_control_reader
+
+    def forward(
+        self,
+        noisy_latents,
+        timesteps,
+        ref_image_latents,
+        clip_image_embeds,
+        pose_img,
+        seed_img,
+        uncond_fwd: bool = False,
+    ):
+        seed_cond_tensor = seed_img.to(device="cuda")
+        seed_fea = self.seed_guider(seed_cond_tensor)
+
+        pose_cond_tensor = pose_img.to(device="cuda")
+        pose_fea = self.pose_guider(pose_cond_tensor)
+
+        if not uncond_fwd:
+            ref_timesteps = torch.zeros_like(timesteps)
+            # bs, c, f h, w = ref_image_latents.shape
+            # ref_image_latents = ref_image_latents.reshape()
+            self.reference_unet(
+                ref_image_latents,
+                ref_timesteps,
+                encoder_hidden_states=clip_image_embeds,
+                return_dict=False,
+            )
+            self.reference_control_reader.update(self.reference_control_writer)
+
+        pose_fea = seed_fea + pose_fea
+
+        model_pred = self.denoising_unet(
+            noisy_latents,
+            timesteps,
+            guidance_fea=pose_fea,
+            encoder_hidden_states=clip_image_embeds,
+        ).sample
+
+        return model_pred
+
+
+
+class VFIShareVAEModel(nn.Module):
+    def __init__(
+        self,
+        reference_unet: UNet2DConditionModel,
+        denoising_unet: UNet3DConditionModel,
+        seed_guider,
+        pose_guider: PoseGuider,
+        reference_control_writer,
+        reference_control_reader,
+    ):
+        super().__init__()
+        self.reference_unet = reference_unet
+        self.denoising_unet = denoising_unet
+        self.pose_guider = pose_guider
+        self.seed_guider = seed_guider
+        self.reference_control_writer = reference_control_writer
+        self.reference_control_reader = reference_control_reader
+
+    def forward(
+        self,
+        noisy_latents,
+        timesteps,
+        ref_image_latents,
+        clip_image_embeds,
+        seed_img = None,
+        pose_img = None,
+        uncond_fwd: bool = False,
+    ):
+        # print(seed_img.shape)
+        # print(pose_img.shape, seed_img.shape)
+        if seed_img is not None:
+            seed_cond_tensor = seed_img.to(device="cuda")
+            seed_fea = self.seed_guider(seed_cond_tensor)
+        else:
+            seed_fea = None
+
+        if pose_img is not None:
+            pose_cond_tensor = pose_img.to(device="cuda")
+            pose_fea = self.pose_guider(pose_cond_tensor)
+            if seed_fea is None:
+                pose_fea = pose_fea
+            else:
+                pose_fea = seed_fea + pose_fea
+        else:
+            pose_fea = seed_fea
+
+        if True:
+            ref_timesteps = torch.zeros_like(timesteps)
+            self.reference_unet(
+                ref_image_latents,
+                ref_timesteps,
+                encoder_hidden_states=clip_image_embeds,
+                return_dict=False,
+            )
+            self.reference_control_reader.update(self.reference_control_writer)
+
+        model_pred = self.denoising_unet(
+            noisy_latents,
+            timesteps,
+            guidance_fea=pose_fea,
+            encoder_hidden_states=clip_image_embeds,
+        ).sample
+
+        return model_pred
+
+
+class VFIShareVAEModel_s1(nn.Module):
+    def __init__(
+        self,
+        reference_unet: UNet2DConditionModel,
+        denoising_unet: UNet3DConditionModel,
+        seed_guider,
+        pose_guider: PoseGuider,
+        reference_control_writer,
+        reference_control_reader,
+    ):
+        super().__init__()
+        self.reference_unet = reference_unet
+        self.denoising_unet = denoising_unet
+        self.pose_guider = pose_guider
+        self.seed_guider = seed_guider
+        self.reference_control_writer = reference_control_writer
+        self.reference_control_reader = reference_control_reader
+
+    def forward(
+        self,
+        noisy_latents,
+        timesteps,
+        ref_image_latents,
+        clip_image_embeds,
+        seed_img = None,
+        pose_img = None,
+        uncond_fwd: bool = False,
+    ):
+        # print(seed_img.shape)
+        # print(pose_img.shape, seed_img.shape)
+        if seed_img is not None:
+            seed_cond_tensor = seed_img.to(device="cuda")
+            seed_fea = self.seed_guider(seed_cond_tensor)
+        else:
+            seed_fea = None
+
+        if pose_img is not None:
+            pose_cond_tensor = pose_img.to(device="cuda")
+            pose_fea = self.pose_guider(pose_cond_tensor)
+            if seed_fea is None:
+                pose_fea = pose_fea
+            else:
+                pose_fea = seed_fea + pose_fea
+        else:
+            pose_fea = seed_fea
+
+        if not uncond_fwd:
+            ref_timesteps = torch.zeros_like(timesteps)
+            self.reference_unet(
+                ref_image_latents,
+                ref_timesteps,
+                encoder_hidden_states=clip_image_embeds,
+                return_dict=False,
+            )
+            self.reference_control_reader.update(self.reference_control_writer)
+
+        model_pred = self.denoising_unet(
+            noisy_latents,
+            timesteps,
+            guidance_fea=pose_fea,
+            encoder_hidden_states=clip_image_embeds,
+        ).sample
+
+        return model_pred
+
+
+class VFINoCFGModel(nn.Module):
+    def __init__(
+        self,
+        reference_unet: UNet2DConditionModel,
+        denoising_unet: UNet3DConditionModel,
+        seed_guider,
+        pose_guider: PoseGuider,
+        reference_control_writer,
+        reference_control_reader,
+    ):
+        super().__init__()
+        self.reference_unet = reference_unet
+        self.denoising_unet = denoising_unet
+        self.pose_guider = pose_guider
+        self.seed_guider = seed_guider
+        self.reference_control_writer = reference_control_writer
+        self.reference_control_reader = reference_control_reader
+
+    def forward(
+        self,
+        noisy_latents,
+        timesteps,
+        ref_image_latents,
+        clip_image_embeds,
+        seed_img = None,
+        pose_img = None,
+        uncond_fwd: bool = False,
+    ):
+        # print(seed_img.shape)
+        # print(pose_img.shape, seed_img.shape)
+        if seed_img is not None:
+            seed_cond_tensor = seed_img.to(device="cuda")
+            seed_fea = self.seed_guider(seed_cond_tensor)
+        else:
+            seed_fea = None
+
+        if pose_img is not None:
+            pose_cond_tensor = pose_img.to(device="cuda")
+            pose_fea = self.pose_guider(pose_cond_tensor)
+            if seed_fea is None:
+                pose_fea = pose_fea
+            else:
+                pose_fea = seed_fea + pose_fea
+        else:
+            pose_fea = seed_fea
+
+        if not uncond_fwd:
+            ref_timesteps = torch.zeros_like(timesteps)
+            self.reference_unet(
+                ref_image_latents,
+                ref_timesteps,
+                encoder_hidden_states=clip_image_embeds,
+                return_dict=False,
+            )
+            self.reference_control_reader.update(self.reference_control_writer)
+
+        model_pred = self.denoising_unet(
+            noisy_latents,
+            timesteps,
+            guidance_fea=pose_fea,
+            encoder_hidden_states=clip_image_embeds,
+        ).sample
+
+        return model_pred
+
+
+class VFINoRefModel(nn.Module):
+    def __init__(
+        self,
+        reference_unet: UNet2DConditionModel,
+        denoising_unet: UNet3DConditionModel,
+        seed_guider,
+        pose_guider: PoseGuider,
+        reference_control_writer,
+        reference_control_reader,
+    ):
+        super().__init__()
+        self.reference_unet = reference_unet
+        self.denoising_unet = denoising_unet
+        self.pose_guider = pose_guider
+        self.seed_guider = seed_guider
+        self.reference_control_writer = reference_control_writer
+        self.reference_control_reader = reference_control_reader
+
+    def forward(
+        self,
+        noisy_latents,
+        timesteps,
+        ref_image_latents,
+        clip_image_embeds,
+        seed_img = None,
+        pose_img = None,
+        uncond_fwd: bool = False,
+    ):
+        # print(seed_img.shape)
+        if seed_img is not None:
+            seed_cond_tensor = seed_img.to(device="cuda")
+            seed_fea = self.seed_guider(seed_cond_tensor)
+        else:
+            seed_fea = None
+
+        if pose_img is not None:
+            pose_cond_tensor = pose_img.to(device="cuda")
+            pose_fea = self.pose_guider(pose_cond_tensor)
+            if seed_fea is None:
+                pose_fea = pose_fea
+            else:
+                pose_fea = seed_fea + pose_fea
+        else:
+            pose_fea = seed_fea
+
+        model_pred = self.denoising_unet(
+            noisy_latents,
+            timesteps,
+            guidance_fea=pose_fea,
+            encoder_hidden_states=clip_image_embeds,
+        ).sample
+
+        return model_pred
+
+
+# v1-dup
+# reference image are cat to noise, re-train convin
+# v2-dup
+# reference add to, 2 convin
+# v1-blank
+# leave middle padding
+# v2-blank
+# all use 256, batch size 4
+
+# v3-add, flexiable reference
+class VFICatModel(nn.Module):
+    def __init__(
+        self,
+        reference_unet: UNet2DConditionModel,
+        denoising_unet: UNet3DConditionModel,
+        pose_guider: PoseGuider,
+        reference_control_writer,
+        reference_control_reader,
+    ):
+        super().__init__()
+        self.reference_unet = reference_unet
+        self.denoising_unet = denoising_unet
+        self.pose_guider = pose_guider
+        self.reference_control_writer = reference_control_writer
+        self.reference_control_reader = reference_control_reader
+
+    def forward(
+        self,
+        noisy_latents,
+        timesteps,
+        ref_image_latents,
+        clip_image_embeds,
+        seed_img = None,
+        pose_img = None,
+        uncond_fwd: bool = False,
+    ):
+        if seed_img is not None:
+            seed_cond_tensor = seed_img.to(device="cuda")
+            noisy_latents = torch.cat([noisy_latents, seed_cond_tensor], dim=1)
+        else:
+            noisy_latents = torch.cat([noisy_latents, torch.zeros_like(noisy_latents).to(device="cuda")], dim=1)
+
+        if pose_img is not None:
+            pose_cond_tensor = pose_img.to(device="cuda")
+            pose_fea = self.pose_guider(pose_cond_tensor)
+        else:
+            pose_fea = None
+        if not uncond_fwd:
+            ref_timesteps = torch.zeros_like(timesteps)
+            self.reference_unet(
+                ref_image_latents,
+                ref_timesteps,
+                encoder_hidden_states=clip_image_embeds,
+                return_dict=False,
+            )
+            self.reference_control_reader.update(self.reference_control_writer)
+
+        model_pred = self.denoising_unet(
+            noisy_latents,
+            timesteps,
+            guidance_fea=pose_fea,
+            encoder_hidden_states=clip_image_embeds,
+        ).sample
+        return model_pred
+
+from einops import rearrange
+class VFIMultiRefModel(nn.Module):
+    def __init__(
+        self,
+        reference_unet: UNet2DConditionModel,
+        denoising_unet: UNet3DConditionModel,
+        seed_guider,
+        pose_guider: PoseGuider,
+        reference_control_writer,
+        reference_control_reader,
+    ):
+        super().__init__()
+        self.reference_unet = reference_unet
+        self.denoising_unet = denoising_unet
+        self.pose_guider = pose_guider
+        self.seed_guider = seed_guider
+        self.reference_control_writer = reference_control_writer
+        self.reference_control_reader = reference_control_reader
+
+    def forward(
+        self,
+        noisy_latents,
+        timesteps,
+        ref_image_latents,
+        clip_image_embeds,
+        seed_img = None,
+        pose_img = None,
+        uncond_fwd: bool = False,
+    ):
+        if seed_img is not None:
+            seed_cond_tensor = seed_img.to(device="cuda")
+            seed_fea = self.seed_guider(seed_cond_tensor)
+        else:
+            seed_fea = None
+
+        if pose_img is not None:
+            pose_cond_tensor = pose_img.to(device="cuda")
+            pose_fea = self.pose_guider(pose_cond_tensor)
+            if seed_fea is None:
+                pose_fea = pose_fea
+            else:
+                pose_fea = seed_fea + pose_fea
+        else:
+            pose_fea = seed_fea
+
+        if not uncond_fwd:
+            # print(timesteps.shape)
+            ref_image_latents = rearrange(
+                ref_image_latents, "b c f h w -> (b f) c h w"
+            )
+            ref_timesteps = torch.zeros(ref_image_latents.shape[0]).to(device="cuda")
+            # print(ref_timesteps.shape, ref_image_latents.shape, clip_image_embeds.shape)
+            self.reference_unet(
+                ref_image_latents,
+                ref_timesteps,
+                encoder_hidden_states=clip_image_embeds,
+                return_dict=False,
+            )
+            self.reference_control_reader.update(self.reference_control_writer)
+
+        model_pred = self.denoising_unet(
+            noisy_latents,
+            timesteps,
+            guidance_fea=pose_fea,
+            encoder_hidden_states=clip_image_embeds,
+        ).sample
+
+        return model_pred
+
+from einops import rearrange
+class VFIMultiRefModel_v4_dup(nn.Module):
+    def __init__(
+        self,
+        reference_unet: UNet2DConditionModel,
+        denoising_unet: UNet3DConditionModel,
+        seed_guider,
+        pose_guider: PoseGuider,
+        reference_control_writer,
+        reference_control_reader,
+    ):
+        super().__init__()
+        self.reference_unet = reference_unet
+        self.denoising_unet = denoising_unet
+        self.pose_guider = pose_guider
+        self.seed_guider = seed_guider
+        self.reference_control_writer = reference_control_writer
+        self.reference_control_reader = reference_control_reader
+
+    def forward(
+        self,
+        noisy_latents,
+        timesteps,
+        ref_image_latents,
+        clip_image_embeds,
+        seed_img = None,
+        pose_img = None,
+        uncond_fwd: bool = False,
+    ):
+        if seed_img is not None:
+            seed_cond_tensor = seed_img.to(device="cuda")
+            seed_fea = self.seed_guider(seed_cond_tensor)
+        else:
+            seed_fea = None
+
+        if pose_img is not None:
+            pose_cond_tensor = pose_img.to(device="cuda")
+            pose_fea = self.pose_guider(pose_cond_tensor)
+            if seed_fea is None:
+                pose_fea = pose_fea
+            else:
+                pose_fea = seed_fea + pose_fea
+        else:
+            pose_fea = seed_fea
+
+        if not uncond_fwd:
+            # print(ref_image_latents.shape)
+            ref_image_latents = rearrange(
+                ref_image_latents, "b c f h w -> (b f) c h w"
+            )
+            clip_image_embeds_a = clip_image_embeds.repeat(noisy_latents.shape[2], 1, 1)
+            # print(clip_image_embeds.shape, clip_image_embeds_a.shape)
+            ref_timesteps = torch.zeros(ref_image_latents.shape[0]).to(device="cuda")
+            self.reference_unet(
+                ref_image_latents,
+                ref_timesteps,
+                encoder_hidden_states=clip_image_embeds_a,
+                return_dict=False,
+            )
+            self.reference_control_reader.update(self.reference_control_writer)
+
+        model_pred = self.denoising_unet(
+            noisy_latents,
+            timesteps,
+            guidance_fea=pose_fea,
+            encoder_hidden_states=clip_image_embeds,
+        ).sample
+
+        return model_pred
+
+class VFIMultiRefModel_v5_dup(nn.Module):
+    def __init__(
+        self,
+        reference_unet: UNet2DConditionModel,
+        denoising_unet: UNet3DConditionModel,
+        seed_guider,
+        pose_guider: PoseGuider,
+        reference_control_writer,
+        reference_control_reader,
+    ):
+        super().__init__()
+        self.reference_unet = reference_unet
+        self.denoising_unet = denoising_unet
+        self.pose_guider = pose_guider
+        self.seed_guider = seed_guider
+        self.reference_control_writer = reference_control_writer
+        self.reference_control_reader = reference_control_reader
+
+    def forward(
+        self,
+        noisy_latents,
+        timesteps,
+        ref_image_latents,
+        clip_image_embeds,
+        seed_img = None,
+        pose_img = None,
+        uncond_fwd: bool = False,
+    ):
+        if pose_img is not None:
+            pose_cond_tensor = pose_img.to(device="cuda")
+            pose_fea = self.pose_guider(pose_cond_tensor)
+        else:
+            pose_fea = None
+
+        # clip_image_embeds_a = rearrange(
+        #         clip_image_embeds, "(b a) c f -> b a c f", a=noisy_latents.shape[2],
+        #     ) # 2 14 1 768
+        # clip_image_embeds_a = clip_image_embeds_a[:, 2]
+        # clip_image_embeds_b = clip_image_embeds_a.repeat(1, noisy_latents.shape[2], 1, 1)
+        # clip_image_embeds_b = rearrange(
+        #         clip_image_embeds_b, "b a c f -> (b a) c f",
+        #     )
+
+        if not uncond_fwd:
+            ref_image_latents = rearrange(
+                ref_image_latents, "b c f h w -> (b f) c h w"
+            )
+            ref_timesteps = torch.zeros(ref_image_latents.shape[0]).to(device="cuda")
+            self.reference_unet(
+                ref_image_latents,
+                ref_timesteps,
+                encoder_hidden_states=clip_image_embeds,
+                return_dict=False,
+            )
+            self.reference_control_reader.update(self.reference_control_writer)
+
+        model_pred = self.denoising_unet(
+            noisy_latents,
+            timesteps,
+            guidance_fea=pose_fea,
+            encoder_hidden_states=clip_image_embeds,
+        ).sample
+
+        return model_pred
+
+class VFIMultiRefModelv2(nn.Module):
+    def __init__(
+        self,
+        reference_unet: UNet2DConditionModel,
+        denoising_unet: UNet3DConditionModel,
+        seed_guider,
+        pose_guider: PoseGuider,
+        reference_control_writer,
+        reference_control_reader,
+    ):
+        super().__init__()
+        self.reference_unet = reference_unet
+        self.denoising_unet = denoising_unet
+        self.pose_guider = pose_guider
+        self.seed_guider = seed_guider
+        self.reference_control_writer = reference_control_writer
+        self.reference_control_reader = reference_control_reader
+
+    def forward(
+        self,
+        noisy_latents,
+        timesteps,
+        ref_image_latents,
+        clip_image_embeds,
+        seed_img = None,
+        pose_img = None,
+        uncond_fwd: bool = False,
+    ):
+        if pose_img is not None:
+            pose_cond_tensor = pose_img.to(device="cuda")
+            pose_fea = self.pose_guider(pose_cond_tensor)
+        else:
+            pose_fea = None
+
+        if not uncond_fwd:
+            # print(timesteps.shape)
+            ref_image_latents = rearrange(
+                ref_image_latents, "b c f h w -> (b f) c h w"
+            )
+            ref_timesteps = torch.zeros(ref_image_latents.shape[0]).to(device="cuda")
+            print(ref_timesteps.shape, ref_image_latents.shape, clip_image_embeds.shape)
+            self.reference_unet(
+                ref_image_latents,
+                ref_timesteps,
+                encoder_hidden_states=clip_image_embeds,
+                return_dict=False,
+            )
+            self.reference_control_reader.update(self.reference_control_writer)
+
+        model_pred = self.denoising_unet(
+            noisy_latents,
+            timesteps,
+            guidance_fea=pose_fea,
+            encoder_hidden_states=clip_image_embeds,
+        ).sample
+
+        return model_pred
+
+class VFIMultiRefModel_v8(nn.Module):
+    def __init__(
+        self,
+        reference_unet: UNet2DConditionModel,
+        denoising_unet: UNet3DConditionModel,
+        seed_guider: PoseGuider,
+        pose_guider: PoseGuider,
+        reference_control_writer,
+        reference_control_reader,
+    ):
+        super().__init__()
+        self.reference_unet = reference_unet
+        self.denoising_unet = denoising_unet
+        self.pose_guider = pose_guider
+        self.seed_guider = seed_guider
+        self.reference_control_writer = reference_control_writer
+        self.reference_control_reader = reference_control_reader
+
+    def forward(
+        self,
+        noisy_latents,
+        timesteps,
+        ref_image_latents,
+        clip_image_embeds,
+        seed_img = None,
+        pose_img = None,
+        uncond_fwd: bool = False,
+    ):
+        if seed_img is not None:
+            seed_cond_tensor = seed_img.to(device="cuda")
+            seed_fea = self.seed_guider(seed_cond_tensor)
+        else:
+            seed_fea = None
+
+        if pose_img is not None:
+            pose_cond_tensor = pose_img.to(device="cuda")
+            pose_fea = self.pose_guider(pose_cond_tensor)
+            if seed_fea is None:
+                pose_fea = pose_fea
+            else:
+                pose_fea = seed_fea + pose_fea
+        else:
+            pose_fea = seed_fea
+
+        # random pick two
+        idx_1 = 2
+        idx_2 = 11
+        ref_image_latents_1 = ref_image_latents[:, :, idx_1]
+        ref_image_latents_2 = ref_image_latents[:, :, idx_2]
+
+        length = noisy_latents.shape[2]
+        confidence_zero_to_one = torch.linspace(0, 1, steps=length).to(device="cuda")
+        confidence_one_to_zero = torch.linspace(1, 0, steps=length).to(device="cuda")
+        confidence_1 = torch.cat([confidence_zero_to_one[-idx_1-1:],confidence_one_to_zero[1:-idx_1]])
+        confidence_2 = torch.cat([confidence_zero_to_one[-idx_2-1:],confidence_one_to_zero[1:-idx_2]])
+
+        if True: #not uncond_fwd:
+            ref_timesteps = torch.zeros_like(timesteps)
+            self.reference_unet(
+                ref_image_latents_1,
+                ref_timesteps,
+                encoder_hidden_states=clip_image_embeds,
+                return_dict=False,
+            )
+            self.reference_control_reader.update(self.reference_control_writer, confidence_1)
+
+            self.reference_unet(
+                ref_image_latents_2,
+                ref_timesteps,
+                encoder_hidden_states=clip_image_embeds,
+                return_dict=False,
+            )
+            self.reference_control_reader.update(self.reference_control_writer, confidence_2)
+
+        model_pred = self.denoising_unet(
+            noisy_latents,
+            timesteps,
+            guidance_fea=pose_fea,
+            encoder_hidden_states=clip_image_embeds,
+        ).sample
+
+        return model_pred
+
+
+class VFIMultiRefModel_16frames(nn.Module):
+    def __init__(
+        self,
+        reference_unet: UNet2DConditionModel,
+        denoising_unet: UNet3DConditionModel,
+        seed_guider: PoseGuider,
+        pose_guider: PoseGuider,
+        reference_control_writer,
+        reference_control_reader,
+    ):
+        super().__init__()
+        self.reference_unet = reference_unet
+        self.denoising_unet = denoising_unet
+        self.pose_guider = pose_guider
+        self.seed_guider = seed_guider
+        self.reference_control_writer = reference_control_writer
+        self.reference_control_reader = reference_control_reader
+
+    def forward(
+        self,
+        noisy_latents,
+        timesteps,
+        ref_image_latents,
+        clip_image_embeds,
+        seed_img = None,
+        pose_img = None,
+        uncond_fwd: bool = False,
+    ):
+        if seed_img is not None:
+            seed_cond_tensor = seed_img.to(device="cuda")
+            seed_fea = self.seed_guider(seed_cond_tensor)
+        else:
+            seed_fea = None
+
+        if pose_img is not None:
+            pose_cond_tensor = pose_img.to(device="cuda")
+            pose_fea = self.pose_guider(pose_cond_tensor)
+            if seed_fea is None:
+                pose_fea = pose_fea
+            else:
+                pose_fea = seed_fea + pose_fea
+        else:
+            pose_fea = seed_fea
+
+        # random pick two
+        idx_1 = 1
+        idx_2 = -1
+        ref_image_latents_1 = ref_image_latents[:, :, idx_1]
+        ref_image_latents_2 = ref_image_latents[:, :, idx_2]
+
+        length = noisy_latents.shape[2]
+        confidence_zero_to_one = torch.linspace(0, 1, steps=length).to(device="cuda")
+        confidence_one_to_zero = torch.linspace(1, 0, steps=length).to(device="cuda")
+        confidence_1 = confidence_one_to_zero
+        confidence_2 = confidence_zero_to_one
+
+        if True: #not uncond_fwd:
+            ref_timesteps = torch.zeros_like(timesteps)
+            self.reference_unet(
+                ref_image_latents_1,
+                ref_timesteps,
+                encoder_hidden_states=clip_image_embeds,
+                return_dict=False,
+            )
+            self.reference_control_reader.update(self.reference_control_writer, confidence_1)
+
+            self.reference_unet(
+                ref_image_latents_2,
+                ref_timesteps,
+                encoder_hidden_states=clip_image_embeds,
+                return_dict=False,
+            )
+            self.reference_control_reader.update(self.reference_control_writer, confidence_2)
+
+        model_pred = self.denoising_unet(
+            noisy_latents,
+            timesteps,
+            guidance_fea=pose_fea,
+            encoder_hidden_states=clip_image_embeds,
+        ).sample
+
+        return model_pred
+
+
+
+class VFIMultiRefModel_8Time(nn.Module):
+    def __init__(
+        self,
+        reference_unet: UNet2DConditionModel,
+        denoising_unet: UNet3DConditionModel,
+        seed_guider: PoseGuider,
+        pose_guider: PoseGuider,
+        reference_control_writer,
+        reference_control_reader,
+    ):
+        super().__init__()
+        self.reference_unet = reference_unet
+        self.denoising_unet = denoising_unet
+        self.pose_guider = pose_guider
+        self.seed_guider = seed_guider
+        self.reference_control_writer = reference_control_writer
+        self.reference_control_reader = reference_control_reader
+
+    def forward(
+        self,
+        noisy_latents,
+        timesteps,
+        ref_image_latents,
+        clip_image_embeds,
+        seed_img = None,
+        pose_img = None,
+        uncond_fwd: bool = False,
+    ):
+        if seed_img is not None:
+            seed_cond_tensor = seed_img.to(device="cuda")
+            seed_fea = self.seed_guider(seed_cond_tensor)
+        else:
+            seed_fea = None
+
+        if pose_img is not None:
+            pose_cond_tensor = pose_img.to(device="cuda")
+            pose_fea = self.pose_guider(pose_cond_tensor)
+            if seed_fea is None:
+                pose_fea = pose_fea
+            else:
+                pose_fea = seed_fea + pose_fea
+        else:
+            pose_fea = seed_fea
+
+
+        length = ref_image_latents.shape[2]
+        confidence_zero_to_one = torch.linspace(0, 1, steps=length).to(
+            dtype=ref_image_latents.dtype, device=ref_image_latents.device
+        )
+        confidence_one_to_zero = torch.linspace(1, 0, steps=length).to(
+            dtype=ref_image_latents.dtype, device=ref_image_latents.device
+        )
+        confidence_lst = []
+        confidence_lst.append(confidence_one_to_zero.to(
+            dtype=ref_image_latents.dtype, device=ref_image_latents.device
+        ))
+        for idx_1 in range(1,length-1):
+            confidence_lst.append(torch.cat([confidence_zero_to_one[-idx_1-1:],confidence_one_to_zero[1:-idx_1]]).to(
+            dtype=ref_image_latents.dtype, device=ref_image_latents.device
+        ))
+        confidence_lst.append(confidence_one_to_zero.to(
+            dtype=ref_image_latents.dtype, device=ref_image_latents.device
+        ))
+        # confidence_lst = [
+        #     x.view(1,length,1,1) for x in confidence_lst
+        # ]
+
+        if True: #not uncond_fwd:
+            ref_timesteps = torch.zeros_like(timesteps)
+            for ref_id in [0, 1, 2, 3, 12, 13, 14, 15]:
+                self.reference_unet(
+                    ref_image_latents[:, :, ref_id],
+                    ref_timesteps,
+                    encoder_hidden_states=clip_image_embeds,
+                    return_dict=False,
+                )
+                self.reference_control_reader.update(self.reference_control_writer, confidence_lst[ref_id])
+
+        model_pred = self.denoising_unet(
+            noisy_latents,
+            timesteps,
+            guidance_fea=pose_fea,
+            encoder_hidden_states=clip_image_embeds,
+        ).sample
+
+        return model_pred
+
+class VFIMultiRefModel_1Time(nn.Module):
+    def __init__(
+        self,
+        reference_unet: UNet2DConditionModel,
+        denoising_unet: UNet3DConditionModel,
+        seed_guider: PoseGuider,
+        pose_guider: PoseGuider,
+        reference_control_writer,
+        reference_control_reader,
+    ):
+        super().__init__()
+        self.reference_unet = reference_unet
+        self.denoising_unet = denoising_unet
+        self.pose_guider = pose_guider
+        self.seed_guider = seed_guider
+        self.reference_control_writer = reference_control_writer
+        self.reference_control_reader = reference_control_reader
+
+    def forward(
+        self,
+        noisy_latents,
+        timesteps,
+        ref_image_latents,
+        clip_image_embeds,
+        seed_img = None,
+        pose_img = None,
+        uncond_fwd: bool = False,
+    ):
+        if seed_img is not None:
+            seed_cond_tensor = seed_img.to(device="cuda")
+            seed_fea = self.seed_guider(seed_cond_tensor)
+        else:
+            seed_fea = None
+
+        if pose_img is not None:
+            pose_cond_tensor = pose_img.to(device="cuda")
+            pose_fea = self.pose_guider(pose_cond_tensor)
+            if seed_fea is None:
+                pose_fea = pose_fea
+            else:
+                pose_fea = seed_fea + pose_fea
+        else:
+            pose_fea = seed_fea
+
+
+        length = ref_image_latents.shape[2]
+        confidence_zero_to_one = torch.linspace(0, 1, steps=length).to(
+            dtype=ref_image_latents.dtype, device=ref_image_latents.device
+        )
+        confidence_one_to_zero = torch.linspace(1, 0, steps=length).to(
+            dtype=ref_image_latents.dtype, device=ref_image_latents.device
+        )
+        confidence_lst = []
+        confidence_lst.append(confidence_one_to_zero.to(
+            dtype=ref_image_latents.dtype, device=ref_image_latents.device
+        ))
+        for idx_1 in range(1,length-1):
+            confidence_lst.append(torch.cat([confidence_zero_to_one[-idx_1-1:],confidence_one_to_zero[1:-idx_1]]).to(
+            dtype=ref_image_latents.dtype, device=ref_image_latents.device
+        ))
+        confidence_lst.append(confidence_one_to_zero.to(
+            dtype=ref_image_latents.dtype, device=ref_image_latents.device
+        ))
+        # confidence_lst = [
+        #     x.view(1,length,1,1) for x in confidence_lst
+        # ]
+
+        if True: #not uncond_fwd:
+            ref_timesteps = torch.zeros_like(timesteps)
+            for ref_id in [0, 1, 2, 3, 12, 13, 14, 15]:
+                self.reference_unet(
+                    ref_image_latents[:, :, ref_id],
+                    ref_timesteps,
+                    encoder_hidden_states=clip_image_embeds,
+                    return_dict=False,
+                )
+                self.reference_control_reader.update(self.reference_control_writer, confidence_lst[ref_id])
+
+        model_pred = self.denoising_unet(
+            noisy_latents,
+            timesteps,
+            guidance_fea=pose_fea,
+            encoder_hidden_states=clip_image_embeds,
+        ).sample
+
+        return model_pred

2610_variant_runtime/diffusion/models/motion_module.py

@@ -0,0 +1,390 @@
+# Adapt from https://github.com/guoyww/AnimateDiff/blob/main/animatediff/models/motion_module.py
+import math
+from dataclasses import dataclass
+from typing import Callable, Optional
+
+import torch
+from diffusers.models.attention import FeedForward
+from diffusers.models.attention_processor import Attention, AttnProcessor
+from diffusers.utils import BaseOutput
+from diffusers.utils.import_utils import is_xformers_available
+from einops import rearrange, repeat
+from torch import nn
+
+
+def zero_module(module):
+    # Zero out the parameters of a module and return it.
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+@dataclass
+class TemporalTransformer3DModelOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
+
+def get_motion_module(in_channels, motion_module_type: str, motion_module_kwargs: dict):
+    if motion_module_type == "Vanilla":
+        return VanillaTemporalModule(
+            in_channels=in_channels,
+            **motion_module_kwargs,
+        )
+    else:
+        raise ValueError
+
+
+class VanillaTemporalModule(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        num_attention_heads=8,
+        num_transformer_block=2,
+        attention_block_types=("Temporal_Self", "Temporal_Self"),
+        cross_frame_attention_mode=None,
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+        temporal_attention_dim_div=1,
+        zero_initialize=True,
+    ):
+        super().__init__()
+
+        self.temporal_transformer = TemporalTransformer3DModel(
+            in_channels=in_channels,
+            num_attention_heads=num_attention_heads,
+            attention_head_dim=in_channels
+            // num_attention_heads
+            // temporal_attention_dim_div,
+            num_layers=num_transformer_block,
+            attention_block_types=attention_block_types,
+            cross_frame_attention_mode=cross_frame_attention_mode,
+            temporal_position_encoding=temporal_position_encoding,
+            temporal_position_encoding_max_len=temporal_position_encoding_max_len,
+        )
+
+        if zero_initialize:
+            self.temporal_transformer.proj_out = zero_module(
+                self.temporal_transformer.proj_out
+            )
+
+    def forward(
+        self,
+        input_tensor,
+        temb,
+        encoder_hidden_states,
+        attention_mask=None,
+        anchor_frame_idx=None,
+    ):
+        hidden_states = input_tensor
+        hidden_states = self.temporal_transformer(
+            hidden_states, encoder_hidden_states, attention_mask
+        )
+
+        output = hidden_states
+        return output
+
+
+class TemporalTransformer3DModel(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        num_attention_heads,
+        attention_head_dim,
+        num_layers,
+        attention_block_types=(
+            "Temporal_Self",
+            "Temporal_Self",
+        ),
+        dropout=0.0,
+        norm_num_groups=32,
+        cross_attention_dim=768,
+        activation_fn="geglu",
+        attention_bias=False,
+        upcast_attention=False,
+        cross_frame_attention_mode=None,
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+    ):
+        super().__init__()
+
+        inner_dim = num_attention_heads * attention_head_dim
+
+        self.norm = torch.nn.GroupNorm(
+            num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True
+        )
+        self.proj_in = nn.Linear(in_channels, inner_dim)
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                TemporalTransformerBlock(
+                    dim=inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    attention_block_types=attention_block_types,
+                    dropout=dropout,
+                    norm_num_groups=norm_num_groups,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    attention_bias=attention_bias,
+                    upcast_attention=upcast_attention,
+                    cross_frame_attention_mode=cross_frame_attention_mode,
+                    temporal_position_encoding=temporal_position_encoding,
+                    temporal_position_encoding_max_len=temporal_position_encoding_max_len,
+                )
+                for d in range(num_layers)
+            ]
+        )
+        self.proj_out = nn.Linear(inner_dim, in_channels)
+
+    def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None):
+        assert (
+            hidden_states.dim() == 5
+        ), f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
+        video_length = hidden_states.shape[2]
+        hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
+
+        batch, channel, height, weight = hidden_states.shape
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        inner_dim = hidden_states.shape[1]
+        hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+            batch, height * weight, inner_dim
+        )
+        hidden_states = self.proj_in(hidden_states)
+
+        # Transformer Blocks
+        for block in self.transformer_blocks:
+            hidden_states = block(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                video_length=video_length,
+            )
+
+        # output
+        hidden_states = self.proj_out(hidden_states)
+        hidden_states = (
+            hidden_states.reshape(batch, height, weight, inner_dim)
+            .permute(0, 3, 1, 2)
+            .contiguous()
+        )
+
+        output = hidden_states + residual
+        output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length)
+
+        return output
+
+
+class TemporalTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim,
+        num_attention_heads,
+        attention_head_dim,
+        attention_block_types=(
+            "Temporal_Self",
+            "Temporal_Self",
+        ),
+        dropout=0.0,
+        norm_num_groups=32,
+        cross_attention_dim=768,
+        activation_fn="geglu",
+        attention_bias=False,
+        upcast_attention=False,
+        cross_frame_attention_mode=None,
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+    ):
+        super().__init__()
+
+        attention_blocks = []
+        norms = []
+
+        for block_name in attention_block_types:
+            attention_blocks.append(
+                VersatileAttention(
+                    attention_mode=block_name.split("_")[0],
+                    cross_attention_dim=(
+                        cross_attention_dim if block_name.endswith("_Cross") else None
+                    ),
+                    query_dim=dim,
+                    heads=num_attention_heads,
+                    dim_head=attention_head_dim,
+                    dropout=dropout,
+                    bias=attention_bias,
+                    upcast_attention=upcast_attention,
+                    cross_frame_attention_mode=cross_frame_attention_mode,
+                    temporal_position_encoding=temporal_position_encoding,
+                    temporal_position_encoding_max_len=temporal_position_encoding_max_len,
+                )
+            )
+            norms.append(nn.LayerNorm(dim))
+
+        self.attention_blocks = nn.ModuleList(attention_blocks)
+        self.norms = nn.ModuleList(norms)
+
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
+        self.ff_norm = nn.LayerNorm(dim)
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        video_length=None,
+    ):
+        for attention_block, norm in zip(self.attention_blocks, self.norms):
+            norm_hidden_states = norm(hidden_states)
+            hidden_states = (
+                attention_block(
+                    norm_hidden_states,
+                    encoder_hidden_states=(
+                        encoder_hidden_states
+                        if attention_block.is_cross_attention
+                        else None
+                    ),
+                    video_length=video_length,
+                )
+                + hidden_states
+            )
+
+        hidden_states = self.ff(self.ff_norm(hidden_states)) + hidden_states
+
+        output = hidden_states
+        return output
+
+
+class PositionalEncoding(nn.Module):
+    def __init__(self, d_model, dropout=0.0, max_len=24):
+        super().__init__()
+        self.dropout = nn.Dropout(p=dropout)
+        position = torch.arange(max_len).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model)
+        )
+        pe = torch.zeros(1, max_len, d_model)
+        pe[0, :, 0::2] = torch.sin(position * div_term)
+        pe[0, :, 1::2] = torch.cos(position * div_term)
+        self.register_buffer("pe", pe)
+
+    def forward(self, x):
+        x = x + self.pe[:, : x.size(1)]
+        return self.dropout(x)
+
+
+class VersatileAttention(Attention):
+    def __init__(
+        self,
+        attention_mode=None,
+        cross_frame_attention_mode=None,
+        temporal_position_encoding=False,
+        temporal_position_encoding_max_len=24,
+        *args,
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+        assert attention_mode == "Temporal"
+
+        self.attention_mode = attention_mode
+        self.is_cross_attention = kwargs["cross_attention_dim"] is not None
+
+        self.pos_encoder = (
+            PositionalEncoding(
+                kwargs["query_dim"],
+                dropout=0.0,
+                max_len=temporal_position_encoding_max_len,
+            )
+            if (temporal_position_encoding and attention_mode == "Temporal")
+            else None
+        )
+
+    def extra_repr(self):
+        return f"(Module Info) Attention_Mode: {self.attention_mode}, Is_Cross_Attention: {self.is_cross_attention}"
+
+    def set_use_memory_efficient_attention_xformers(
+        self,
+        use_memory_efficient_attention_xformers: bool,
+        attention_op: Optional[Callable] = None,
+    ):
+        if use_memory_efficient_attention_xformers:
+            if not is_xformers_available():
+                raise ModuleNotFoundError(
+                    (
+                        "Refer to https://github.com/facebookresearch/xformers for more information on how to install"
+                        " xformers"
+                    ),
+                    name="xformers",
+                )
+            elif not torch.cuda.is_available():
+                raise ValueError(
+                    "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is"
+                    " only available for GPU "
+                )
+            else:
+                try:
+                    # Make sure we can run the memory efficient attention
+                    _ = xformers.ops.memory_efficient_attention(
+                        torch.randn((1, 2, 40), device="cuda"),
+                        torch.randn((1, 2, 40), device="cuda"),
+                        torch.randn((1, 2, 40), device="cuda"),
+                    )
+                except Exception as e:
+                    raise e
+
+            # XFormersAttnProcessor corrupts video generation and work with Pytorch 1.13.
+            # Pytorch 2.0.1 AttnProcessor works the same as XFormersAttnProcessor in Pytorch 1.13.
+            # You don't need XFormersAttnProcessor here.
+            # processor = XFormersAttnProcessor(
+            #     attention_op=attention_op,
+            # )
+            processor = AttnProcessor()
+        else:
+            processor = AttnProcessor()
+
+        self.set_processor(processor)
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        video_length=None,
+        **cross_attention_kwargs,
+    ):
+        if self.attention_mode == "Temporal":
+            d = hidden_states.shape[1]  # d means HxW
+            hidden_states = rearrange(
+                hidden_states, "(b f) d c -> (b d) f c", f=video_length
+            )
+
+            if self.pos_encoder is not None:
+                hidden_states = self.pos_encoder(hidden_states)
+
+            encoder_hidden_states = (
+                repeat(encoder_hidden_states, "b n c -> (b d) n c", d=d)
+                if encoder_hidden_states is not None
+                else encoder_hidden_states
+            )
+
+        else:
+            raise NotImplementedError
+
+        hidden_states = self.processor(
+            self,
+            hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+
+        if self.attention_mode == "Temporal":
+            hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)
+
+        return hidden_states

2610_variant_runtime/diffusion/models/mutual_self_attention.py

@@ -0,0 +1,365 @@
+# Adapted from https://github.com/magic-research/magic-animate/blob/main/magicanimate/models/mutual_self_attention.py
+from typing import Any, Dict, Optional
+
+import torch
+from einops import rearrange
+
+from models.attention import TemporalBasicTransformerBlock
+
+from .attention import BasicTransformerBlock
+
+
+def torch_dfs(model: torch.nn.Module):
+    result = [model]
+    for child in model.children():
+        result += torch_dfs(child)
+    return result
+
+
+class ReferenceAttentionControl:
+    def __init__(
+        self,
+        unet,
+        mode="write",
+        do_classifier_free_guidance=False,
+        attention_auto_machine_weight=float("inf"),
+        gn_auto_machine_weight=1.0,
+        style_fidelity=1.0,
+        reference_attn=True,
+        reference_adain=False,
+        fusion_blocks="midup",
+        batch_size=1,
+    ) -> None:
+        # 10. Modify self attention and group norm
+        self.unet = unet
+        assert mode in ["read", "write"]
+        assert fusion_blocks in ["midup", "full"]
+        self.reference_attn = reference_attn
+        self.reference_adain = reference_adain
+        self.fusion_blocks = fusion_blocks
+        self.register_reference_hooks(
+            mode,
+            do_classifier_free_guidance,
+            attention_auto_machine_weight,
+            gn_auto_machine_weight,
+            style_fidelity,
+            reference_attn,
+            reference_adain,
+            fusion_blocks,
+            batch_size=batch_size,
+        )
+
+    def register_reference_hooks(
+        self,
+        mode,
+        do_classifier_free_guidance,
+        attention_auto_machine_weight,
+        gn_auto_machine_weight,
+        style_fidelity,
+        reference_attn,
+        reference_adain,
+        dtype=torch.float16,
+        batch_size=1,
+        num_images_per_prompt=1,
+        device=torch.device("cpu"),
+        fusion_blocks="midup",
+    ):
+        MODE = mode
+        do_classifier_free_guidance = do_classifier_free_guidance
+        attention_auto_machine_weight = attention_auto_machine_weight
+        gn_auto_machine_weight = gn_auto_machine_weight
+        style_fidelity = style_fidelity
+        reference_attn = reference_attn
+        reference_adain = reference_adain
+        fusion_blocks = fusion_blocks
+        num_images_per_prompt = num_images_per_prompt
+        dtype = dtype
+        if do_classifier_free_guidance:
+            uc_mask = (
+                torch.Tensor(
+                    [1] * batch_size * num_images_per_prompt * 16
+                    + [0] * batch_size * num_images_per_prompt * 16
+                )
+                .to(device)
+                .bool()
+            )
+        else:
+            uc_mask = (
+                torch.Tensor([0] * batch_size * num_images_per_prompt * 2)
+                .to(device)
+                .bool()
+            )
+
+        def hacked_basic_transformer_inner_forward(
+            self,
+            hidden_states: torch.FloatTensor,
+            attention_mask: Optional[torch.FloatTensor] = None,
+            encoder_hidden_states: Optional[torch.FloatTensor] = None,
+            encoder_attention_mask: Optional[torch.FloatTensor] = None,
+            timestep: Optional[torch.LongTensor] = None,
+            cross_attention_kwargs: Dict[str, Any] = None,
+            class_labels: Optional[torch.LongTensor] = None,
+            video_length=None,
+        ):
+            if self.use_ada_layer_norm:  # False
+                norm_hidden_states = self.norm1(hidden_states, timestep)
+            elif self.use_ada_layer_norm_zero:
+                (
+                    norm_hidden_states,
+                    gate_msa,
+                    shift_mlp,
+                    scale_mlp,
+                    gate_mlp,
+                ) = self.norm1(
+                    hidden_states,
+                    timestep,
+                    class_labels,
+                    hidden_dtype=hidden_states.dtype,
+                )
+            else:
+                norm_hidden_states = self.norm1(hidden_states)
+
+            # 1. Self-Attention
+            # self.only_cross_attention = False
+            cross_attention_kwargs = (
+                cross_attention_kwargs if cross_attention_kwargs is not None else {}
+            )
+            if self.only_cross_attention:
+                attn_output = self.attn1(
+                    norm_hidden_states,
+                    encoder_hidden_states=(
+                        encoder_hidden_states if self.only_cross_attention else None
+                    ),
+                    attention_mask=attention_mask,
+                    **cross_attention_kwargs,
+                )
+            else:
+                if MODE == "write":
+                    self.bank.append(norm_hidden_states.clone())
+                    attn_output = self.attn1(
+                        norm_hidden_states,
+                        encoder_hidden_states=(
+                            encoder_hidden_states if self.only_cross_attention else None
+                        ),
+                        attention_mask=attention_mask,
+                        **cross_attention_kwargs,
+                    )
+                if MODE == "read":
+                    #print("before: ", self.bank[0].shape)
+                    bank_fea = [
+                        rearrange(
+                            d.unsqueeze(1).repeat(1, video_length, 1, 1),
+                            "b t l c -> (b t) l c",
+                        )
+                        for d in self.bank
+                    ]
+                    #print("after: ", bank_fea[0].shape)
+                    modify_norm_hidden_states = torch.cat(
+                        [norm_hidden_states] + bank_fea, dim=1
+                    )
+                    hidden_states_uc = (
+                        self.attn1(
+                            norm_hidden_states,
+                            encoder_hidden_states=modify_norm_hidden_states,
+                            attention_mask=attention_mask,
+                        )
+                        + hidden_states
+                    )
+                    if do_classifier_free_guidance:
+                        hidden_states_c = hidden_states_uc.clone()
+                        _uc_mask = uc_mask.clone()
+                        if hidden_states.shape[0] != _uc_mask.shape[0]:
+                            _uc_mask = (
+                                torch.Tensor(
+                                    [1] * (hidden_states.shape[0] // 2)
+                                    + [0] * (hidden_states.shape[0] // 2)
+                                )
+                                .to(device)
+                                .bool()
+                            )
+                        hidden_states_c[_uc_mask] = (
+                            self.attn1(
+                                norm_hidden_states[_uc_mask],
+                                encoder_hidden_states=norm_hidden_states[_uc_mask],
+                                attention_mask=attention_mask,
+                            )
+                            + hidden_states[_uc_mask]
+                        )
+                        hidden_states = hidden_states_c.clone()
+                    else:
+                        hidden_states = hidden_states_uc
+
+                    # self.bank.clear()
+                    if self.attn2 is not None:
+                        # Cross-Attention
+                        norm_hidden_states = (
+                            self.norm2(hidden_states, timestep)
+                            if self.use_ada_layer_norm
+                            else self.norm2(hidden_states)
+                        )
+                        hidden_states = (
+                            self.attn2(
+                                norm_hidden_states,
+                                encoder_hidden_states=encoder_hidden_states,
+                                attention_mask=attention_mask,
+                            )
+                            + hidden_states
+                        )
+
+                    # Feed-forward
+                    hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
+
+                    # Temporal-Attention
+                    if self.unet_use_temporal_attention:
+                        d = hidden_states.shape[1]
+                        hidden_states = rearrange(
+                            hidden_states, "(b f) d c -> (b d) f c", f=video_length
+                        )
+                        norm_hidden_states = (
+                            self.norm_temp(hidden_states, timestep)
+                            if self.use_ada_layer_norm
+                            else self.norm_temp(hidden_states)
+                        )
+                        hidden_states = (
+                            self.attn_temp(norm_hidden_states) + hidden_states
+                        )
+                        hidden_states = rearrange(
+                            hidden_states, "(b d) f c -> (b f) d c", d=d
+                        )
+
+                    return hidden_states
+
+            if self.use_ada_layer_norm_zero:
+                attn_output = gate_msa.unsqueeze(1) * attn_output
+            hidden_states = attn_output + hidden_states
+
+            if self.attn2 is not None:
+                norm_hidden_states = (
+                    self.norm2(hidden_states, timestep)
+                    if self.use_ada_layer_norm
+                    else self.norm2(hidden_states)
+                )
+
+                # 2. Cross-Attention
+                attn_output = self.attn2(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=encoder_attention_mask,
+                    **cross_attention_kwargs,
+                )
+                hidden_states = attn_output + hidden_states
+
+            # 3. Feed-forward
+            norm_hidden_states = self.norm3(hidden_states)
+
+            if self.use_ada_layer_norm_zero:
+                norm_hidden_states = (
+                    norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+                )
+
+            ff_output = self.ff(norm_hidden_states)
+
+            if self.use_ada_layer_norm_zero:
+                ff_output = gate_mlp.unsqueeze(1) * ff_output
+
+            hidden_states = ff_output + hidden_states
+
+            return hidden_states
+
+        if self.reference_attn:
+            if self.fusion_blocks == "midup":
+                attn_modules = [
+                    module
+                    for module in (
+                        torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
+                    )
+                    if isinstance(module, BasicTransformerBlock)
+                    or isinstance(module, TemporalBasicTransformerBlock)
+                ]
+            elif self.fusion_blocks == "full":
+                attn_modules = [
+                    module
+                    for module in torch_dfs(self.unet)
+                    if isinstance(module, BasicTransformerBlock)
+                    or isinstance(module, TemporalBasicTransformerBlock)
+                ]
+            attn_modules = sorted(
+                attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+
+            for i, module in enumerate(attn_modules):
+                module._original_inner_forward = module.forward
+                if isinstance(module, BasicTransformerBlock):
+                    module.forward = hacked_basic_transformer_inner_forward.__get__(
+                        module, BasicTransformerBlock
+                    )
+                if isinstance(module, TemporalBasicTransformerBlock):
+                    module.forward = hacked_basic_transformer_inner_forward.__get__(
+                        module, TemporalBasicTransformerBlock
+                    )
+
+                module.bank = []
+                module.attn_weight = float(i) / float(len(attn_modules))
+
+    def update(self, writer, dtype=torch.float16):
+        if self.reference_attn:
+            if self.fusion_blocks == "midup":
+                reader_attn_modules = [
+                    module
+                    for module in (
+                        torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
+                    )
+                    if isinstance(module, TemporalBasicTransformerBlock)
+                ]
+                writer_attn_modules = [
+                    module
+                    for module in (
+                        torch_dfs(writer.unet.mid_block)
+                        + torch_dfs(writer.unet.up_blocks)
+                    )
+                    if isinstance(module, BasicTransformerBlock)
+                ]
+            elif self.fusion_blocks == "full":
+                reader_attn_modules = [
+                    module
+                    for module in torch_dfs(self.unet)
+                    if isinstance(module, TemporalBasicTransformerBlock)
+                ]
+                writer_attn_modules = [
+                    module
+                    for module in torch_dfs(writer.unet)
+                    if isinstance(module, BasicTransformerBlock)
+                ]
+            reader_attn_modules = sorted(
+                reader_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+            writer_attn_modules = sorted(
+                writer_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+            for r, w in zip(reader_attn_modules, writer_attn_modules):
+                r.bank = [v.clone().to(dtype) for v in w.bank]
+                # w.bank.clear()
+
+    def clear(self):
+        if self.reference_attn:
+            if self.fusion_blocks == "midup":
+                reader_attn_modules = [
+                    module
+                    for module in (
+                        torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
+                    )
+                    if isinstance(module, BasicTransformerBlock)
+                    or isinstance(module, TemporalBasicTransformerBlock)
+                ]
+            elif self.fusion_blocks == "full":
+                reader_attn_modules = [
+                    module
+                    for module in torch_dfs(self.unet)
+                    if isinstance(module, BasicTransformerBlock)
+                    or isinstance(module, TemporalBasicTransformerBlock)
+                ]
+            reader_attn_modules = sorted(
+                reader_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
+            )
+            for r in reader_attn_modules:
+                r.bank.clear()

2610_variant_runtime/diffusion/models/pose_guider.py

@@ -0,0 +1,57 @@
+from typing import Tuple
+
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.nn.init as init
+from diffusers.models.modeling_utils import ModelMixin
+
+from models.motion_module import zero_module
+from models.resnet import InflatedConv3d
+
+
+class PoseGuider(ModelMixin):
+    def __init__(
+        self,
+        conditioning_embedding_channels: int,
+        conditioning_channels: int = 3,
+        block_out_channels: Tuple[int] = (16, 32, 64, 128),
+    ):
+        super().__init__()
+        self.conv_in = InflatedConv3d(
+            conditioning_channels, block_out_channels[0], kernel_size=3, padding=1
+        )
+
+        self.blocks = nn.ModuleList([])
+
+        for i in range(len(block_out_channels) - 1):
+            channel_in = block_out_channels[i]
+            channel_out = block_out_channels[i + 1]
+            self.blocks.append(
+                InflatedConv3d(channel_in, channel_in, kernel_size=3, padding=1)
+            )
+            self.blocks.append(
+                InflatedConv3d(
+                    channel_in, channel_out, kernel_size=3, padding=1, stride=2
+                )
+            )
+
+        self.conv_out = zero_module(
+            InflatedConv3d(
+                block_out_channels[-1],
+                conditioning_embedding_channels,
+                kernel_size=3,
+                padding=1,
+            )
+        )
+
+    def forward(self, conditioning):
+        embedding = self.conv_in(conditioning)
+        embedding = F.silu(embedding)
+
+        for block in self.blocks:
+            embedding = block(embedding)
+            embedding = F.silu(embedding)
+
+        embedding = self.conv_out(embedding)
+
+        return embedding

2610_variant_runtime/diffusion/models/resnet.py

@@ -0,0 +1,252 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/resnet.py
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+
+
+class InflatedConv3d(nn.Conv2d):
+    def forward(self, x):
+        video_length = x.shape[2]
+
+        x = rearrange(x, "b c f h w -> (b f) c h w")
+        x = super().forward(x)
+        x = rearrange(x, "(b f) c h w -> b c f h w", f=video_length)
+
+        return x
+
+
+class InflatedGroupNorm(nn.GroupNorm):
+    def forward(self, x):
+        video_length = x.shape[2]
+
+        x = rearrange(x, "b c f h w -> (b f) c h w")
+        x = super().forward(x)
+        x = rearrange(x, "(b f) c h w -> b c f h w", f=video_length)
+
+        return x
+
+
+class Upsample3D(nn.Module):
+    def __init__(
+        self,
+        channels,
+        use_conv=False,
+        use_conv_transpose=False,
+        out_channels=None,
+        name="conv",
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_conv_transpose = use_conv_transpose
+        self.name = name
+
+        conv = None
+        if use_conv_transpose:
+            raise NotImplementedError
+        elif use_conv:
+            self.conv = InflatedConv3d(self.channels, self.out_channels, 3, padding=1)
+
+    def forward(self, hidden_states, output_size=None):
+        assert hidden_states.shape[1] == self.channels
+
+        if self.use_conv_transpose:
+            raise NotImplementedError
+
+        # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16
+        dtype = hidden_states.dtype
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(torch.float32)
+
+        # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
+        if hidden_states.shape[0] >= 64:
+            hidden_states = hidden_states.contiguous()
+
+        # if `output_size` is passed we force the interpolation output
+        # size and do not make use of `scale_factor=2`
+        if output_size is None:
+            hidden_states = F.interpolate(
+                hidden_states, scale_factor=[1.0, 2.0, 2.0], mode="nearest"
+            )
+        else:
+            hidden_states = F.interpolate(
+                hidden_states, size=output_size, mode="nearest"
+            )
+
+        # If the input is bfloat16, we cast back to bfloat16
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(dtype)
+
+        # if self.use_conv:
+        #     if self.name == "conv":
+        #         hidden_states = self.conv(hidden_states)
+        #     else:
+        #         hidden_states = self.Conv2d_0(hidden_states)
+        hidden_states = self.conv(hidden_states)
+
+        return hidden_states
+
+
+class Downsample3D(nn.Module):
+    def __init__(
+        self, channels, use_conv=False, out_channels=None, padding=1, name="conv"
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.padding = padding
+        stride = 2
+        self.name = name
+
+        if use_conv:
+            self.conv = InflatedConv3d(
+                self.channels, self.out_channels, 3, stride=stride, padding=padding
+            )
+        else:
+            raise NotImplementedError
+
+    def forward(self, hidden_states):
+        assert hidden_states.shape[1] == self.channels
+        if self.use_conv and self.padding == 0:
+            raise NotImplementedError
+
+        assert hidden_states.shape[1] == self.channels
+        hidden_states = self.conv(hidden_states)
+
+        return hidden_states
+
+
+class ResnetBlock3D(nn.Module):
+    def __init__(
+        self,
+        *,
+        in_channels,
+        out_channels=None,
+        conv_shortcut=False,
+        dropout=0.0,
+        temb_channels=512,
+        groups=32,
+        groups_out=None,
+        pre_norm=True,
+        eps=1e-6,
+        non_linearity="swish",
+        time_embedding_norm="default",
+        output_scale_factor=1.0,
+        use_in_shortcut=None,
+        use_inflated_groupnorm=None,
+    ):
+        super().__init__()
+        self.pre_norm = pre_norm
+        self.pre_norm = True
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+        self.time_embedding_norm = time_embedding_norm
+        self.output_scale_factor = output_scale_factor
+
+        if groups_out is None:
+            groups_out = groups
+
+        assert use_inflated_groupnorm != None
+        if use_inflated_groupnorm:
+            self.norm1 = InflatedGroupNorm(
+                num_groups=groups, num_channels=in_channels, eps=eps, affine=True
+            )
+        else:
+            self.norm1 = torch.nn.GroupNorm(
+                num_groups=groups, num_channels=in_channels, eps=eps, affine=True
+            )
+
+        self.conv1 = InflatedConv3d(
+            in_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+
+        if temb_channels is not None:
+            if self.time_embedding_norm == "default":
+                time_emb_proj_out_channels = out_channels
+            elif self.time_embedding_norm == "scale_shift":
+                time_emb_proj_out_channels = out_channels * 2
+            else:
+                raise ValueError(
+                    f"unknown time_embedding_norm : {self.time_embedding_norm} "
+                )
+
+            self.time_emb_proj = torch.nn.Linear(
+                temb_channels, time_emb_proj_out_channels
+            )
+        else:
+            self.time_emb_proj = None
+
+        if use_inflated_groupnorm:
+            self.norm2 = InflatedGroupNorm(
+                num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True
+            )
+        else:
+            self.norm2 = torch.nn.GroupNorm(
+                num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True
+            )
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = InflatedConv3d(
+            out_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+
+        if non_linearity == "swish":
+            self.nonlinearity = lambda x: F.silu(x)
+        elif non_linearity == "mish":
+            self.nonlinearity = Mish()
+        elif non_linearity == "silu":
+            self.nonlinearity = nn.SiLU()
+
+        self.use_in_shortcut = (
+            self.in_channels != self.out_channels
+            if use_in_shortcut is None
+            else use_in_shortcut
+        )
+
+        self.conv_shortcut = None
+        if self.use_in_shortcut:
+            self.conv_shortcut = InflatedConv3d(
+                in_channels, out_channels, kernel_size=1, stride=1, padding=0
+            )
+
+    def forward(self, input_tensor, temb):
+        hidden_states = input_tensor
+
+        hidden_states = self.norm1(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.conv1(hidden_states)
+
+        if temb is not None:
+            temb = self.time_emb_proj(self.nonlinearity(temb))[:, :, None, None, None]
+
+        if temb is not None and self.time_embedding_norm == "default":
+            hidden_states = hidden_states + temb
+
+        hidden_states = self.norm2(hidden_states)
+
+        if temb is not None and self.time_embedding_norm == "scale_shift":
+            scale, shift = torch.chunk(temb, 2, dim=1)
+            hidden_states = hidden_states * (1 + scale) + shift
+
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+
+        if self.conv_shortcut is not None:
+            input_tensor = self.conv_shortcut(input_tensor)
+
+        output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
+
+        return output_tensor
+
+
+class Mish(torch.nn.Module):
+    def forward(self, hidden_states):
+        return hidden_states * torch.tanh(torch.nn.functional.softplus(hidden_states))

2610_variant_runtime/diffusion/models/transformer_2d.py

@@ -0,0 +1,396 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformer_2d.py
+from dataclasses import dataclass
+from typing import Any, Dict, Optional
+
+import torch
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models.embeddings import CaptionProjection
+from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import AdaLayerNormSingle
+from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, is_torch_version
+from torch import nn
+
+from .attention import BasicTransformerBlock
+
+
+@dataclass
+class Transformer2DModelOutput(BaseOutput):
+    """
+    The output of [`Transformer2DModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
+            The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
+            distributions for the unnoised latent pixels.
+    """
+
+    sample: torch.FloatTensor
+    ref_feature: torch.FloatTensor
+
+
+class Transformer2DModel(ModelMixin, ConfigMixin):
+    """
+    A 2D Transformer model for image-like data.
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            The number of channels in the input and output (specify if the input is **continuous**).
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+        sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
+            This is fixed during training since it is used to learn a number of position embeddings.
+        num_vector_embeds (`int`, *optional*):
+            The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
+            Includes the class for the masked latent pixel.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
+        num_embeds_ada_norm ( `int`, *optional*):
+            The number of diffusion steps used during training. Pass if at least one of the norm_layers is
+            `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
+            added to the hidden states.
+
+            During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
+        attention_bias (`bool`, *optional*):
+            Configure if the `TransformerBlocks` attention should contain a bias parameter.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        out_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        sample_size: Optional[int] = None,
+        num_vector_embeds: Optional[int] = None,
+        patch_size: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_type: str = "layer_norm",
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        attention_type: str = "default",
+        caption_channels: int = None,
+    ):
+        super().__init__()
+        self.use_linear_projection = use_linear_projection
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+
+        conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+        linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
+
+        # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
+        # Define whether input is continuous or discrete depending on configuration
+        self.is_input_continuous = (in_channels is not None) and (patch_size is None)
+        self.is_input_vectorized = num_vector_embeds is not None
+        self.is_input_patches = in_channels is not None and patch_size is not None
+
+        if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
+            deprecation_message = (
+                f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
+                " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
+                " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
+                " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
+                " would be very nice if you could open a Pull request for the `transformer/config.json` file"
+            )
+            deprecate(
+                "norm_type!=num_embeds_ada_norm",
+                "1.0.0",
+                deprecation_message,
+                standard_warn=False,
+            )
+            norm_type = "ada_norm"
+
+        if self.is_input_continuous and self.is_input_vectorized:
+            raise ValueError(
+                f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
+                " sure that either `in_channels` or `num_vector_embeds` is None."
+            )
+        elif self.is_input_vectorized and self.is_input_patches:
+            raise ValueError(
+                f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
+                " sure that either `num_vector_embeds` or `num_patches` is None."
+            )
+        elif (
+            not self.is_input_continuous
+            and not self.is_input_vectorized
+            and not self.is_input_patches
+        ):
+            raise ValueError(
+                f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
+                f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
+            )
+
+        # 2. Define input layers
+        self.in_channels = in_channels
+
+        self.norm = torch.nn.GroupNorm(
+            num_groups=norm_num_groups,
+            num_channels=in_channels,
+            eps=1e-6,
+            affine=True,
+        )
+        if use_linear_projection:
+            self.proj_in = linear_cls(in_channels, inner_dim)
+        else:
+            self.proj_in = conv_cls(
+                in_channels, inner_dim, kernel_size=1, stride=1, padding=0
+            )
+
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                    only_cross_attention=only_cross_attention,
+                    double_self_attention=double_self_attention,
+                    upcast_attention=upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                    attention_type=attention_type,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        self.out_channels = in_channels if out_channels is None else out_channels
+        # TODO: should use out_channels for continuous projections
+        if use_linear_projection:
+            self.proj_out = linear_cls(inner_dim, in_channels)
+        else:
+            self.proj_out = conv_cls(
+                inner_dim, in_channels, kernel_size=1, stride=1, padding=0
+            )
+
+        # 5. PixArt-Alpha blocks.
+        self.adaln_single = None
+        self.use_additional_conditions = False
+        if norm_type == "ada_norm_single":
+            self.use_additional_conditions = self.config.sample_size == 128
+            # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
+            # additional conditions until we find better name
+            self.adaln_single = AdaLayerNormSingle(
+                inner_dim, use_additional_conditions=self.use_additional_conditions
+            )
+
+        self.caption_projection = None
+        if caption_channels is not None:
+            self.caption_projection = CaptionProjection(
+                in_features=caption_channels, hidden_size=inner_dim
+            )
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Dict[str, torch.Tensor] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ):
+        """
+        The [`Transformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
+                Input `hidden_states`.
+            encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            timestep ( `torch.LongTensor`, *optional*):
+                Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
+            class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
+                Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
+                `AdaLayerZeroNorm`.
+            cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            attention_mask ( `torch.Tensor`, *optional*):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            encoder_attention_mask ( `torch.Tensor`, *optional*):
+                Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
+
+                    * Mask `(batch, sequence_length)` True = keep, False = discard.
+                    * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
+
+                If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
+                above. This bias will be added to the cross-attention scores.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
+        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None and attention_mask.ndim == 2:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (
+                1 - encoder_attention_mask.to(hidden_states.dtype)
+            ) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # Retrieve lora scale.
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+
+        # 1. Input
+        batch, _, height, width = hidden_states.shape
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        if not self.use_linear_projection:
+            hidden_states = (
+                self.proj_in(hidden_states, scale=lora_scale)
+                if not USE_PEFT_BACKEND
+                else self.proj_in(hidden_states)
+            )
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+                batch, height * width, inner_dim
+            )
+        else:
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+                batch, height * width, inner_dim
+            )
+            hidden_states = (
+                self.proj_in(hidden_states, scale=lora_scale)
+                if not USE_PEFT_BACKEND
+                else self.proj_in(hidden_states)
+            )
+
+        # 2. Blocks
+        if self.caption_projection is not None:
+            batch_size = hidden_states.shape[0]
+            encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states.view(
+                batch_size, -1, hidden_states.shape[-1]
+            )
+
+        ref_feature = hidden_states.reshape(batch, height, width, inner_dim)
+        for block in self.transformer_blocks:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    cross_attention_kwargs,
+                    class_labels,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    timestep=timestep,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    class_labels=class_labels,
+                )
+
+        # 3. Output
+        if self.is_input_continuous:
+            if not self.use_linear_projection:
+                hidden_states = (
+                    hidden_states.reshape(batch, height, width, inner_dim)
+                    .permute(0, 3, 1, 2)
+                    .contiguous()
+                )
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+            else:
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+                hidden_states = (
+                    hidden_states.reshape(batch, height, width, inner_dim)
+                    .permute(0, 3, 1, 2)
+                    .contiguous()
+                )
+
+            output = hidden_states + residual
+        if not return_dict:
+            return (output, ref_feature)
+
+        return Transformer2DModelOutput(sample=output, ref_feature=ref_feature)

2610_variant_runtime/diffusion/models/transformer_3d.py

@@ -0,0 +1,170 @@
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models import ModelMixin
+from diffusers.utils import BaseOutput
+from diffusers.utils.import_utils import is_xformers_available
+from einops import rearrange, repeat
+from torch import nn
+
+from .attention import TemporalBasicTransformerBlock
+
+
+@dataclass
+class Transformer3DModelOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
+
+class Transformer3DModel(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+    ):
+        super().__init__()
+        self.use_linear_projection = use_linear_projection
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+
+        # Define input layers
+        self.in_channels = in_channels
+
+        self.norm = torch.nn.GroupNorm(
+            num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True
+        )
+        if use_linear_projection:
+            self.proj_in = nn.Linear(in_channels, inner_dim)
+        else:
+            self.proj_in = nn.Conv2d(
+                in_channels, inner_dim, kernel_size=1, stride=1, padding=0
+            )
+
+        # Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                TemporalBasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        if use_linear_projection:
+            self.proj_out = nn.Linear(in_channels, inner_dim)
+        else:
+            self.proj_out = nn.Conv2d(
+                inner_dim, in_channels, kernel_size=1, stride=1, padding=0
+            )
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        timestep=None,
+        return_dict: bool = True,
+    ):
+        # Input
+        assert (
+            hidden_states.dim() == 5
+        ), f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
+        video_length = hidden_states.shape[2]
+        hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
+        if encoder_hidden_states is not None:
+            if encoder_hidden_states.shape[0] != hidden_states.shape[0]:
+                encoder_hidden_states = repeat(
+                    encoder_hidden_states, "b n c -> (b f) n c", f=video_length
+                )
+
+        batch, channel, height, weight = hidden_states.shape
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        if not self.use_linear_projection:
+            hidden_states = self.proj_in(hidden_states)
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+                batch, height * weight, inner_dim
+            )
+        else:
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+                batch, height * weight, inner_dim
+            )
+            hidden_states = self.proj_in(hidden_states)
+
+        # Blocks
+        for i, block in enumerate(self.transformer_blocks):
+            hidden_states = block(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                timestep=timestep,
+                video_length=video_length,
+            )
+
+        # Output
+        if not self.use_linear_projection:
+            hidden_states = (
+                hidden_states.reshape(batch, height, weight, inner_dim)
+                .permute(0, 3, 1, 2)
+                .contiguous()
+            )
+            hidden_states = self.proj_out(hidden_states)
+        else:
+            hidden_states = self.proj_out(hidden_states)
+            hidden_states = (
+                hidden_states.reshape(batch, height, weight, inner_dim)
+                .permute(0, 3, 1, 2)
+                .contiguous()
+            )
+
+        output = hidden_states + residual
+
+        output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length)
+        if not return_dict:
+            return (output,)
+
+        return Transformer3DModelOutput(sample=output)

2610_variant_runtime/diffusion/models/unet_2d_blocks.py

@@ -0,0 +1,1076 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_blocks.py
+from typing import Any, Dict, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import Attention
+from diffusers.models.dual_transformer_2d import DualTransformer2DModel
+from diffusers.models.resnet import Downsample2D, ResnetBlock2D, Upsample2D
+from diffusers.utils import is_torch_version, logging
+from diffusers.utils.torch_utils import apply_freeu
+from torch import nn
+
+from .transformer_2d import Transformer2DModel
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def get_down_block(
+    down_block_type: str,
+    num_layers: int,
+    in_channels: int,
+    out_channels: int,
+    temb_channels: int,
+    add_downsample: bool,
+    resnet_eps: float,
+    resnet_act_fn: str,
+    transformer_layers_per_block: int = 1,
+    num_attention_heads: Optional[int] = None,
+    resnet_groups: Optional[int] = None,
+    cross_attention_dim: Optional[int] = None,
+    downsample_padding: Optional[int] = None,
+    dual_cross_attention: bool = False,
+    use_linear_projection: bool = False,
+    only_cross_attention: bool = False,
+    upcast_attention: bool = False,
+    resnet_time_scale_shift: str = "default",
+    attention_type: str = "default",
+    resnet_skip_time_act: bool = False,
+    resnet_out_scale_factor: float = 1.0,
+    cross_attention_norm: Optional[str] = None,
+    attention_head_dim: Optional[int] = None,
+    downsample_type: Optional[str] = None,
+    dropout: float = 0.0,
+):
+    # If attn head dim is not defined, we default it to the number of heads
+    if attention_head_dim is None:
+        logger.warn(
+            f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
+        )
+        attention_head_dim = num_attention_heads
+
+    down_block_type = (
+        down_block_type[7:]
+        if down_block_type.startswith("UNetRes")
+        else down_block_type
+    )
+    if down_block_type == "DownBlock2D":
+        return DownBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            dropout=dropout,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif down_block_type == "CrossAttnDownBlock2D":
+        if cross_attention_dim is None:
+            raise ValueError(
+                "cross_attention_dim must be specified for CrossAttnDownBlock2D"
+            )
+        return CrossAttnDownBlock2D(
+            num_layers=num_layers,
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            dropout=dropout,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            cross_attention_dim=cross_attention_dim,
+            num_attention_heads=num_attention_heads,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            attention_type=attention_type,
+        )
+    raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_up_block(
+    up_block_type: str,
+    num_layers: int,
+    in_channels: int,
+    out_channels: int,
+    prev_output_channel: int,
+    temb_channels: int,
+    add_upsample: bool,
+    resnet_eps: float,
+    resnet_act_fn: str,
+    resolution_idx: Optional[int] = None,
+    transformer_layers_per_block: int = 1,
+    num_attention_heads: Optional[int] = None,
+    resnet_groups: Optional[int] = None,
+    cross_attention_dim: Optional[int] = None,
+    dual_cross_attention: bool = False,
+    use_linear_projection: bool = False,
+    only_cross_attention: bool = False,
+    upcast_attention: bool = False,
+    resnet_time_scale_shift: str = "default",
+    attention_type: str = "default",
+    resnet_skip_time_act: bool = False,
+    resnet_out_scale_factor: float = 1.0,
+    cross_attention_norm: Optional[str] = None,
+    attention_head_dim: Optional[int] = None,
+    upsample_type: Optional[str] = None,
+    dropout: float = 0.0,
+) -> nn.Module:
+    # If attn head dim is not defined, we default it to the number of heads
+    if attention_head_dim is None:
+        logger.warn(
+            f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
+        )
+        attention_head_dim = num_attention_heads
+
+    up_block_type = (
+        up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
+    )
+    if up_block_type == "UpBlock2D":
+        return UpBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            resolution_idx=resolution_idx,
+            dropout=dropout,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif up_block_type == "CrossAttnUpBlock2D":
+        if cross_attention_dim is None:
+            raise ValueError(
+                "cross_attention_dim must be specified for CrossAttnUpBlock2D"
+            )
+        return CrossAttnUpBlock2D(
+            num_layers=num_layers,
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            resolution_idx=resolution_idx,
+            dropout=dropout,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            cross_attention_dim=cross_attention_dim,
+            num_attention_heads=num_attention_heads,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            attention_type=attention_type,
+        )
+
+    raise ValueError(f"{up_block_type} does not exist.")
+
+
+class AutoencoderTinyBlock(nn.Module):
+    """
+    Tiny Autoencoder block used in [`AutoencoderTiny`]. It is a mini residual module consisting of plain conv + ReLU
+    blocks.
+
+    Args:
+        in_channels (`int`): The number of input channels.
+        out_channels (`int`): The number of output channels.
+        act_fn (`str`):
+            ` The activation function to use. Supported values are `"swish"`, `"mish"`, `"gelu"`, and `"relu"`.
+
+    Returns:
+        `torch.FloatTensor`: A tensor with the same shape as the input tensor, but with the number of channels equal to
+        `out_channels`.
+    """
+
+    def __init__(self, in_channels: int, out_channels: int, act_fn: str):
+        super().__init__()
+        act_fn = get_activation(act_fn)
+        self.conv = nn.Sequential(
+            nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
+            act_fn,
+            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
+            act_fn,
+            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
+        )
+        self.skip = (
+            nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
+            if in_channels != out_channels
+            else nn.Identity()
+        )
+        self.fuse = nn.ReLU()
+
+    def forward(self, x: torch.FloatTensor) -> torch.FloatTensor:
+        return self.fuse(self.conv(x) + self.skip(x))
+
+
+class UNetMidBlock2D(nn.Module):
+    """
+    A 2D UNet mid-block [`UNetMidBlock2D`] with multiple residual blocks and optional attention blocks.
+
+    Args:
+        in_channels (`int`): The number of input channels.
+        temb_channels (`int`): The number of temporal embedding channels.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout rate.
+        num_layers (`int`, *optional*, defaults to 1): The number of residual blocks.
+        resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks.
+        resnet_time_scale_shift (`str`, *optional*, defaults to `default`):
+            The type of normalization to apply to the time embeddings. This can help to improve the performance of the
+            model on tasks with long-range temporal dependencies.
+        resnet_act_fn (`str`, *optional*, defaults to `swish`): The activation function for the resnet blocks.
+        resnet_groups (`int`, *optional*, defaults to 32):
+            The number of groups to use in the group normalization layers of the resnet blocks.
+        attn_groups (`Optional[int]`, *optional*, defaults to None): The number of groups for the attention blocks.
+        resnet_pre_norm (`bool`, *optional*, defaults to `True`):
+            Whether to use pre-normalization for the resnet blocks.
+        add_attention (`bool`, *optional*, defaults to `True`): Whether to add attention blocks.
+        attention_head_dim (`int`, *optional*, defaults to 1):
+            Dimension of a single attention head. The number of attention heads is determined based on this value and
+            the number of input channels.
+        output_scale_factor (`float`, *optional*, defaults to 1.0): The output scale factor.
+
+    Returns:
+        `torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size,
+        in_channels, height, width)`.
+
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",  # default, spatial
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        attn_groups: Optional[int] = None,
+        resnet_pre_norm: bool = True,
+        add_attention: bool = True,
+        attention_head_dim: int = 1,
+        output_scale_factor: float = 1.0,
+    ):
+        super().__init__()
+        resnet_groups = (
+            resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+        )
+        self.add_attention = add_attention
+
+        if attn_groups is None:
+            attn_groups = (
+                resnet_groups if resnet_time_scale_shift == "default" else None
+            )
+
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock2D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+            )
+        ]
+        attentions = []
+
+        if attention_head_dim is None:
+            logger.warn(
+                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}."
+            )
+            attention_head_dim = in_channels
+
+        for _ in range(num_layers):
+            if self.add_attention:
+                attentions.append(
+                    Attention(
+                        in_channels,
+                        heads=in_channels // attention_head_dim,
+                        dim_head=attention_head_dim,
+                        rescale_output_factor=output_scale_factor,
+                        eps=resnet_eps,
+                        norm_num_groups=attn_groups,
+                        spatial_norm_dim=(
+                            temb_channels
+                            if resnet_time_scale_shift == "spatial"
+                            else None
+                        ),
+                        residual_connection=True,
+                        bias=True,
+                        upcast_softmax=True,
+                        _from_deprecated_attn_block=True,
+                    )
+                )
+            else:
+                attentions.append(None)
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+    def forward(
+        self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None
+    ) -> torch.FloatTensor:
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            if attn is not None:
+                hidden_states = attn(hidden_states, temb=temb)
+            hidden_states = resnet(hidden_states, temb)
+
+        return hidden_states
+
+
+class UNetMidBlock2DCrossAttn(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads: int = 1,
+        output_scale_factor: float = 1.0,
+        cross_attention_dim: int = 1280,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        upcast_attention: bool = False,
+        attention_type: str = "default",
+    ):
+        super().__init__()
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+        resnet_groups = (
+            resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+        )
+
+        # support for variable transformer layers per block
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock2D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+            )
+        ]
+        attentions = []
+
+        for i in range(num_layers):
+            if not dual_cross_attention:
+                attentions.append(
+                    Transformer2DModel(
+                        num_attention_heads,
+                        in_channels // num_attention_heads,
+                        in_channels=in_channels,
+                        num_layers=transformer_layers_per_block[i],
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        upcast_attention=upcast_attention,
+                        attention_type=attention_type,
+                    )
+                )
+            else:
+                attentions.append(
+                    DualTransformer2DModel(
+                        num_attention_heads,
+                        in_channels // num_attention_heads,
+                        in_channels=in_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
+                )
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+        hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+
+        return hidden_states
+
+
+class CrossAttnDownBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads: int = 1,
+        cross_attention_dim: int = 1280,
+        output_scale_factor: float = 1.0,
+        downsample_padding: int = 1,
+        add_downsample: bool = True,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        attention_type: str = "default",
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            if not dual_cross_attention:
+                attentions.append(
+                    Transformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=transformer_layers_per_block[i],
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        only_cross_attention=only_cross_attention,
+                        upcast_attention=upcast_attention,
+                        attention_type=attention_type,
+                    )
+                )
+            else:
+                attentions.append(
+                    DualTransformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
+                )
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample2D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        additional_residuals: Optional[torch.FloatTensor] = None,
+    ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+        output_states = ()
+
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+
+        blocks = list(zip(self.resnets, self.attentions))
+
+        for i, (resnet, attn) in enumerate(blocks):
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+
+            # apply additional residuals to the output of the last pair of resnet and attention blocks
+            if i == len(blocks) - 1 and additional_residuals is not None:
+                hidden_states = hidden_states + additional_residuals
+
+            output_states = output_states + (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states, scale=lora_scale)
+
+            output_states = output_states + (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class DownBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor: float = 1.0,
+        add_downsample: bool = True,
+        downsample_padding: int = 1,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample2D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        scale: float = 1.0,
+    ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+        output_states = ()
+
+        for resnet in self.resnets:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                if is_torch_version(">=", "1.11.0"):
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet),
+                        hidden_states,
+                        temb,
+                        use_reentrant=False,
+                    )
+                else:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet), hidden_states, temb
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=scale)
+
+            output_states = output_states + (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states, scale=scale)
+
+            output_states = output_states + (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class CrossAttnUpBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        prev_output_channel: int,
+        temb_channels: int,
+        resolution_idx: Optional[int] = None,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads: int = 1,
+        cross_attention_dim: int = 1280,
+        output_scale_factor: float = 1.0,
+        add_upsample: bool = True,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+        attention_type: str = "default",
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            if not dual_cross_attention:
+                attentions.append(
+                    Transformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=transformer_layers_per_block[i],
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        only_cross_attention=only_cross_attention,
+                        upcast_attention=upcast_attention,
+                        attention_type=attention_type,
+                    )
+                )
+            else:
+                attentions.append(
+                    DualTransformer2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=1,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                    )
+                )
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]
+            )
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+        self.resolution_idx = resolution_idx
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        upsample_size: Optional[int] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+        is_freeu_enabled = (
+            getattr(self, "s1", None)
+            and getattr(self, "s2", None)
+            and getattr(self, "b1", None)
+            and getattr(self, "b2", None)
+        )
+
+        for resnet, attn in zip(self.resnets, self.attentions):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+            # FreeU: Only operate on the first two stages
+            if is_freeu_enabled:
+                hidden_states, res_hidden_states = apply_freeu(
+                    self.resolution_idx,
+                    hidden_states,
+                    res_hidden_states,
+                    s1=self.s1,
+                    s2=self.s2,
+                    b1=self.b1,
+                    b2=self.b2,
+                )
+
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = (
+                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+                hidden_states, ref_feature = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(
+                    hidden_states, upsample_size, scale=lora_scale
+                )
+
+        return hidden_states
+
+
+class UpBlock2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        prev_output_channel: int,
+        out_channels: int,
+        temb_channels: int,
+        resolution_idx: Optional[int] = None,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor: float = 1.0,
+        add_upsample: bool = True,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]
+            )
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+        self.resolution_idx = resolution_idx
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+        temb: Optional[torch.FloatTensor] = None,
+        upsample_size: Optional[int] = None,
+        scale: float = 1.0,
+    ) -> torch.FloatTensor:
+        is_freeu_enabled = (
+            getattr(self, "s1", None)
+            and getattr(self, "s2", None)
+            and getattr(self, "b1", None)
+            and getattr(self, "b2", None)
+        )
+
+        for resnet in self.resnets:
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+            # FreeU: Only operate on the first two stages
+            if is_freeu_enabled:
+                hidden_states, res_hidden_states = apply_freeu(
+                    self.resolution_idx,
+                    hidden_states,
+                    res_hidden_states,
+                    s1=self.s1,
+                    s2=self.s2,
+                    b1=self.b1,
+                    b2=self.b2,
+                )
+
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                if is_torch_version(">=", "1.11.0"):
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet),
+                        hidden_states,
+                        temb,
+                        use_reentrant=False,
+                    )
+                else:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(resnet), hidden_states, temb
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb, scale=scale)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size, scale=scale)
+
+        return hidden_states

2610_variant_runtime/diffusion/models/unet_2d_condition.py

@@ -0,0 +1,1313 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_condition.py
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import UNet2DConditionLoadersMixin
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from diffusers.models.embeddings import (
+    GaussianFourierProjection,
+    ImageHintTimeEmbedding,
+    ImageProjection,
+    ImageTimeEmbedding,
+    PositionNet,
+    TextImageProjection,
+    TextImageTimeEmbedding,
+    TextTimeEmbedding,
+    TimestepEmbedding,
+    Timesteps,
+)
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.utils import (
+    USE_PEFT_BACKEND,
+    BaseOutput,
+    deprecate,
+    logging,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+
+from .unet_2d_blocks import (
+    UNetMidBlock2D,
+    UNetMidBlock2DCrossAttn,
+    get_down_block,
+    get_up_block,
+)
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class UNet2DConditionOutput(BaseOutput):
+    """
+    The output of [`UNet2DConditionModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+    """
+
+    sample: torch.FloatTensor = None
+    ref_features: Tuple[torch.FloatTensor] = None
+
+
+class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+    r"""
+    A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
+    shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+            Height and width of input/output sample.
+        in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+        flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
+            Whether to flip the sin to cos in the time embedding.
+        freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+            Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
+            `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+            The tuple of upsample blocks to use.
+        only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
+            Whether to include self-attention in the basic transformer blocks, see
+            [`~models.attention.BasicTransformerBlock`].
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+        downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+        mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+            If `None`, normalization and activation layers is skipped in post-processing.
+        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+            The dimension of the cross attention features.
+        transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+       reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
+            blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        encoder_hid_dim (`int`, *optional*, defaults to None):
+            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+            dimension to `cross_attention_dim`.
+        encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+            If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
+        num_attention_heads (`int`, *optional*):
+            The number of attention heads. If not defined, defaults to `attention_head_dim`
+        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+            for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+        class_embed_type (`str`, *optional*, defaults to `None`):
+            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+        addition_embed_type (`str`, *optional*, defaults to `None`):
+            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+            "text". "text" will use the `TextTimeEmbedding` layer.
+        addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
+            Dimension for the timestep embeddings.
+        num_class_embeds (`int`, *optional*, defaults to `None`):
+            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+            class conditioning with `class_embed_type` equal to `None`.
+        time_embedding_type (`str`, *optional*, defaults to `positional`):
+            The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
+        time_embedding_dim (`int`, *optional*, defaults to `None`):
+            An optional override for the dimension of the projected time embedding.
+        time_embedding_act_fn (`str`, *optional*, defaults to `None`):
+            Optional activation function to use only once on the time embeddings before they are passed to the rest of
+            the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
+        timestep_post_act (`str`, *optional*, defaults to `None`):
+            The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
+        time_cond_proj_dim (`int`, *optional*, defaults to `None`):
+            The dimension of `cond_proj` layer in the timestep embedding.
+        conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`,
+        *optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`,
+        *optional*): The dimension of the `class_labels` input when
+            `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
+        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+            embeddings with the class embeddings.
+        mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
+            Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
+            `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
+            `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
+            otherwise.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 4,
+        out_channels: int = 4,
+        center_input_sample: bool = False,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "DownBlock2D",
+        ),
+        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+        up_block_types: Tuple[str] = (
+            "UpBlock2D",
+            "CrossAttnUpBlock2D",
+            "CrossAttnUpBlock2D",
+            "CrossAttnUpBlock2D",
+        ),
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        layers_per_block: Union[int, Tuple[int]] = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        dropout: float = 0.0,
+        act_fn: str = "silu",
+        norm_num_groups: Optional[int] = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: Union[int, Tuple[int]] = 1280,
+        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
+        reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
+        encoder_hid_dim: Optional[int] = None,
+        encoder_hid_dim_type: Optional[str] = None,
+        attention_head_dim: Union[int, Tuple[int]] = 8,
+        num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        addition_embed_type: Optional[str] = None,
+        addition_time_embed_dim: Optional[int] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        resnet_skip_time_act: bool = False,
+        resnet_out_scale_factor: int = 1.0,
+        time_embedding_type: str = "positional",
+        time_embedding_dim: Optional[int] = None,
+        time_embedding_act_fn: Optional[str] = None,
+        timestep_post_act: Optional[str] = None,
+        time_cond_proj_dim: Optional[int] = None,
+        conv_in_kernel: int = 3,
+        conv_out_kernel: int = 3,
+        projection_class_embeddings_input_dim: Optional[int] = None,
+        attention_type: str = "default",
+        class_embeddings_concat: bool = False,
+        mid_block_only_cross_attention: Optional[bool] = None,
+        cross_attention_norm: Optional[str] = None,
+        addition_embed_type_num_heads=64,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+
+        if num_attention_heads is not None:
+            raise ValueError(
+                "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
+            )
+
+        # If `num_attention_heads` is not defined (which is the case for most models)
+        # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+        # The reason for this behavior is to correct for incorrectly named variables that were introduced
+        # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+        # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+        # which is why we correct for the naming here.
+        num_attention_heads = num_attention_heads or attention_head_dim
+
+        # Check inputs
+        if len(down_block_types) != len(up_block_types):
+            raise ValueError(
+                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+            )
+
+        if len(block_out_channels) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(only_cross_attention, bool) and len(
+            only_cross_attention
+        ) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(
+            down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(
+            down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(
+            down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(layers_per_block, int) and len(layers_per_block) != len(
+            down_block_types
+        ):
+            raise ValueError(
+                f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+            )
+        if (
+            isinstance(transformer_layers_per_block, list)
+            and reverse_transformer_layers_per_block is None
+        ):
+            for layer_number_per_block in transformer_layers_per_block:
+                if isinstance(layer_number_per_block, list):
+                    raise ValueError(
+                        "Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet."
+                    )
+
+        # input
+        conv_in_padding = (conv_in_kernel - 1) // 2
+        self.conv_in = nn.Conv2d(
+            in_channels,
+            block_out_channels[0],
+            kernel_size=conv_in_kernel,
+            padding=conv_in_padding,
+        )
+
+        # time
+        if time_embedding_type == "fourier":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
+            if time_embed_dim % 2 != 0:
+                raise ValueError(
+                    f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}."
+                )
+            self.time_proj = GaussianFourierProjection(
+                time_embed_dim // 2,
+                set_W_to_weight=False,
+                log=False,
+                flip_sin_to_cos=flip_sin_to_cos,
+            )
+            timestep_input_dim = time_embed_dim
+        elif time_embedding_type == "positional":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
+
+            self.time_proj = Timesteps(
+                block_out_channels[0], flip_sin_to_cos, freq_shift
+            )
+            timestep_input_dim = block_out_channels[0]
+        else:
+            raise ValueError(
+                f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
+            )
+
+        self.time_embedding = TimestepEmbedding(
+            timestep_input_dim,
+            time_embed_dim,
+            act_fn=act_fn,
+            post_act_fn=timestep_post_act,
+            cond_proj_dim=time_cond_proj_dim,
+        )
+
+        if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+            encoder_hid_dim_type = "text_proj"
+            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+            logger.info(
+                "encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined."
+            )
+
+        if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+            )
+
+        if encoder_hid_dim_type == "text_proj":
+            self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+        elif encoder_hid_dim_type == "text_image_proj":
+            # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+            self.encoder_hid_proj = TextImageProjection(
+                text_embed_dim=encoder_hid_dim,
+                image_embed_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2
+            self.encoder_hid_proj = ImageProjection(
+                image_embed_dim=encoder_hid_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+            )
+        else:
+            self.encoder_hid_proj = None
+
+        # class embedding
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(
+                timestep_input_dim, time_embed_dim, act_fn=act_fn
+            )
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        elif class_embed_type == "projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+            # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+            # 2. it projects from an arbitrary input dimension.
+            #
+            # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+            # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+            # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+            self.class_embedding = TimestepEmbedding(
+                projection_class_embeddings_input_dim, time_embed_dim
+            )
+        elif class_embed_type == "simple_projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            self.class_embedding = nn.Linear(
+                projection_class_embeddings_input_dim, time_embed_dim
+            )
+        else:
+            self.class_embedding = None
+
+        if addition_embed_type == "text":
+            if encoder_hid_dim is not None:
+                text_time_embedding_from_dim = encoder_hid_dim
+            else:
+                text_time_embedding_from_dim = cross_attention_dim
+
+            self.add_embedding = TextTimeEmbedding(
+                text_time_embedding_from_dim,
+                time_embed_dim,
+                num_heads=addition_embed_type_num_heads,
+            )
+        elif addition_embed_type == "text_image":
+            # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
+            self.add_embedding = TextImageTimeEmbedding(
+                text_embed_dim=cross_attention_dim,
+                image_embed_dim=cross_attention_dim,
+                time_embed_dim=time_embed_dim,
+            )
+        elif addition_embed_type == "text_time":
+            self.add_time_proj = Timesteps(
+                addition_time_embed_dim, flip_sin_to_cos, freq_shift
+            )
+            self.add_embedding = TimestepEmbedding(
+                projection_class_embeddings_input_dim, time_embed_dim
+            )
+        elif addition_embed_type == "image":
+            # Kandinsky 2.2
+            self.add_embedding = ImageTimeEmbedding(
+                image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim
+            )
+        elif addition_embed_type == "image_hint":
+            # Kandinsky 2.2 ControlNet
+            self.add_embedding = ImageHintTimeEmbedding(
+                image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim
+            )
+        elif addition_embed_type is not None:
+            raise ValueError(
+                f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'."
+            )
+
+        if time_embedding_act_fn is None:
+            self.time_embed_act = None
+        else:
+            self.time_embed_act = get_activation(time_embedding_act_fn)
+
+        self.down_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            if mid_block_only_cross_attention is None:
+                mid_block_only_cross_attention = only_cross_attention
+
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if mid_block_only_cross_attention is None:
+            mid_block_only_cross_attention = False
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        if isinstance(cross_attention_dim, int):
+            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+        if isinstance(layers_per_block, int):
+            layers_per_block = [layers_per_block] * len(down_block_types)
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(
+                down_block_types
+            )
+
+        if class_embeddings_concat:
+            # The time embeddings are concatenated with the class embeddings. The dimension of the
+            # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+            # regular time embeddings
+            blocks_time_embed_dim = time_embed_dim * 2
+        else:
+            blocks_time_embed_dim = time_embed_dim
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block[i],
+                transformer_layers_per_block=transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim[i],
+                num_attention_heads=num_attention_heads[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=(
+                    attention_head_dim[i]
+                    if attention_head_dim[i] is not None
+                    else output_channel
+                ),
+                dropout=dropout,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        if mid_block_type == "UNetMidBlock2DCrossAttn":
+            self.mid_block = UNetMidBlock2DCrossAttn(
+                transformer_layers_per_block=transformer_layers_per_block[-1],
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                dropout=dropout,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                cross_attention_dim=cross_attention_dim[-1],
+                num_attention_heads=num_attention_heads[-1],
+                resnet_groups=norm_num_groups,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                upcast_attention=upcast_attention,
+                attention_type=attention_type,
+            )
+        elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
+            raise NotImplementedError(f"Unsupport mid_block_type: {mid_block_type}")
+        elif mid_block_type == "UNetMidBlock2D":
+            self.mid_block = UNetMidBlock2D(
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                dropout=dropout,
+                num_layers=0,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_groups=norm_num_groups,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                add_attention=False,
+            )
+        elif mid_block_type is None:
+            self.mid_block = None
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+        # count how many layers upsample the images
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_num_attention_heads = list(reversed(num_attention_heads))
+        reversed_layers_per_block = list(reversed(layers_per_block))
+        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+        reversed_transformer_layers_per_block = (
+            list(reversed(transformer_layers_per_block))
+            if reverse_transformer_layers_per_block is None
+            else reverse_transformer_layers_per_block
+        )
+        only_cross_attention = list(reversed(only_cross_attention))
+
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[
+                min(i + 1, len(block_out_channels) - 1)
+            ]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=reversed_layers_per_block[i] + 1,
+                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resolution_idx=i,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=reversed_cross_attention_dim[i],
+                num_attention_heads=reversed_num_attention_heads[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=(
+                    attention_head_dim[i]
+                    if attention_head_dim[i] is not None
+                    else output_channel
+                ),
+                dropout=dropout,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        if norm_num_groups is not None:
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=block_out_channels[0],
+                num_groups=norm_num_groups,
+                eps=norm_eps,
+            )
+
+            self.conv_act = get_activation(act_fn)
+
+        else:
+            self.conv_norm_out = None
+            self.conv_act = None
+        self.conv_norm_out = None
+
+        conv_out_padding = (conv_out_kernel - 1) // 2
+        # self.conv_out = nn.Conv2d(
+        #     block_out_channels[0],
+        #     out_channels,
+        #     kernel_size=conv_out_kernel,
+        #     padding=conv_out_padding,
+        # )
+
+        if attention_type in ["gated", "gated-text-image"]:
+            positive_len = 768
+            if isinstance(cross_attention_dim, int):
+                positive_len = cross_attention_dim
+            elif isinstance(cross_attention_dim, tuple) or isinstance(
+                cross_attention_dim, list
+            ):
+                positive_len = cross_attention_dim[0]
+
+            feature_type = "text-only" if attention_type == "gated" else "text-image"
+            self.position_net = PositionNet(
+                positive_len=positive_len,
+                out_dim=cross_attention_dim,
+                feature_type=feature_type,
+            )
+
+    @property
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(
+            name: str,
+            module: torch.nn.Module,
+            processors: Dict[str, AttentionProcessor],
+        ):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor(
+                    return_deprecated_lora=True
+                )
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attn_processor(
+        self,
+        processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]],
+        _remove_lora=False,
+    ):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor, _remove_lora=_remove_lora)
+                else:
+                    module.set_processor(
+                        processor.pop(f"{name}.processor"), _remove_lora=_remove_lora
+                    )
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(
+            proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS
+            for proc in self.attn_processors.values()
+        ):
+            processor = AttnAddedKVProcessor()
+        elif all(
+            proc.__class__ in CROSS_ATTENTION_PROCESSORS
+            for proc in self.attn_processors.values()
+        ):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor, _remove_lora=True)
+
+    def set_attention_slice(self, slice_size):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+        several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_sliceable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_sliceable_dims(module)
+
+        num_sliceable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_sliceable_layers * [1]
+
+        slice_size = (
+            num_sliceable_layers * [slice_size]
+            if not isinstance(slice_size, list)
+            else slice_size
+        )
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(
+            module: torch.nn.Module, slice_size: List[int]
+        ):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def enable_freeu(self, s1, s2, b1, b2):
+        r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stage blocks where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
+        are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        for i, upsample_block in enumerate(self.up_blocks):
+            setattr(upsample_block, "s1", s1)
+            setattr(upsample_block, "s2", s2)
+            setattr(upsample_block, "b1", b1)
+            setattr(upsample_block, "b2", b2)
+
+    def disable_freeu(self):
+        """Disables the FreeU mechanism."""
+        freeu_keys = {"s1", "s2", "b1", "b2"}
+        for i, upsample_block in enumerate(self.up_blocks):
+            for k in freeu_keys:
+                if (
+                    hasattr(upsample_block, k)
+                    or getattr(upsample_block, k, None) is not None
+                ):
+                    setattr(upsample_block, k, None)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[UNet2DConditionOutput, Tuple]:
+        r"""
+        The [`UNet2DConditionModel`] forward method.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The noisy input tensor with the following shape `(batch, channel, height, width)`.
+            timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.FloatTensor`):
+                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
+                Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
+                through the `self.time_embedding` layer to obtain the timestep embeddings.
+            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
+                A tuple of tensors that if specified are added to the residuals of down unet blocks.
+            mid_block_additional_residual: (`torch.Tensor`, *optional*):
+                A tensor that if specified is added to the residual of the middle unet block.
+            encoder_attention_mask (`torch.Tensor`):
+                A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
+                `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
+                which adds large negative values to the attention scores corresponding to "discard" tokens.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added to UNet long skip connections from down blocks to up blocks for
+                example from ControlNet side model(s)
+            mid_block_additional_residual (`torch.Tensor`, *optional*):
+                additional residual to be added to UNet mid block output, for example from ControlNet side model
+            down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
+
+        Returns:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
+                a `tuple` is returned where the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        for dim in sample.shape[-2:]:
+            if dim % default_overall_up_factor != 0:
+                # Forward upsample size to force interpolation output size.
+                forward_upsample_size = True
+                break
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None:
+            encoder_attention_mask = (
+                1 - encoder_attention_mask.to(sample.dtype)
+            ) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 0. center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+
+        # `Timesteps` does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=sample.dtype)
+
+        emb = self.time_embedding(t_emb, timestep_cond)
+        aug_emb = None
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError(
+                    "class_labels should be provided when num_class_embeds > 0"
+                )
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+                # `Timesteps` does not contain any weights and will always return f32 tensors
+                # there might be better ways to encapsulate this.
+                class_labels = class_labels.to(dtype=sample.dtype)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
+
+            if self.config.class_embeddings_concat:
+                emb = torch.cat([emb, class_emb], dim=-1)
+            else:
+                emb = emb + class_emb
+
+        if self.config.addition_embed_type == "text":
+            aug_emb = self.add_embedding(encoder_hidden_states)
+        elif self.config.addition_embed_type == "text_image":
+            # Kandinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+
+            image_embs = added_cond_kwargs.get("image_embeds")
+            text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
+            aug_emb = self.add_embedding(text_embs, image_embs)
+        elif self.config.addition_embed_type == "text_time":
+            # SDXL - style
+            if "text_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+                )
+            text_embeds = added_cond_kwargs.get("text_embeds")
+            if "time_ids" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+                )
+            time_ids = added_cond_kwargs.get("time_ids")
+            time_embeds = self.add_time_proj(time_ids.flatten())
+            time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+            add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+            add_embeds = add_embeds.to(emb.dtype)
+            aug_emb = self.add_embedding(add_embeds)
+        elif self.config.addition_embed_type == "image":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            aug_emb = self.add_embedding(image_embs)
+        elif self.config.addition_embed_type == "image_hint":
+            # Kandinsky 2.2 - style
+            if (
+                "image_embeds" not in added_cond_kwargs
+                or "hint" not in added_cond_kwargs
+            ):
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            hint = added_cond_kwargs.get("hint")
+            aug_emb, hint = self.add_embedding(image_embs, hint)
+            sample = torch.cat([sample, hint], dim=1)
+
+        emb = emb + aug_emb if aug_emb is not None else emb
+
+        if self.time_embed_act is not None:
+            emb = self.time_embed_act(emb)
+
+        if (
+            self.encoder_hid_proj is not None
+            and self.config.encoder_hid_dim_type == "text_proj"
+        ):
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+        elif (
+            self.encoder_hid_proj is not None
+            and self.config.encoder_hid_dim_type == "text_image_proj"
+        ):
+            # Kadinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(
+                encoder_hidden_states, image_embeds
+            )
+        elif (
+            self.encoder_hid_proj is not None
+            and self.config.encoder_hid_dim_type == "image_proj"
+        ):
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(image_embeds)
+        elif (
+            self.encoder_hid_proj is not None
+            and self.config.encoder_hid_dim_type == "ip_image_proj"
+        ):
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            image_embeds = self.encoder_hid_proj(image_embeds).to(
+                encoder_hidden_states.dtype
+            )
+            encoder_hidden_states = torch.cat(
+                [encoder_hidden_states, image_embeds], dim=1
+            )
+
+        # 2. pre-process
+        sample = self.conv_in(sample)
+
+        # 2.5 GLIGEN position net
+        if (
+            cross_attention_kwargs is not None
+            and cross_attention_kwargs.get("gligen", None) is not None
+        ):
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            gligen_args = cross_attention_kwargs.pop("gligen")
+            cross_attention_kwargs["gligen"] = {
+                "objs": self.position_net(**gligen_args)
+            }
+
+        # 3. down
+        lora_scale = (
+            cross_attention_kwargs.get("scale", 1.0)
+            if cross_attention_kwargs is not None
+            else 1.0
+        )
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+
+        is_controlnet = (
+            mid_block_additional_residual is not None
+            and down_block_additional_residuals is not None
+        )
+        # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
+        is_adapter = down_intrablock_additional_residuals is not None
+        # maintain backward compatibility for legacy usage, where
+        #       T2I-Adapter and ControlNet both use down_block_additional_residuals arg
+        #       but can only use one or the other
+        if (
+            not is_adapter
+            and mid_block_additional_residual is None
+            and down_block_additional_residuals is not None
+        ):
+            deprecate(
+                "T2I should not use down_block_additional_residuals",
+                "1.3.0",
+                "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
+                       and will be removed in diffusers 1.3.0.  `down_block_additional_residuals` should only be used \
+                       for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
+                standard_warn=False,
+            )
+            down_intrablock_additional_residuals = down_block_additional_residuals
+            is_adapter = True
+
+        down_block_res_samples = (sample,)
+        tot_referece_features = ()
+        for downsample_block in self.down_blocks:
+            if (
+                hasattr(downsample_block, "has_cross_attention")
+                and downsample_block.has_cross_attention
+            ):
+                # For t2i-adapter CrossAttnDownBlock2D
+                additional_residuals = {}
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    additional_residuals["additional_residuals"] = (
+                        down_intrablock_additional_residuals.pop(0)
+                    )
+
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                    **additional_residuals,
+                )
+            else:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample, temb=emb, scale=lora_scale
+                )
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    sample += down_intrablock_additional_residuals.pop(0)
+
+            down_block_res_samples += res_samples
+
+        if is_controlnet:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_res_sample = (
+                    down_block_res_sample + down_block_additional_residual
+                )
+                new_down_block_res_samples = new_down_block_res_samples + (
+                    down_block_res_sample,
+                )
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            if (
+                hasattr(self.mid_block, "has_cross_attention")
+                and self.mid_block.has_cross_attention
+            ):
+                sample = self.mid_block(
+                    sample,
+                    emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = self.mid_block(sample, emb)
+
+            # To support T2I-Adapter-XL
+            if (
+                is_adapter
+                and len(down_intrablock_additional_residuals) > 0
+                and sample.shape == down_intrablock_additional_residuals[0].shape
+            ):
+                sample += down_intrablock_additional_residuals.pop(0)
+
+        if is_controlnet:
+            sample = sample + mid_block_additional_residual
+
+        # 5. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[
+                : -len(upsample_block.resnets)
+            ]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if (
+                hasattr(upsample_block, "has_cross_attention")
+                and upsample_block.has_cross_attention
+            ):
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
+                    scale=lora_scale,
+                )
+
+        # 6. post-process
+        # if self.conv_norm_out:
+        #     sample = self.conv_norm_out(sample)
+        #     sample = self.conv_act(sample)
+        # sample = self.conv_out(sample)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet2DConditionOutput(sample=sample)

2610_variant_runtime/diffusion/models/unet_3d.py

@@ -0,0 +1,668 @@
+# Adapted from https://github.com/guoyww/AnimateDiff/blob/main/animatediff/models/unet_blocks.py
+
+from collections import OrderedDict
+from dataclasses import dataclass
+from os import PathLike
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models.attention_processor import AttentionProcessor
+from diffusers.models.embeddings import TimestepEmbedding, Timesteps
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.utils import SAFETENSORS_WEIGHTS_NAME, WEIGHTS_NAME, BaseOutput, logging
+from safetensors.torch import load_file
+
+from .resnet import InflatedConv3d, InflatedGroupNorm
+from .unet_3d_blocks import UNetMidBlock3DCrossAttn, get_down_block, get_up_block
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class UNet3DConditionOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+class UNet3DConditionModel(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 4,
+        out_channels: int = 4,
+        center_input_sample: bool = False,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "DownBlock3D",
+        ),
+        mid_block_type: str = "UNetMidBlock3DCrossAttn",
+        up_block_types: Tuple[str] = (
+            "UpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+        ),
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        layers_per_block: int = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        act_fn: str = "silu",
+        norm_num_groups: int = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: int = 1280,
+        attention_head_dim: Union[int, Tuple[int]] = 8,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        use_inflated_groupnorm=False,
+        # Additional
+        use_motion_module=False,
+        motion_module_resolutions=(1, 2, 4, 8),
+        motion_module_mid_block=False,
+        motion_module_decoder_only=False,
+        motion_module_type=None,
+        motion_module_kwargs={},
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+        time_embed_dim = block_out_channels[0] * 4
+
+        # input
+        self.conv_in = InflatedConv3d(
+            in_channels, block_out_channels[0], kernel_size=3, padding=(1, 1)
+        )
+
+        # time
+        self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+        timestep_input_dim = block_out_channels[0]
+
+        self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+
+        # class embedding
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        else:
+            self.class_embedding = None
+
+        self.down_blocks = nn.ModuleList([])
+        self.mid_block = None
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            res = 2**i
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=attention_head_dim[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                unet_use_temporal_attention=unet_use_temporal_attention,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+                use_motion_module=use_motion_module
+                and (res in motion_module_resolutions)
+                and (not motion_module_decoder_only),
+                motion_module_type=motion_module_type,
+                motion_module_kwargs=motion_module_kwargs,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        if mid_block_type == "UNetMidBlock3DCrossAttn":
+            self.mid_block = UNetMidBlock3DCrossAttn(
+                in_channels=block_out_channels[-1],
+                temb_channels=time_embed_dim,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=attention_head_dim[-1],
+                resnet_groups=norm_num_groups,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                upcast_attention=upcast_attention,
+                unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                unet_use_temporal_attention=unet_use_temporal_attention,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+                use_motion_module=use_motion_module and motion_module_mid_block,
+                motion_module_type=motion_module_type,
+                motion_module_kwargs=motion_module_kwargs,
+            )
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+        # count how many layers upsample the videos
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_attention_head_dim = list(reversed(attention_head_dim))
+        only_cross_attention = list(reversed(only_cross_attention))
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            res = 2 ** (3 - i)
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[
+                min(i + 1, len(block_out_channels) - 1)
+            ]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=layers_per_block + 1,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim,
+                attn_num_head_channels=reversed_attention_head_dim[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                unet_use_temporal_attention=unet_use_temporal_attention,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+                use_motion_module=use_motion_module
+                and (res in motion_module_resolutions),
+                motion_module_type=motion_module_type,
+                motion_module_kwargs=motion_module_kwargs,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        if use_inflated_groupnorm:
+            self.conv_norm_out = InflatedGroupNorm(
+                num_channels=block_out_channels[0],
+                num_groups=norm_num_groups,
+                eps=norm_eps,
+            )
+        else:
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=block_out_channels[0],
+                num_groups=norm_num_groups,
+                eps=norm_eps,
+            )
+        self.conv_act = nn.SiLU()
+        self.conv_out = InflatedConv3d(
+            block_out_channels[0], out_channels, kernel_size=3, padding=1
+        )
+
+    @property
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(
+            name: str,
+            module: torch.nn.Module,
+            processors: Dict[str, AttentionProcessor],
+        ):
+            if hasattr(module, "set_processor"):
+                processors[f"{name}.processor"] = module.processor
+
+            for sub_name, child in module.named_children():
+                if "temporal_transformer" not in sub_name:
+                    fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            if "temporal_transformer" not in name:
+                fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attention_slice(self, slice_size):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
+        in several steps. This is useful to save some memory in exchange for a small speed decrease.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+                `"max"`, maxium amount of memory will be saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_slicable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_slicable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_slicable_dims(module)
+
+        num_slicable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_slicable_layers * [1]
+
+        slice_size = (
+            num_slicable_layers * [slice_size]
+            if not isinstance(slice_size, list)
+            else slice_size
+        )
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(
+            module: torch.nn.Module, slice_size: List[int]
+        ):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(
+        self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]
+    ):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                if "temporal_transformer" not in sub_name:
+                    fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            if "temporal_transformer" not in name:
+                fn_recursive_attn_processor(name, module, processor)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        class_labels: Optional[torch.Tensor] = None,
+        guidance_fea: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[UNet3DConditionOutput, Tuple]:
+        r"""
+        Args:
+            sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor
+            timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps
+            encoder_hidden_states (`torch.FloatTensor`): (batch, sequence_length, feature_dim) encoder hidden states
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+            [`~models.unet_2d_condition.UNet2DConditionOutput`] if `return_dict` is True, otherwise a `tuple`. When
+            returning a tuple, the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
+            logger.info("Forward upsample size to force interpolation output size.")
+            forward_upsample_size = True
+
+        # prepare attention_mask
+        if attention_mask is not None:
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+
+        # timesteps does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=self.dtype)
+        emb = self.time_embedding(t_emb)
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError(
+                    "class_labels should be provided when num_class_embeds > 0"
+                )
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
+            emb = emb + class_emb
+
+        # pre-process
+        sample = self.conv_in(sample)
+        if guidance_fea is not None:
+            sample = sample + guidance_fea
+
+        # down
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if (
+                hasattr(downsample_block, "has_cross_attention")
+                and downsample_block.has_cross_attention
+            ):
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                )
+            else:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                )
+
+            down_block_res_samples += res_samples
+
+        if down_block_additional_residuals is not None:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_res_sample = (
+                    down_block_res_sample + down_block_additional_residual
+                )
+                new_down_block_res_samples += (down_block_res_sample,)
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # mid
+        sample = self.mid_block(
+            sample,
+            emb,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=attention_mask,
+        )
+
+        if mid_block_additional_residual is not None:
+            sample = sample + mid_block_additional_residual
+
+        # up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[
+                : -len(upsample_block.resnets)
+            ]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if (
+                hasattr(upsample_block, "has_cross_attention")
+                and upsample_block.has_cross_attention
+            ):
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
+                    encoder_hidden_states=encoder_hidden_states,
+                )
+
+        # post-process
+        sample = self.conv_norm_out(sample)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet3DConditionOutput(sample=sample)
+
+    @classmethod
+    def from_pretrained_2d(
+        cls,
+        pretrained_model_path: PathLike,
+        motion_module_path: PathLike,
+        subfolder=None,
+        unet_additional_kwargs=None,
+        mm_zero_proj_out=False,
+    ):
+        pretrained_model_path = Path(pretrained_model_path)
+        motion_module_path = Path(motion_module_path)
+        if subfolder is not None:
+            pretrained_model_path = pretrained_model_path.joinpath(subfolder)
+        logger.info(
+            f"loaded temporal unet's pretrained weights from {pretrained_model_path} ..."
+        )
+
+        config_file = pretrained_model_path / "config.json"
+        if not (config_file.exists() and config_file.is_file()):
+            raise RuntimeError(f"{config_file} does not exist or is not a file")
+
+        unet_config = cls.load_config(config_file)
+        unet_config["_class_name"] = cls.__name__
+        unet_config["down_block_types"] = [
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "CrossAttnDownBlock3D",
+            "DownBlock3D",
+        ]
+        unet_config["up_block_types"] = [
+            "UpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+            "CrossAttnUpBlock3D",
+        ]
+        unet_config["mid_block_type"] = "UNetMidBlock3DCrossAttn"
+
+        model = cls.from_config(unet_config, **unet_additional_kwargs)
+        # load the vanilla weights
+        if pretrained_model_path.joinpath(SAFETENSORS_WEIGHTS_NAME).exists():
+            logger.debug(
+                f"loading safeTensors weights from {pretrained_model_path} ..."
+            )
+            state_dict = load_file(
+                pretrained_model_path.joinpath(SAFETENSORS_WEIGHTS_NAME), device="cpu"
+            )
+
+        elif pretrained_model_path.joinpath(WEIGHTS_NAME).exists():
+            logger.debug(f"loading weights from {pretrained_model_path} ...")
+            state_dict = torch.load(
+                pretrained_model_path.joinpath(WEIGHTS_NAME),
+                map_location="cpu",
+                weights_only=True,
+            )
+        else:
+            raise FileNotFoundError(f"no weights file found in {pretrained_model_path}")
+
+        # load the motion module weights
+        if motion_module_path.exists() and motion_module_path.is_file():
+            if motion_module_path.suffix.lower() in [".pth", ".pt", ".ckpt"]:
+                logger.info(f"Load motion module params from {motion_module_path}")
+                motion_state_dict = torch.load(
+                    motion_module_path, map_location="cpu", weights_only=True
+                )
+            elif motion_module_path.suffix.lower() == ".safetensors":
+                motion_state_dict = load_file(motion_module_path, device="cpu")
+            else:
+                raise RuntimeError(
+                    f"unknown file format for motion module weights: {motion_module_path.suffix}"
+                )
+            if mm_zero_proj_out:
+                logger.info(f"Zero initialize proj_out layers in motion module...")
+                new_motion_state_dict = OrderedDict()
+                for k in motion_state_dict:
+                    if "proj_out" in k:
+                        continue
+                    new_motion_state_dict[k] = motion_state_dict[k]
+                motion_state_dict = new_motion_state_dict
+
+            # merge the state dicts
+            state_dict.update(motion_state_dict)
+
+        # load the weights into the model
+        m, u = model.load_state_dict(state_dict, strict=False)
+        logger.debug(f"### missing keys: {len(m)}; \n### unexpected keys: {len(u)};")
+
+        params = [
+            p.numel() if "temporal" in n else 0 for n, p in model.named_parameters()
+        ]
+        logger.info(f"Loaded {sum(params) / 1e6}M-parameter motion module")
+
+        return model

2610_variant_runtime/diffusion/models/unet_3d_blocks.py

@@ -0,0 +1,863 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_blocks.py
+
+import pdb
+
+import torch
+from torch import nn
+
+from .motion_module import get_motion_module
+
+# from .motion_module import get_motion_module
+from .resnet import Downsample3D, ResnetBlock3D, Upsample3D
+from .transformer_3d import Transformer3DModel
+
+
+def get_down_block(
+    down_block_type,
+    num_layers,
+    in_channels,
+    out_channels,
+    temb_channels,
+    add_downsample,
+    resnet_eps,
+    resnet_act_fn,
+    attn_num_head_channels,
+    resnet_groups=None,
+    cross_attention_dim=None,
+    downsample_padding=None,
+    dual_cross_attention=False,
+    use_linear_projection=False,
+    only_cross_attention=False,
+    upcast_attention=False,
+    resnet_time_scale_shift="default",
+    unet_use_cross_frame_attention=None,
+    unet_use_temporal_attention=None,
+    use_inflated_groupnorm=None,
+    use_motion_module=None,
+    motion_module_type=None,
+    motion_module_kwargs=None,
+):
+    down_block_type = (
+        down_block_type[7:]
+        if down_block_type.startswith("UNetRes")
+        else down_block_type
+    )
+    if down_block_type == "DownBlock3D":
+        return DownBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_type=motion_module_type,
+            motion_module_kwargs=motion_module_kwargs,
+        )
+    elif down_block_type == "CrossAttnDownBlock3D":
+        if cross_attention_dim is None:
+            raise ValueError(
+                "cross_attention_dim must be specified for CrossAttnDownBlock3D"
+            )
+        return CrossAttnDownBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            cross_attention_dim=cross_attention_dim,
+            attn_num_head_channels=attn_num_head_channels,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+            unet_use_temporal_attention=unet_use_temporal_attention,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_type=motion_module_type,
+            motion_module_kwargs=motion_module_kwargs,
+        )
+    raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_up_block(
+    up_block_type,
+    num_layers,
+    in_channels,
+    out_channels,
+    prev_output_channel,
+    temb_channels,
+    add_upsample,
+    resnet_eps,
+    resnet_act_fn,
+    attn_num_head_channels,
+    resnet_groups=None,
+    cross_attention_dim=None,
+    dual_cross_attention=False,
+    use_linear_projection=False,
+    only_cross_attention=False,
+    upcast_attention=False,
+    resnet_time_scale_shift="default",
+    unet_use_cross_frame_attention=None,
+    unet_use_temporal_attention=None,
+    use_inflated_groupnorm=None,
+    use_motion_module=None,
+    motion_module_type=None,
+    motion_module_kwargs=None,
+):
+    up_block_type = (
+        up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
+    )
+    if up_block_type == "UpBlock3D":
+        return UpBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_type=motion_module_type,
+            motion_module_kwargs=motion_module_kwargs,
+        )
+    elif up_block_type == "CrossAttnUpBlock3D":
+        if cross_attention_dim is None:
+            raise ValueError(
+                "cross_attention_dim must be specified for CrossAttnUpBlock3D"
+            )
+        return CrossAttnUpBlock3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            cross_attention_dim=cross_attention_dim,
+            attn_num_head_channels=attn_num_head_channels,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+            unet_use_temporal_attention=unet_use_temporal_attention,
+            use_inflated_groupnorm=use_inflated_groupnorm,
+            use_motion_module=use_motion_module,
+            motion_module_type=motion_module_type,
+            motion_module_kwargs=motion_module_kwargs,
+        )
+    raise ValueError(f"{up_block_type} does not exist.")
+
+
+class UNetMidBlock3DCrossAttn(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        attn_num_head_channels=1,
+        output_scale_factor=1.0,
+        cross_attention_dim=1280,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        upcast_attention=False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        use_inflated_groupnorm=None,
+        use_motion_module=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+    ):
+        super().__init__()
+
+        self.has_cross_attention = True
+        self.attn_num_head_channels = attn_num_head_channels
+        resnet_groups = (
+            resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+        )
+
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock3D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+                use_inflated_groupnorm=use_inflated_groupnorm,
+            )
+        ]
+        attentions = []
+        motion_modules = []
+
+        for _ in range(num_layers):
+            if dual_cross_attention:
+                raise NotImplementedError
+            attentions.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    in_channels // attn_num_head_channels,
+                    in_channels=in_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    upcast_attention=upcast_attention,
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=in_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+    def forward(
+        self,
+        hidden_states,
+        temb=None,
+        encoder_hidden_states=None,
+        attention_mask=None,
+    ):
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for attn, resnet, motion_module in zip(
+            self.attentions, self.resnets[1:], self.motion_modules
+        ):
+            hidden_states = attn(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+            ).sample
+            hidden_states = (
+                motion_module(
+                    hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                )
+                if motion_module is not None
+                else hidden_states
+            )
+            hidden_states = resnet(hidden_states, temb)
+
+        return hidden_states
+
+
+class CrossAttnDownBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        attn_num_head_channels=1,
+        cross_attention_dim=1280,
+        output_scale_factor=1.0,
+        downsample_padding=1,
+        add_downsample=True,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        only_cross_attention=False,
+        upcast_attention=False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        use_inflated_groupnorm=None,
+        use_motion_module=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+        motion_modules = []
+
+        self.has_cross_attention = True
+        self.attn_num_head_channels = attn_num_head_channels
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+            if dual_cross_attention:
+                raise NotImplementedError
+            attentions.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    out_channels // attn_num_head_channels,
+                    in_channels=out_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=out_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample3D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        temb=None,
+        encoder_hidden_states=None,
+        attention_mask=None,
+    ):
+        output_states = ()
+
+        for i, (resnet, attn, motion_module) in enumerate(
+            zip(self.resnets, self.attentions, self.motion_modules)
+        ):
+            # self.gradient_checkpointing = False
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(attn, return_dict=False),
+                    hidden_states,
+                    encoder_hidden_states,
+                    use_reentrant=False,
+                )[0]
+
+                # add motion module
+                hidden_states = (
+                    motion_module(
+                        hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                    )
+                    if motion_module is not None
+                    else hidden_states
+                )
+
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                ).sample
+
+                # add motion module
+                hidden_states = (
+                    motion_module(
+                        hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                    )
+                    if motion_module is not None
+                    else hidden_states
+                )
+
+            output_states += (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+            output_states += (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class DownBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor=1.0,
+        add_downsample=True,
+        downsample_padding=1,
+        use_inflated_groupnorm=None,
+        use_motion_module=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+    ):
+        super().__init__()
+        resnets = []
+        motion_modules = []
+
+        # use_motion_module = False
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=out_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample3D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(self, hidden_states, temb=None, encoder_hidden_states=None):
+        output_states = ()
+
+        for resnet, motion_module in zip(self.resnets, self.motion_modules):
+            # print(f"DownBlock3D {self.gradient_checkpointing = }")
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb
+                )
+                if motion_module is not None:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(motion_module),
+                        hidden_states.requires_grad_(),
+                        temb,
+                        encoder_hidden_states,
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb)
+
+                # add motion module
+                hidden_states = (
+                    motion_module(
+                        hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                    )
+                    if motion_module is not None
+                    else hidden_states
+                )
+
+            output_states += (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+            output_states += (hidden_states,)
+
+        return hidden_states, output_states
+
+
+class CrossAttnUpBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        prev_output_channel: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        attn_num_head_channels=1,
+        cross_attention_dim=1280,
+        output_scale_factor=1.0,
+        add_upsample=True,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        only_cross_attention=False,
+        upcast_attention=False,
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+        use_motion_module=None,
+        use_inflated_groupnorm=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+        motion_modules = []
+
+        self.has_cross_attention = True
+        self.attn_num_head_channels = attn_num_head_channels
+
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+            if dual_cross_attention:
+                raise NotImplementedError
+            attentions.append(
+                Transformer3DModel(
+                    attn_num_head_channels,
+                    out_channels // attn_num_head_channels,
+                    in_channels=out_channels,
+                    num_layers=1,
+                    cross_attention_dim=cross_attention_dim,
+                    norm_num_groups=resnet_groups,
+                    use_linear_projection=use_linear_projection,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=out_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [Upsample3D(out_channels, use_conv=True, out_channels=out_channels)]
+            )
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        res_hidden_states_tuple,
+        temb=None,
+        encoder_hidden_states=None,
+        upsample_size=None,
+        attention_mask=None,
+    ):
+        for i, (resnet, attn, motion_module) in enumerate(
+            zip(self.resnets, self.attentions, self.motion_modules)
+        ):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb
+                )
+                hidden_states = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                ).sample
+                if motion_module is not None:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(motion_module),
+                        hidden_states.requires_grad_(),
+                        temb,
+                        encoder_hidden_states,
+                    )
+
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                ).sample
+
+                # add motion module
+                hidden_states = (
+                    motion_module(
+                        hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                    )
+                    if motion_module is not None
+                    else hidden_states
+                )
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size)
+
+        return hidden_states
+
+
+class UpBlock3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        prev_output_channel: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor=1.0,
+        add_upsample=True,
+        use_inflated_groupnorm=None,
+        use_motion_module=None,
+        motion_module_type=None,
+        motion_module_kwargs=None,
+    ):
+        super().__init__()
+        resnets = []
+        motion_modules = []
+
+        # use_motion_module = False
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock3D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                    use_inflated_groupnorm=use_inflated_groupnorm,
+                )
+            )
+            motion_modules.append(
+                get_motion_module(
+                    in_channels=out_channels,
+                    motion_module_type=motion_module_type,
+                    motion_module_kwargs=motion_module_kwargs,
+                )
+                if use_motion_module
+                else None
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+        self.motion_modules = nn.ModuleList(motion_modules)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [Upsample3D(out_channels, use_conv=True, out_channels=out_channels)]
+            )
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        res_hidden_states_tuple,
+        temb=None,
+        upsample_size=None,
+        encoder_hidden_states=None,
+    ):
+        for resnet, motion_module in zip(self.resnets, self.motion_modules):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            # print(f"UpBlock3D {self.gradient_checkpointing = }")
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs)
+
+                    return custom_forward
+
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(resnet), hidden_states, temb
+                )
+                if motion_module is not None:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(motion_module),
+                        hidden_states.requires_grad_(),
+                        temb,
+                        encoder_hidden_states,
+                    )
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = (
+                    motion_module(
+                        hidden_states, temb, encoder_hidden_states=encoder_hidden_states
+                    )
+                    if motion_module is not None
+                    else hidden_states
+                )
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size)
+
+        return hidden_states

2610_variant_runtime/diffusion/pipelines/__init__.py

@@ -0,0 +1 @@
+

2610_variant_runtime/diffusion/pipelines/pipeline_vfi2video_share_vae.py

@@ -0,0 +1,509 @@
+import inspect
+from dataclasses import dataclass
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import torch
+from diffusers import DiffusionPipeline
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.schedulers import (
+    DDIMScheduler,
+    DPMSolverMultistepScheduler,
+    EulerAncestralDiscreteScheduler,
+    EulerDiscreteScheduler,
+    LMSDiscreteScheduler,
+    PNDMScheduler,
+)
+from diffusers.utils import BaseOutput, is_accelerate_available
+from diffusers.utils.torch_utils import randn_tensor
+from einops import rearrange
+from tqdm import tqdm
+from transformers import CLIPImageProcessor
+from einops import rearrange
+
+from models.mutual_self_attention import ReferenceAttentionControl
+
+
+@dataclass
+class Pose2VideoPipelineOutput(BaseOutput):
+    videos: Union[torch.Tensor, np.ndarray]
+
+class Pose2VideoPipeline(DiffusionPipeline):
+    _optional_components = []
+
+    def __init__(
+        self,
+        vae,
+        image_encoder,
+        reference_unet,
+        denoising_unet,
+        pose_guider,
+        seed_guider,
+        scheduler: Union[
+            DDIMScheduler,
+            PNDMScheduler,
+            LMSDiscreteScheduler,
+            EulerDiscreteScheduler,
+            EulerAncestralDiscreteScheduler,
+            DPMSolverMultistepScheduler,
+        ],
+        image_proj_model=None,
+        tokenizer=None,
+        text_encoder=None,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            image_encoder=image_encoder,
+            reference_unet=reference_unet,
+            denoising_unet=denoising_unet,
+            pose_guider=pose_guider,
+            scheduler=scheduler,
+            seed_guider=seed_guider,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.clip_image_processor = CLIPImageProcessor()
+        self.ref_image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True
+        )
+        self.cond_image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor,
+            do_convert_rgb=True,
+            do_normalize=False,
+        )
+
+
+    def enable_vae_slicing(self):
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        self.vae.disable_slicing()
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+    @property
+    def _execution_device(self):
+        if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    def decode_latents(self, latents):
+        video_length = latents.shape[2]
+        latents = 1 / 0.18215 * latents
+        latents = rearrange(latents, "b c f h w -> (b f) c h w")
+        video = []
+        for frame_idx in tqdm(range(latents.shape[0])):
+            video.append(self.vae.decode(latents[frame_idx : frame_idx + 1]).sample)
+        video = torch.cat(video)
+        video = rearrange(video, "(b f) c h w -> b c f h w", f=video_length)
+        video = (video / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
+        video = video.cpu().float().numpy()
+        return video
+
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(
+            inspect.signature(self.scheduler.step).parameters.keys()
+        )
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(
+            inspect.signature(self.scheduler.step).parameters.keys()
+        )
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        width,
+        height,
+        video_length,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        shape = (
+            batch_size,
+            num_channels_latents,
+            video_length,
+            height // self.vae_scale_factor,
+            width // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(
+                shape, generator=generator, device=device, dtype=dtype
+            )
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def _encode_prompt(
+        self,
+        prompt,
+        device,
+        num_videos_per_prompt,
+        do_classifier_free_guidance,
+        negative_prompt,
+    ):
+        batch_size = len(prompt) if isinstance(prompt, list) else 1
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=self.tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer(
+            prompt, padding="longest", return_tensors="pt"
+        ).input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+            text_input_ids, untruncated_ids
+        ):
+            removed_text = self.tokenizer.batch_decode(
+                untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+            )
+
+        if (
+            hasattr(self.text_encoder.config, "use_attention_mask")
+            and self.text_encoder.config.use_attention_mask
+        ):
+            attention_mask = text_inputs.attention_mask.to(device)
+        else:
+            attention_mask = None
+
+        text_embeddings = self.text_encoder(
+            text_input_ids.to(device),
+            attention_mask=attention_mask,
+        )
+        text_embeddings = text_embeddings[0]
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        bs_embed, seq_len, _ = text_embeddings.shape
+        text_embeddings = text_embeddings.repeat(1, num_videos_per_prompt, 1)
+        text_embeddings = text_embeddings.view(
+            bs_embed * num_videos_per_prompt, seq_len, -1
+        )
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""] * batch_size
+            elif type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            max_length = text_input_ids.shape[-1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+
+            if (
+                hasattr(self.text_encoder.config, "use_attention_mask")
+                and self.text_encoder.config.use_attention_mask
+            ):
+                attention_mask = uncond_input.attention_mask.to(device)
+            else:
+                attention_mask = None
+
+            uncond_embeddings = self.text_encoder(
+                uncond_input.input_ids.to(device),
+                attention_mask=attention_mask,
+            )
+            uncond_embeddings = uncond_embeddings[0]
+
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = uncond_embeddings.shape[1]
+            uncond_embeddings = uncond_embeddings.repeat(1, num_videos_per_prompt, 1)
+            uncond_embeddings = uncond_embeddings.view(
+                batch_size * num_videos_per_prompt, seq_len, -1
+            )
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+        return text_embeddings
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        ref_image,
+        pose_images,
+        seed_images,
+        width,
+        height,
+        video_length,
+        num_inference_steps,
+        guidance_scale,
+        num_images_per_prompt=1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        output_type: Optional[str] = "tensor",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: Optional[int] = 1,
+        seed_mask=None,
+        **kwargs,
+    ):
+        # Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+        device = self._execution_device
+
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        batch_size = 1
+
+        # Prepare clip image embeds
+        clip_image = self.clip_image_processor.preprocess(
+            ref_image.resize((224, 224)), return_tensors="pt"
+        ).pixel_values
+        clip_image_embeds = self.image_encoder(
+            clip_image.to(device, dtype=self.image_encoder.dtype)
+        ).image_embeds
+        encoder_hidden_states = clip_image_embeds.unsqueeze(1)
+        uncond_encoder_hidden_states = torch.zeros_like(encoder_hidden_states)
+
+        if do_classifier_free_guidance:
+            encoder_hidden_states = torch.cat(
+                [uncond_encoder_hidden_states, encoder_hidden_states], dim=0
+            )
+        reference_control_writer = ReferenceAttentionControl(
+            self.reference_unet,
+            do_classifier_free_guidance=do_classifier_free_guidance,
+            mode="write",
+            batch_size=batch_size,
+            fusion_blocks="full",
+        )
+        reference_control_reader = ReferenceAttentionControl(
+            self.denoising_unet,
+            do_classifier_free_guidance=do_classifier_free_guidance,
+            mode="read",
+            batch_size=batch_size,
+            fusion_blocks="full",
+        )
+
+        num_channels_latents = self.denoising_unet.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            width,
+            height,
+            video_length,
+            clip_image_embeds.dtype,
+            device,
+            generator,
+        )
+
+        # Prepare extra step kwargs.
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # Prepare ref image latents
+        ref_image_tensor = self.ref_image_processor.preprocess(
+            ref_image, height=height, width=width
+        )  # (bs, c, width, height)
+        ref_image_tensor = ref_image_tensor.to(
+            dtype=self.vae.dtype, device=self.vae.device
+        )
+        ref_image_latents = self.vae.encode(ref_image_tensor).latent_dist.mean
+        ref_image_latents = ref_image_latents * 0.18215  # (b, 4, h, w)
+
+
+        # seed_features
+        if seed_images is not None:
+            seed_cond_tensor_list = []
+            for seed_image in seed_images:
+                seed_cond_tensor = self.ref_image_processor.preprocess(
+                    seed_image, height=height, width=width
+                )
+                seed_cond_tensor = seed_cond_tensor.unsqueeze(2)  # (bs, c, 1, h, w)
+                seed_cond_tensor_list.append(seed_cond_tensor)
+            seed_cond_tensor = torch.cat(seed_cond_tensor_list, dim=2)  # (bs, c, t, h, w)
+
+            seed_cond_tensor = rearrange(
+                    seed_cond_tensor, "b c f h w -> (b f) c h w"
+                )
+            seed_cond_tensor = seed_cond_tensor.to(
+                dtype=self.vae.dtype, device=self.vae.device
+            )
+            seed_image_latents = self.vae.encode(seed_cond_tensor).latent_dist.mean
+            seed_image_latents = seed_image_latents * 0.18215  # (b, 4, h, w)
+            seed_image_latents = rearrange(
+                    seed_image_latents, "(b f) c h w -> b c f h w",  f=video_length,
+                )
+            seed_image_latents = seed_image_latents.to(
+                device=device, dtype=self.denoising_unet.dtype
+            )
+            if seed_mask is not None:
+                seed_image_latents = torch.cat(
+                    [seed_image_latents, seed_mask.to(seed_image_latents)], dim=1
+                )
+            seed_fea = self.seed_guider(seed_image_latents)
+            seed_fea = (
+                torch.cat([seed_fea] * 2) if do_classifier_free_guidance else seed_fea
+            )
+        else:
+            seed_fea = None
+
+
+        if pose_images is not None:
+            pose_cond_tensor_list = []
+            for pose_image in pose_images:
+                pose_cond_tensor = self.cond_image_processor.preprocess(
+                    pose_image, height=height, width=width
+                )
+                pose_cond_tensor = pose_cond_tensor.unsqueeze(2)  # (bs, c, 1, h, w)
+                pose_cond_tensor_list.append(pose_cond_tensor)
+            pose_cond_tensor = torch.cat(pose_cond_tensor_list, dim=2)  # (bs, c, t, h, w)
+            pose_cond_tensor = pose_cond_tensor.to(
+                device=device, dtype=self.pose_guider.dtype
+            )
+            pose_fea = self.pose_guider(pose_cond_tensor)
+            pose_cond_tensor_zero = torch.zeros_like(pose_cond_tensor).to(
+                dtype=pose_fea.dtype, device=pose_fea.device
+            )
+            pose_fea_zero = self.pose_guider(pose_cond_tensor_zero)
+            pose_fea = (
+                torch.cat([pose_fea_zero, pose_fea]) if do_classifier_free_guidance else pose_fea
+            )
+            if seed_fea is None:
+                pose_fea = pose_fea
+            else:
+                pose_fea = pose_fea + seed_fea
+        else:
+            pose_fea = seed_fea
+
+        # denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                # 1. Forward reference image
+                if i == 0:
+                    self.reference_unet(
+                        ref_image_latents.repeat(
+                            (2 if do_classifier_free_guidance else 1), 1, 1, 1
+                        ),
+                        torch.zeros_like(t),
+                        # t,
+                        encoder_hidden_states=encoder_hidden_states,
+                        return_dict=False,
+                    )
+                    reference_control_reader.update(reference_control_writer)
+
+                # 3.1 expand the latents if we are doing classifier free guidance
+                latent_model_input = (
+                    torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                )
+                latent_model_input = self.scheduler.scale_model_input(
+                    latent_model_input, t
+                )
+
+                # print(pose_fea.shape, latent_model_input.shape)
+                noise_pred = self.denoising_unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=encoder_hidden_states,
+                    guidance_fea=pose_fea,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (
+                        noise_pred_text - noise_pred_uncond
+                    )
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(
+                    noise_pred, t, latents, **extra_step_kwargs, return_dict=False
+                )[0]
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or (
+                    (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0
+                ):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+            reference_control_reader.clear()
+            reference_control_writer.clear()
+
+        # Post-processing
+        images = self.decode_latents(latents)  # (b, c, f, h, w)
+
+        # Convert to tensor
+        if output_type == "tensor":
+            images = torch.from_numpy(images)
+
+        if not return_dict:
+            return images
+
+        return Pose2VideoPipelineOutput(videos=images)

2610_variant_runtime/diffusion/utils/__init__.py

@@ -0,0 +1 @@
+

2610_variant_runtime/diffusion/utils/util.py

@@ -0,0 +1,168 @@
+import importlib
+import os
+import os.path as osp
+import shutil
+import sys
+from pathlib import Path
+import time
+import av
+import numpy as np
+import torch
+import torchvision
+from einops import rearrange
+from PIL import Image
+import cv2
+
+def ensure_file_written(file_path, retries=10, delay=0.1):
+    for i in range(retries):
+        if os.path.exists(file_path):
+            image = Image.open(file_path)
+            if image is not None:
+                return True
+            else:
+                print(i)
+        time.sleep(delay)
+    return False
+
+def compute_snr(noise_scheduler, timesteps):
+    """
+    Computes SNR as per
+    https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849
+    """
+    alphas_cumprod = noise_scheduler.alphas_cumprod
+    sqrt_alphas_cumprod = alphas_cumprod**0.5
+    sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5
+
+    # Expand the tensors.
+    # Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026
+    sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[
+        timesteps
+    ].float()
+    while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape):
+        sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None]
+    alpha = sqrt_alphas_cumprod.expand(timesteps.shape)
+
+    sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(
+        device=timesteps.device
+    )[timesteps].float()
+    while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape):
+        sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None]
+    sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape)
+
+    # Compute SNR.
+    snr = (alpha / sigma) ** 2
+    return snr
+
+def seed_everything(seed):
+    import random
+
+    import numpy as np
+
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+    np.random.seed(seed % (2**32))
+    random.seed(seed)
+
+
+def import_filename(filename):
+    spec = importlib.util.spec_from_file_location("mymodule", filename)
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[spec.name] = module
+    spec.loader.exec_module(module)
+    return module
+
+
+def delete_additional_ckpt(base_path, num_keep):
+    dirs = []
+    for d in os.listdir(base_path):
+        if d.startswith("checkpoint-"):
+            dirs.append(d)
+    num_tot = len(dirs)
+    if num_tot <= num_keep:
+        return
+    # ensure ckpt is sorted and delete the ealier!
+    del_dirs = sorted(dirs, key=lambda x: int(x.split("-")[-1]))[: num_tot - num_keep]
+    for d in del_dirs:
+        path_to_dir = osp.join(base_path, d)
+        if osp.exists(path_to_dir):
+            shutil.rmtree(path_to_dir)
+
+
+def save_videos_from_pil(pil_images, path, fps=8):
+    import av
+
+    save_fmt = Path(path).suffix
+    os.makedirs(os.path.dirname(path), exist_ok=True)
+    width, height = pil_images[0].size
+
+    if save_fmt == ".mp4":
+        codec = "libx264"
+        container = av.open(path, "w")
+        stream = container.add_stream(codec, rate=fps)
+
+        stream.width = width
+        stream.height = height
+
+        for pil_image in pil_images:
+            # pil_image = Image.fromarray(image_arr).convert("RGB")
+            av_frame = av.VideoFrame.from_image(pil_image)
+            container.mux(stream.encode(av_frame))
+        container.mux(stream.encode())
+        container.close()
+
+    elif save_fmt == ".gif":
+        pil_images[0].save(
+            fp=path,
+            format="GIF",
+            append_images=pil_images[1:],
+            save_all=True,
+            duration=(1 / fps * 1000),
+            loop=0,
+        )
+    else:
+        raise ValueError("Unsupported file type. Use .mp4 or .gif.")
+
+
+def save_videos_grid(videos: torch.Tensor, path: str, rescale=False, n_rows=6, fps=8):
+    videos = rearrange(videos, "b c t h w -> t b c h w")
+    height, width = videos.shape[-2:]
+    outputs = []
+
+    for x in videos:
+        x = torchvision.utils.make_grid(x, nrow=n_rows)  # (c h w)
+        x = x.transpose(0, 1).transpose(1, 2).squeeze(-1)  # (h w c)
+        if rescale:
+            x = (x + 1.0) / 2.0  # -1,1 -> 0,1
+        x = (x * 255).numpy().astype(np.uint8)
+        x = Image.fromarray(x)
+
+        outputs.append(x)
+
+    os.makedirs(os.path.dirname(path), exist_ok=True)
+
+    save_videos_from_pil(outputs, path, fps)
+
+
+def read_frames(video_path):
+    container = av.open(video_path)
+
+    video_stream = next(s for s in container.streams if s.type == "video")
+    frames = []
+    for packet in container.demux(video_stream):
+        for frame in packet.decode():
+            image = Image.frombytes(
+                "RGB",
+                (frame.width, frame.height),
+                frame.to_rgb().to_ndarray(),
+            )
+            frames.append(image)
+
+    return frames
+
+
+def get_fps(video_path):
+    container = av.open(video_path)
+    video_stream = next(s for s in container.streams if s.type == "video")
+    fps = video_stream.average_rate
+    container.close()
+    return fps

2610_variant_runtime/diffusion/utils/video_utils.py

@@ -0,0 +1,89 @@
+import os
+import numpy as np
+import torch
+import torchvision
+import torch.nn.functional as F
+from PIL import Image
+from pathlib import Path
+import imageio
+from einops import rearrange
+import torchvision.transforms as transforms
+
+def concat_pil(img_pil_lst):
+    # horizontally concat PIL images
+    # NOTE(ZSH): assume all images are of same size
+    W, H = img_pil_lst[0].size
+    num_img = len(img_pil_lst)
+    new_width = num_img * W
+    new_image = Image.new("RGB", (new_width, H), color=0)
+    for img_idx, img in enumerate(img_pil_lst):
+        new_image.paste(img, (W * img_idx, 0))
+
+    return new_image
+
+def save_videos_from_pil(pil_images, path, fps=24, crf=23):
+
+    save_fmt = Path(path).suffix
+    os.makedirs(os.path.dirname(path), exist_ok=True)
+
+    if save_fmt == ".mp4":
+        with imageio.get_writer(path, fps=fps) as writer:
+            for img in pil_images:
+                img_array = np.array(img)  # Convert PIL Image to numpy array
+                writer.append_data(img_array)
+
+    elif save_fmt == ".gif":
+        pil_images[0].save(
+            fp=path,
+            format="GIF",
+            append_images=pil_images[1:],
+            save_all=True,
+            duration=(1 / fps * 1000),
+            loop=0,
+        )
+    else:
+        raise ValueError("Unsupported file type. Use .mp4 or .gif.")
+
+
+def save_videos_grid(videos: torch.Tensor, path: str, rescale=False, n_rows=6, fps=24):
+    videos = rearrange(videos, "b c t h w -> t b c h w")
+    height, width = videos.shape[-2:]
+    outputs = []
+
+    for x in videos:
+        x = torchvision.utils.make_grid(x, nrow=n_rows)  # (c h w)
+        x = x.transpose(0, 1).transpose(1, 2).squeeze(-1)  # (h w c)
+        if rescale:
+            x = (x + 1.0) / 2.0  # -1,1 -> 0,1
+        x = (x * 255).numpy().astype(np.uint8)
+        x = Image.fromarray(x)
+
+        outputs.append(x)
+
+    os.makedirs(os.path.dirname(path), exist_ok=True)
+
+    save_videos_from_pil(outputs, path, fps)
+
+
+def resize_tensor_frames(video_tensor, new_size):
+    B, C, video_length, H, W = video_tensor.shape
+    # Reshape video tensor to combine batch and frame dimensions: (B*F, C, H, W)
+    video_tensor_reshaped = video_tensor.reshape(-1, C, H, W)
+    # Resize using interpolate
+    resized_frames = F.interpolate(
+        video_tensor_reshaped, size=new_size, mode="bilinear", align_corners=False
+    )
+    resized_video = resized_frames.reshape(B, C, video_length, new_size[0], new_size[1])
+
+    return resized_video
+
+
+def pil_list_to_tensor(image_list, size=None):
+    to_tensor = transforms.ToTensor()
+    if size is not None:
+        tensor_list = [to_tensor(img.resize(size[::-1])) for img in image_list]
+    else:
+        tensor_list = [to_tensor(img) for img in image_list]
+    stacked_tensor = torch.stack(tensor_list, dim=0)
+    tensor = stacked_tensor.permute(1, 0, 2, 3)
+    return tensor

2610_variant_runtime/licenses/AnimateDiff.txt

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

2610_variant_runtime/licenses/DWPose.txt

@@ -0,0 +1,219 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2023 IDEA. All Rights Reserved.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+      http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+
+
+
+   MMPose (https://github.com/open-mmlab/mmpose)
+
+   Copyright 2018-2020 Open-MMLab.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+      http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

2610_variant_runtime/licenses/OpenRAIL-M.txt

@@ -0,0 +1,82 @@
+Copyright (c) 2022 Robin Rombach and Patrick Esser and contributors
+
+CreativeML Open RAIL-M
+dated August 22, 2022
+
+Section I: PREAMBLE
+
+Multimodal generative models are being widely adopted and used, and have the potential to transform the way artists, among other individuals, conceive and benefit from AI or ML technologies as a tool for content creation.
+
+Notwithstanding the current and potential benefits that these artifacts can bring to society at large, there are also concerns about potential misuses of them, either due to their technical limitations or ethical considerations.
+
+In short, this license strives for both the open and responsible downstream use of the accompanying model. When it comes to the open character, we took inspiration from open source permissive licenses regarding the grant of IP rights. Referring to the downstream responsible use, we added use-based restrictions not permitting the use of the Model in very specific scenarios, in order for the licensor to be able to enforce the license in case potential misuses of the Model may occur. At the same time, we strive to promote open and responsible research on generative models for art and content generation.
+
+Even though downstream derivative versions of the model could be released under different licensing terms, the latter will always have to include - at minimum - the same use-based restrictions as the ones in the original license (this license). We believe in the intersection between open and responsible AI development; thus, this License aims to strike a balance between both in order to enable responsible open-science in the field of AI.
+
+This License governs the use of the model (and its derivatives) and is informed by the model card associated with the model.
+
+NOW THEREFORE, You and Licensor agree as follows:
+
+1. Definitions
+
+- "License" means the terms and conditions for use, reproduction, and Distribution as defined in this document.
+- "Data" means a collection of information and/or content extracted from the dataset used with the Model, including to train, pretrain, or otherwise evaluate the Model. The Data is not licensed under this License.
+- "Output" means the results of operating a Model as embodied in informational content resulting therefrom.
+- "Model" means any accompanying machine-learning based assemblies (including checkpoints), consisting of learnt weights, parameters (including optimizer states), corresponding to the model architecture as embodied in the Complementary Material, that have been trained or tuned, in whole or in part on the Data, using the Complementary Material.
+- "Derivatives of the Model" means all modifications to the Model, works based on the Model, or any other model which is created or initialized by transfer of patterns of the weights, parameters, activations or output of the Model, to the other model, in order to cause the other model to perform similarly to the Model, including - but not limited to - distillation methods entailing the use of intermediate data representations or methods based on the generation of synthetic data by the Model for training the other model.
+- "Complementary Material" means the accompanying source code and scripts used to define, run, load, benchmark or evaluate the Model, and used to prepare data for training or evaluation, if any. This includes any accompanying documentation, tutorials, examples, etc, if any.
+- "Distribution" means any transmission, reproduction, publication or other sharing of the Model or Derivatives of the Model to a third party, including providing the Model as a hosted service made available by electronic or other remote means - e.g. API-based or web access.
+- "Licensor" means the copyright owner or entity authorized by the copyright owner that is granting the License, including the persons or entities that may have rights in the Model and/or distributing the Model.
+- "You" (or "Your") means an individual or Legal Entity exercising permissions granted by this License and/or making use of the Model for whichever purpose and in any field of use, including usage of the Model in an end-use application - e.g. chatbot, translator, image generator.
+- "Third Parties" means individuals or legal entities that are not under common control with Licensor or You.
+- "Contribution" means any work of authorship, including the original version of the Model and any modifications or additions to that Model or Derivatives of the Model thereof, that is intentionally submitted to Licensor for inclusion in the Model by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, "submitted" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Model, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as "Not a Contribution."
+- "Contributor" means Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Model.
+
+Section II: INTELLECTUAL PROPERTY RIGHTS
+
+Both copyright and patent grants apply to the Model, Derivatives of the Model and Complementary Material. The Model and Derivatives of the Model are subject to additional terms as described in Section III.
+
+2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare, publicly display, publicly perform, sublicense, and distribute the Complementary Material, the Model, and Derivatives of the Model.
+3. Grant of Patent License. Subject to the terms and conditions of this License and where and as applicable, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this paragraph) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Model and the Complementary Material, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Model to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Model and/or Complementary Material or a Contribution incorporated within the Model and/or Complementary Material constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for the Model and/or Work shall terminate as of the date such litigation is asserted or filed.
+
+Section III: CONDITIONS OF USAGE, DISTRIBUTION AND REDISTRIBUTION
+
+4. Distribution and Redistribution. You may host for Third Party remote access purposes (e.g. software-as-a-service), reproduce and distribute copies of the Model or Derivatives of the Model thereof in any medium, with or without modifications, provided that You meet the following conditions:
+Use-based restrictions as referenced in paragraph 5 MUST be included as an enforceable provision by You in any type of legal agreement (e.g. a license) governing the use and/or distribution of the Model or Derivatives of the Model, and You shall give notice to subsequent users You Distribute to, that the Model or Derivatives of the Model are subject to paragraph 5. This provision does not apply to the use of Complementary Material.
+You must give any Third Party recipients of the Model or Derivatives of the Model a copy of this License;
+You must cause any modified files to carry prominent notices stating that You changed the files;
+You must retain all copyright, patent, trademark, and attribution notices excluding those notices that do not pertain to any part of the Model, Derivatives of the Model.
+You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions - respecting paragraph 4.a. - for use, reproduction, or Distribution of Your modifications, or for any such Derivatives of the Model as a whole, provided Your use, reproduction, and Distribution of the Model otherwise complies with the conditions stated in this License.
+5. Use-based restrictions. The restrictions set forth in Attachment A are considered Use-based restrictions. Therefore You cannot use the Model and the Derivatives of the Model for the specified restricted uses. You may use the Model subject to this License, including only for lawful purposes and in accordance with the License. Use may include creating any content with, finetuning, updating, running, training, evaluating and/or reparametrizing the Model. You shall require all of Your users who use the Model or a Derivative of the Model to comply with the terms of this paragraph (paragraph 5).
+6. The Output You Generate. Except as set forth herein, Licensor claims no rights in the Output You generate using the Model. You are accountable for the Output you generate and its subsequent uses. No use of the output can contravene any provision as stated in the License.
+
+Section IV: OTHER PROVISIONS
+
+7. Updates and Runtime Restrictions. To the maximum extent permitted by law, Licensor reserves the right to restrict (remotely or otherwise) usage of the Model in violation of this License, update the Model through electronic means, or modify the Output of the Model based on updates. You shall undertake reasonable efforts to use the latest version of the Model.
+8. Trademarks and related. Nothing in this License permits You to make use of Licensors’ trademarks, trade names, logos or to otherwise suggest endorsement or misrepresent the relationship between the parties; and any rights not expressly granted herein are reserved by the Licensors.
+9. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Model and the Complementary Material (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Model, Derivatives of the Model, and the Complementary Material and assume any risks associated with Your exercise of permissions under this License.
+10. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Model and the Complementary Material (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages.
+11. Accepting Warranty or Additional Liability. While redistributing the Model, Derivatives of the Model and the Complementary Material thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability.
+12. If any provision of this License is held to be invalid, illegal or unenforceable, the remaining provisions shall be unaffected thereby and remain valid as if such provision had not been set forth herein.
+
+END OF TERMS AND CONDITIONS
+
+
+
+
+Attachment A
+
+Use Restrictions
+
+You agree not to use the Model or Derivatives of the Model:
+- In any way that violates any applicable national, federal, state, local or international law or regulation;
+- For the purpose of exploiting, harming or attempting to exploit or harm minors in any way;
+- To generate or disseminate verifiably false information and/or content with the purpose of harming others;
+- To generate or disseminate personal identifiable information that can be used to harm an individual;
+- To defame, disparage or otherwise harass others;
+- For fully automated decision making that adversely impacts an individual’s legal rights or otherwise creates or modifies a binding, enforceable obligation;
+- For any use intended to or which has the effect of discriminating against or harming individuals or groups based on online or offline social behavior or known or predicted personal or personality characteristics;
+- To exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm;
+- For any use intended to or which has the effect of discriminating against individuals or groups based on legally protected characteristics or categories;
+- To provide medical advice and medical results interpretation;
+- To generate or disseminate information for the purpose to be used for administration of justice, law enforcement, immigration or asylum processes, such as predicting an individual will commit fraud/crime commitment (e.g. by text profiling, drawing causal relationships between assertions made in documents, indiscriminate and arbitrarily-targeted use).

2610_variant_runtime/licenses/PixArt-alpha.txt

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2024 Junsong Chen, Jincheng Yu, Enze Xie
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

2610_variant_runtime/licenses/diffusers.txt

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, Any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

2610_variant_runtime/licenses/magic-animate.txt

@@ -0,0 +1,28 @@
+BSD 3-Clause License
+
+Copyright 2023 MagicAnimate Team All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+
+3. Neither the name of the copyright holder nor the names of its
+   contributors may be used to endorse or promote products derived from
+   this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

2610_variant_runtime/requirements.lock.txt

@@ -0,0 +1,133 @@
+absl-py==2.5.1
+accelerate==0.21.0
+antlr4-python3-runtime==4.9.3
+asttokens==3.0.2
+audioread==3.1.0
+av==12.2.0
+bitsandbytes==0.43.3
+certifi==2026.7.22
+cffi==2.1.1
+charset-normalizer==3.5.2
+click==8.5.0
+cloudpickle==3.1.2
+coloredlogs==15.0.1
+configargparse==1.8.0
+contourpy==1.3.2
+controlnet-aux==0.0.7
+cycler==0.12.1
+decorator==4.4.2
+decord==0.6.0
+diffusers==0.24.0
+docker-pycreds==0.4.0
+einops==0.8.2
+exceptiongroup==1.3.1
+executing==2.3.0
+filelock==4.1.0
+flatbuffers==25.12.19
+fonttools==4.65.0
+fsspec==2026.9.0
+gitdb==4.0.12
+gitpython==3.2.0
+grpcio==1.84.0
+h5py==3.16.0
+huggingface-hub==0.25.2
+humanfriendly==10.0
+idna==3.20
+igraph==0.11.6
+imageio-ffmpeg==0.4.9
+imageio==2.33.0
+importlib-metadata==9.0.1
+ipython==8.39.0
+jedi==0.20.1
+jinja2==3.1.6
+joblib==1.6.0
+kiwisolver==1.5.1
+lazy-loader==0.6
+librosa==0.10.2.post1
+llvmlite==0.50.0
+lmdb==3.0.0
+loguru==0.7.3
+markdown==3.10.3
+markupsafe==3.0.4
+matplotlib-inline==0.2.2
+matplotlib==3.10.9
+moviepy==1.0.3
+mpmath==1.3.0
+msgpack==1.2.3
+networkx==3.4.2
+numba==0.68.0
+numpy==1.26.4
+nvidia-cublas-cu12==12.1.3.1
+nvidia-cuda-cupti-cu12==12.1.105
+nvidia-cuda-nvrtc-cu12==12.1.105
+nvidia-cuda-runtime-cu12==12.1.105
+nvidia-cudnn-cu12==8.9.2.26
+nvidia-cufft-cu12==11.0.2.54
+nvidia-curand-cu12==10.3.2.106
+nvidia-cusolver-cu12==11.4.5.107
+nvidia-cusparse-cu12==12.1.0.106
+nvidia-nccl-cu12==2.19.3
+nvidia-nvjitlink-cu12==12.9.86
+nvidia-nvtx-cu12==12.1.105
+omegaconf==2.3.0
+onnxruntime-gpu @ https://aiinfra.pkgs.visualstudio.com/2692857e-05ef-43b4-ba9c-ccf1c22c437c/_packaging/9387c3aa-d9ad-4513-968c-383f6f7f53b8/pypi/download/onnxruntime-gpu/1.18/onnxruntime_gpu-1.18.0-cp310-cp310-manylinux_2_28_x86_64.whl#sha256=039be9a6b7f71c6739e97eec79f4bf240793a7c0c4108a09e0e1a27b4c33dbca
+opencv-python-headless==4.9.0.80
+packaging==26.3
+pandas==2.3.3
+parso==0.8.7
+pexpect==4.9.0
+pillow==12.3.0
+platformdirs==4.12.4
+pooch==1.9.0
+proglog==0.1.12
+prompt-toolkit==3.0.53
+protobuf==5.29.6
+psutil==7.2.2
+ptyprocess==0.7.0
+pure-eval==0.2.4
+pyarrow==25.0.1
+pycparser==3.11
+pygments==2.21.0
+pyparsing==3.3.3
+python-dateutil==2.9.0.post0
+pytz==2026.5
+pyyaml==6.0.3
+regex==2026.9.29
+requests==2.34.2
+safetensors==0.8.0
+scikit-image==0.25.2
+scikit-learn==1.7.2
+scipy==1.15.3
+sentry-sdk==2.72.0
+setproctitle==1.3.8
+setuptools==80.10.2
+six==1.17.0
+smmap==5.0.3
+smplx==0.1.28
+soundfile==0.14.0
+soxr==1.1.0
+stack-data==0.6.3
+sympy==1.14.0
+tensorboard-data-server==0.7.2
+tensorboard==2.18.0
+termcolor==3.3.0
+textgrid==1.6.1
+texttable==1.7.1
+threadpoolctl==3.7.0
+tifffile==2025.5.10
+timm==1.0.30
+tokenizers==0.15.2
+torch==2.2.2
+torchvision==0.17.2
+tqdm==4.70.1
+traitlets==5.16.1
+transformers==4.36.2
+triton==2.2.0
+typing-extensions==4.16.0
+tzdata==2026.5
+urllib3==2.8.0
+wandb==0.17.9
+wcwidth==0.9.2
+werkzeug==3.1.9
+xformers==0.0.25.post1
+zipp==4.1.1

2610_variant_runtime/scripts/download_assets.py

@@ -0,0 +1,50 @@
+import argparse
+import hashlib
+import json
+import shutil
+import urllib.request
+from pathlib import Path
+
+
+def verify(path, asset):
+    if not path.is_file() or path.stat().st_size != asset["bytes"]:
+        return False
+    digest = hashlib.sha256()
+    with path.open("rb") as stream:
+        for chunk in iter(lambda: stream.read(8 * 1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest() == asset["sha256"]
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Download and verify VMG models and raw examples.")
+    parser.add_argument("--group", choices=["all", "base", "preprocessing", "checkpoint", "examples"], default="all")
+    parser.add_argument("--verify-only", action="store_true")
+    args = parser.parse_args()
+    root = Path(__file__).resolve().parents[1]
+    assets = json.loads((root / "asset_manifest.json").read_text())["assets"]
+    missing = []
+    for asset in assets:
+        if args.group != "all" and asset["group"] != args.group:
+            continue
+        path = root / asset["path"]
+        if verify(path, asset):
+            print(f"Verified {asset['path']}", flush=True)
+            continue
+        if args.verify_only:
+            missing.append(asset["path"])
+            continue
+        path.parent.mkdir(parents=True, exist_ok=True)
+        temporary = path.with_name(path.name + ".part")
+        print(f"Downloading {asset['path']} ({asset['bytes']:,} bytes)", flush=True)
+        with urllib.request.urlopen(asset["url"], timeout=120) as response, temporary.open("wb") as output:
+            shutil.copyfileobj(response, output, length=8 * 1024 * 1024)
+        if not verify(temporary, asset):
+            raise RuntimeError(f"Checksum mismatch: {asset['path']}")
+        temporary.replace(path)
+    if missing:
+        raise SystemExit("Missing or invalid assets: " + ", ".join(missing))
+
+
+if __name__ == "__main__":
+    main()

README.md

@@ -27,3 +27,16 @@ Checkpoint SHA256: `9cc0f0354c5ec4b33085196ba5786f845e0901fd1dc7159e57be14d972c3
 Starting with only the raw 3,051-frame example video and music, the release code rebuilt pose and segmentation, constructed a 762-node / 1,059-edge graph, and generated three new diffusion transitions. The resulting video is 18 seconds, 768×768, 30 fps with audio. This run completed in 12 minutes 2 seconds on one H100 80GB.
 
 <video controls src="https://huggingface.co/H-Liu1997/Video-Motion-Graphs/resolve/main/examples/generated_demo.mp4"></video>
+## Historical interpolation variants
+
+The following dated archival variants are separate from the original weighted-pose checkpoint above. Their downloaded source files passed full SHA256 checks and strict CPU key/shape checks. **GPU forward execution, rendered-video evaluation and fresh-Hub-download verification are pending.** The CPU checks did not perform a forward pass.
+
+| Material Date | Historical Run | Step | Details |
+| --- | --- | ---: | --- |
+| 2024-09 | `v7_dup_14_continue_2` | 655000 | [2409 historical variant](2409_variants/v7_dup_14_continue_2_step655000/README.md) |
+| 2024-11 | `v7_dup_14_s2_8frames_512` | 9000 | [2411 8-frame variant](2411_variants/v7_dup_14_s2_8frames_512_step9000/README.md) |
+| 2024-11 | `v7_dup_14_s2_12frames_512` | 30000 | [2411 12-frame variant](2411_variants/v7_dup_14_s2_12frames_512_step30000/README.md) |
+
+The [standalone runtime](2610_variant_runtime/README.md) contains the matching five-channel standard-VAE entrypoint, three configurations, pinned requirements, source licenses and a base-asset downloader. Its source snapshot is recorded in the [runtime hash manifest](2610_runtime_asset_manifest.json); model and configuration hashes are in the [variant asset manifest](2610_variant_asset_manifest.json).
+
+Date prefixes describe the earliest confirmed source-material month. The relationship of step 655000 to the final TANGO or ACInterp model is unconfirmed; none of these checkpoints is labeled best or paper-final. Only model weights are retained, without optimizer or exact-resume state. The existing `finetune_weighted/checkpoint-8000` model and its previously verified example remain unchanged.