Cierpliwy/krea2-inpaint-edit

Krea 2 inpaint edit

Prompt

masked image / this LoRA / yijunwang2-krea2-anypaint

LoRAs for Krea-2-Raw that perform inpainting by masked editing: you provide an image with the region to (re)generate filled in black, plus a text prompt describing what should go there — the model paints the masked region while leaving the rest of the image untouched.

The gallery below compares the results of this LoRA against yijunwang2/krea2-anypaint (column 3), with per-row difference maps.

Files

Recommended: krea2_inpaint_edit_v2_17500.safetensors (repo root) — the best checkpoint of the v2 512px training run. It covers everything the original LoRAs did (masked inpainting, editing) plus outpainting (extending a scene into black padding), and generalizes to other resolutions.

All training runs are archived in subfolders, best checkpoints only:

Path	File	Notes
./	krea2_inpaint_edit_v2_17500.safetensors	recommended — v2 512px run, step 17500 (rank 64)
v2_512/	inpaint_p1_512_v2_000017500.safetensors	same file as the recommended one (original checkpoint name)
v2_512/	inpaint_p1_512_v2_000020000.safetensors	end-of-run v2 checkpoint (step 20000), grades nearly identically
v1_1024/	inpaint_p1_1024_000017500.safetensors	best 1024px-run checkpoint (step 17500)
v1_1024/	inpaint_p1_1024_000019000.safetensors	second-best 1024px-run checkpoint (step 19000)
v1_512/	krea2_inpaint_edit.safetensors	v1 512px — default strength
v1_512/	krea2_inpaint_edit_mild.safetensors	v1 512px — less changes to non-masked areas; better at inpainting areas covering the main focus point
v1_512/	krea2_inpaint_edit_strong.safetensors	v1 512px — better at outpainting; adapts more strongly to the prompt

Notes on the runs:

• v1 512 (v1_512/): the original series — random black masks, no outpainting examples.
• v1 1024 (v1_1024/): same recipe at 1024px; the two checkpoints above were the top of that run.
• v2 512 (v2_512/): 512px retrain (~10k–20k steps, rank 64) where ~10% of the masks are outpainting masks (a rectangular island of the image visible, the rest black). This is what gives it real outpainting capability. Side by side, the v2 512 checkpoints graded clearly ahead of both the v1 series and the v1 1024 checkpoints on inpainting, outpainting, rotation and custom edits.

Requirements

• ComfyUI
• ComfyUI-Krea2-Ostris-Edit — Krea 2 edit nodes (TextEncodeKrea2OstrisEdit, Krea2OstrisEditModelPatch). kv_cache on Krea2OstrisEditModelPatch must be enabled (it is disabled by default) — this is important for correct masked editing.
  • Inputs above 1MP: until ostris/ComfyUI-Krea2-Ostris-Edit#10 is merged, install the vlm-ref-size-inputs branch of the fork instead, which allows reference images higher than 1MP.
• Krea-2-Raw base model weights: krea/Krea-2-Raw

The example workflows only use official ComfyUI nodes (including comfy_extras) plus the Ostris edit nodes above.

Quick start

1. Download the LoRA you want (Files & versions tab) and put it in ComfyUI/models/loras/.
2. Mask your image: fill the region you want to (re)generate with solid black — the mask can cover part or all of the image.
3. Load any example workflow from the workflows/ folder into ComfyUI (Workflow → Open).
4. Connect your masked image to the image1 input of the inpaint TextEncodeKrea2OstrisEdit node (the one after the LoRA loader) — replacing the auto-generated reference the workflow feeds it by default.
5. In that TextEncodeKrea2OstrisEdit node, write a prompt describing what should appear in the masked region (style, colors, content).
6. Make sure kv_cache is enabled on the Krea2OstrisEditModelPatch node (it is disabled by default).
7. Queue the prompt.

Example workflows

All example workflows live in the workflows/ folder. Each one first generates a base image with stock Krea-2, builds the black-masked reference from it, then runs the inpaint-edit LoRA to fill the black region.

Workflow	What it does	Extra models needed
workflows/outpaint_example.json	Generates an image, pads it onto a larger black canvas (Image Composite Masked) and extends the scene into the black border	—
workflows/background_replacement_example.json	Generates an image, removes the background (BiRefNet via RemoveBackground) and composites the foreground onto black, then inpaints a new background	birefnet.safetensors → models/bg_removal/
workflows/moge_rotate_inpaint_example.json	Generates an image, estimates geometry with MoGe-3, converts it to a textured mesh, rotates the mesh and renders it on a black background, then inpaints the scene around the new viewpoint	moge_3_vitl_fp16.safetensors → models/geometry_estimation/
workflows/sam3_detect_example.json	Generates an image, detects the subject with SAM 3 open-vocabulary text detection, composites the detected object onto black, then inpaints a new background around it	sam3.safetensors → models/checkpoints/ (loaded via CheckpointLoaderSimple, which also provides the SAM 3 text encoder)

If you change the base generation size or the canvas size in any of these workflows, keep the composite offsets / canvas dimensions in sync.