BennyDaBall/PiD-1.5-flux2-nvfp4-comfy

🚀 PiD 1.5 FLUX.2 — NVFP4 (ComfyUI-ready)

The FLUX.2 half of the first quantized PiD decoder pair on HF. 2.61 GB → 1.09 GB, drops straight into a stock UNETLoader, decodes faster than the bf16 original. No custom loader. No core patches. No Blackwell requirement either — this runs anywhere the bf16 file runs.

🧠 What is this?

NVIDIA's PiD v1.5 decoder for the FLUX.2 latent family (FLUX.2 dev + Klein 4B/9B, 128-channel latents) — the 4-step pixel-diffusion decode that replaces your VAE decode and hands back 4× the resolution. Same decoder, 42% of the size, a little faster.

Running qwen-image models instead? That build is already out: PiD-1.5-qwenimage-nvfp4-comfy.

👀 Same seed, same latent, both builds

100% crops — faces, lettering, bokeh all hold:

📋 Specs

Attribute	Details
Base	nvidia/PiD v1.5 flux2 (4-step distill), via the Comfy-Org/PixelDiT repackage
Quantization	NVFP4 (group 16), quality-critical layers kept bf16
File size	1.09 GB (bf16: 2.61 GB)
Requirements	ComfyUI ≥ 0.32 (tested on 0.33.0) · any GPU that runs PiD bf16
Text encoder	gemma_2_2b_it_elm_bf16.safetensors from Comfy-Org/PixelDiT (CLIPLoader type pixeldit)

💻 Usage

Drop the model in ComfyUI/models/diffusion_models/, select it in UNETLoader, done. Four workflows included — pick your Klein and your size:

Simple (core nodes only) — PiD-1.5-flux2-nvfp4_simple_klein4b_1mp_to_16mp.json and ..._simple_klein9b_1mp_to_16mp.json. Klein stage-1 at ~1MP → single-shot PiD 4× → 5376×3072, built from nothing but core nodes (PiDConditioning + a 4-step KSampler). Drop the included pixel_space_vae.safetensors into ComfyUI/models/vae/ for the final decode step. Keep these near the trained envelope — single-shot PiD collapses past ~4K output.

Tiled (the big guns) — PiD-1.5-flux2-nvfp4_tiled_klein4b.json and ..._tiled_klein9b.json. Stage-1 → seam-free 7680×4352 in one queue via ComfyUI-Latent-Tiled-PiD: install from ComfyUI Manager (search "Latent-Tiled-PiD") or the Comfy Registry. Klein 9B stays coherent up to the ~59 MP rungs; let 4B own the monsters.

Every included workflow was executed through the actual ComfyUI frontend against this exact file before upload. FLUX.2-family latents only (128-ch; the nodes auto-detect them under flux/flux2).

⚖️ License & credits

NVIDIA created PiD (paper); Comfy-Org published the bf16 ComfyUI repackage this converts. NSCLv1 — non-commercial research/evaluation use only, derivatives included. Read the nvidia/PiD card before you build anything on it. Not affiliated with NVIDIA or Comfy-Org. Conversion and validation by BennyDaBall_OG.