BennyDaBall/PiD-1.5-flux1-nvfp4-comfy

🚀 PiD 1.5 FLUX.1 / Z-Image — NVFP4 (ComfyUI-ready)

The FLUX.1 leg of the quantized PiD decoder trio on HF. 2.61 GB → 1.84 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.1 latent family (FLUX.1 dev/schnell + Z-Image-Turbo, 16-channel flux latents) — the 4-step pixel-diffusion decode that replaces your VAE decode and hands back 4× the resolution. Same decoder, 70% of the size, a little faster.

This family holds more of its quality in a handful of layers than its siblings do, so the converter keeps more of them in bf16 — that's why this file is bigger than the qwen/flux2 builds. What you get for it: pixel-identical character to bf16 at 100% zoom.

Running other stage-1 families? The rest of the trio: PiD-1.5-qwenimage-nvfp4-comfy · PiD-1.5-flux2-nvfp4-comfy

👀 Same seed, same latent, both builds

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

📋 Specs

Attribute	Details
Base	nvidia/PiD v1.5 flux1 (4-step distill), via the Comfy-Org/PixelDiT repackage
Quantization	NVFP4 (group 16), quality-critical layers kept bf16
File size	1.84 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. Both included workflows run Z-Image-Turbo as the stage-1 (qwen_3_4b.safetensors text encoder on CLIPLoader type lumina2, ModelSamplingAuraFlow shift 3.0) — swap in any FLUX.1-family stage-1 and the decode side doesn't change.

Simple (core nodes only) — PiD-1.5-flux1-nvfp4_simple_zimage_1mp_to_16mp.json. Z-Image-Turbo at 1024×1024 → single-shot PiD 4× → 4096×4096, built from nothing but core nodes (PiDConditioning + a 4-step sampler on the exact distill sigmas). Drop the included pixel_space_vae.safetensors into ComfyUI/models/vae/ for the final decode step. ⚠ flux1's single-shot envelope is tighter than its siblings: keep the output at 4096 on the long edge and feed it a ~1MP stage-1 — square is the sweet spot. For 16:9 or anything bigger, use the tiled workflow.

Tiled (the big guns) — PiD-1.5-flux1-nvfp4_tiled_zimage.json. Z-Image stage-1 at 1920×1088 → seam-free 7680×4352 in one queue via ComfyUI-Latent-Tiled-PiD: install from ComfyUI Manager (search "Latent-Tiled-PiD") or the Comfy Registry. Set latent_format to flux and every 16-channel flux latent decodes clean to the big rungs.

Every included workflow was executed through the actual ComfyUI frontend against this exact file before upload.

⚖️ 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.