ComfyUI-PiD

⚠️ Archived — use ComfyUI core instead.

PiD (and PixelDiT base T2I) now have official ComfyUI core support via kijai's PR. It ships native nodes, auto model detection, proper VRAM management, and uses Comfy-Org's repackaged weights. That's the version you should use.

This repo predates the upstream support and is kept around as a learning artifact / reference implementation. Don't install it for production use.

---

NVIDIA PiD (Pixel Diffusion Decoder) as native ComfyUI nodes. Decodes latents straight to 4× resolution in a single 4-step pass. Works with Flux, Flux2, and Z-Image. No new pip dependencies.

Nodes

• PiD Decoder Loader — loads a PiD checkpoint + Gemma text encoder.
• PiD Decode — LATENT → IMAGE. Drop-in for VAE Decode, but 4× larger.
• PiD Super-Resolve — IMAGE → IMAGE. 4× upscale an existing image.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/nynxz/ComfyUI-PiD.git

Grab two files:

Model	Source	Goes in
Gemma-2-2b-it text encoder (~5GB)	https://huggingface.co/nynxz/PiD	ComfyUI/models/text_encoders/
PiD decoder checkpoint (~2.7GB each)	https://huggingface.co/nvidia/PiD	ComfyUI/models/pid/

Pick the decoder for your generation resolution:

KSampler at	Output	Checkpoint
512px (Flux / Z-Image)	2048px	PiD_res2k_sr4x_official_flux_distill_4step
1024px (Flux / Z-Image)	4K	PiD_res2kto4k_sr4x_official_flux_distill_4step
Flux2	2K / 4K	swap _flux_ → _flux2_

Filename matters — the loader auto-picks the backbone from _flux_ / _flux2_ substrings. Z-Image uses the Flux checkpoint (shared VAE).

Use

Replace VAE Decode with PiD Decode after your KSampler. Wire PiD Decoder Loader into the decoder input. Feed PiD Decode the same prompt you fed the KSampler — it uses it for guidance.

Example workflow: examples/PiD_flux2.json. Same prompt + seed, only the decoder changes:

Flux2 VAE Decode (512×512)	PiD Decode (2048×2048)

Click through to see the fine texture (hair, fabric weave, skin pores) PiD reconstructs.

License

Apache 2.0 — see LICENSE and NOTICE. Portions derived from nvidia/PiD. Bundled tokenizer is Google's official Gemma-2-2b SentencePiece model (sha256 61a7b147… — same as in google/gemma-2-2b-it).