nv-tlabs/PiD

Health v2 · maintenance only

Flags: no_releases young no_license

Adoption not part of the score

1045 stars · 61 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

PiD is a plug-and-play pixel diffusion decoder from NVIDIA that replaces VAE/RAE decoders, decoding latent representations directly into high-resolution pixels in a single pass. It unifies decoding and upsampling (4x-8x) into one generative module and supports latents from FLUX, SD3, SDXL, Qwen-Image, DINOv2, and SigLIP.

Use cases

• decode latent diffusion model outputs to high-resolution images
• upscale 512x512 latents to 2048x2048 pixels quickly
• replace VAE decoder in text-to-image pipelines
• decode DINOv2 or SigLIP semantic latents into images
• integrate high-resolution decoding into ComfyUI workflows
• early-exit latent diffusion by decoding partially denoised latents

When to choose

• you need fast, high-resolution image decoding from latent diffusion models
• you want to replace the decode-then-super-resolve cascade with lower latency
• you work with FLUX, SD3, SDXL, or Qwen-Image latents and need 4K output
• you need a distilled few-step decoder for consumer GPUs

When to avoid

• you need a lightweight deterministic decoder without diffusion sampling
• your environment lacks a CUDA-capable GPU with sufficient VRAM
• you need non-image modalities or non-PyTorch frameworks

Facets

library · maturity active

image-processing machine-learning deep-learning image-processing machine-learning artificial-intelligence python windows diffusion-decoder pixel-diffusion super-resolution latent-decoding text-to-image nvidia comfyui gpu linux macos

Member repositories

For agents

markdown · JSON · MCP: product_card(name="nv-tlabs/PiD")