tsolful/comfyui-pid

PiD stands for Pixel Diffusion Decoder, a research project from NVIDIA. It is a decoder that sits at the end of AI image-generation pipelines and replaces the standard component that converts a compressed internal representation into actual pixels. Most image generators work in two steps: they create a compressed sketch (called a latent), then decode it into a full image. PiD takes over that decoding step and simultaneously upscales the image to much higher resolution, doing both in a single pass instead of separately. This repository is a ComfyUI custom node that lets you plug PiD into the ComfyUI workflow. ComfyUI is a popular visual tool for building image-generation pipelines. Adding this node gives you PiD's decoding and upscaling capability directly inside that interface without needing to write any code. PiD supports several underlying image-generation backbones, including Flux, Flux2, Stable Diffusion 3, Z-Image, DINOv2, and SigLIP. For each backbone there are two usage paths: starting from an existing image (encode it, then let PiD decode it at higher resolution), or generating from a text prompt (run the text-to-image model partway through its process, then hand off to PiD for the final decode and upscale). Pre-trained model weights are hosted on Hugging Face. The repository also includes a standalone Python package with command-line inference scripts for users who do not use ComfyUI. Running either the scripts or the node requires PyTorch with CUDA support and a handful of other machine-learning libraries. The pre-trained checkpoints come in two variants: one trained at 2048 pixels and one capable of decoding up to 4K resolution. Both support non-square aspect ratios. This is a research code release tied to an academic paper. It targets researchers and advanced developers working on image generation, not general users looking for a simple photo upscaling tool.