What is comfyui-pid?

ComfyUI-PiD is a plugin for ComfyUI, a node-based image generation interface, that adds support for NVIDIA's PiD (pixel diffusion) decoding method. PiD is an alternative way to convert the internal mathematical representation that diffusion models work with (called a latent) into a final visible image. Unlike the standard decoder bundled with most image generation models, PiD takes additional inputs: a text caption describing the image content and a numerical value called sigma that controls the decoding behavior. The plugin works with several image generation model families including Flux, Flux2, Stable Diffusion 3, and others. Depending on the backbone, the plugin selects an appropriate PiD checkpoint trained for that architecture. The scale option controls how much the output is enlarged compared to the base latent size, with common settings producing images at 2048x2048 or 4096x4096 pixels. For systems with limited graphics memory, the plugin offers a staged workflow that splits the decoding process across several nodes. The most memory-intensive computation step runs in a separate Python subprocess so its GPU memory is released when it finishes, rather than held alongside everything else already loaded in ComfyUI. Additional options allow individual model blocks to be processed sequentially rather than all at once, trading speed for lower peak memory usage. Installation involves cloning the repository into ComfyUI's custom_nodes folder and installing the Python dependencies listed in requirements.txt. The plugin can download the necessary PiD source code and model weights automatically on first use if the auto_download option is enabled. PyTorch is not included in the requirements because ComfyUI already provides it. The plugin is a community-built wrapper around NVIDIA's publicly released PiD code and is not an official NVIDIA or ComfyUI project. NVIDIA's model weights may have their own usage terms separate from this plugin's MIT license.