viralvfx/comfyui-int4-fast

ComfyUI-INT4-Fast is an add-on for ComfyUI, a visual tool many people use to run AI image generation models. This add-on lets those image generation models run in a compressed format called INT4, which shrinks each number the model uses down to just 4 bits instead of the usual larger sizes. The payoff is that models load faster, use less memory, and run more quickly on graphics cards that support a feature called Tensor Cores, without needing a separate, more powerful setup. It builds directly on an earlier project by another developer, called ComfyUI-INT8-Fast, adapting that same approach from 8-bit compression down to the more aggressive 4-bit version. The add-on can take a standard model file and shrink it to this format automatically the first time you load it, or you can save the shrunk version to disk so future loads are instant. It also handles the more delicate parts of a model, like its first and last layers, by keeping those in a slightly less aggressive 8-bit format so quality does not suffer where it matters most. If you use LoRA files, small add-on tweaks that customize how a model draws, this tool automatically adjusts them so they still work correctly with the compressed model. The package adds two new building blocks inside ComfyUI: one to load a model in this compressed format, and one to save a model out in that same format after processing. The developer notes that your very first image generation after loading a model will take longer, since the system needs to set up some behind the scenes optimizations, but every generation after that should be noticeably faster. To use it, you need ComfyUI already installed and updated, plus a required companion package called comfy-kitchen. This tool is meant for people who are already comfortable running ComfyUI and want faster, lighter image generation on supported hardware.