document97/comfy-int4-hiptest

This project is a plugin for ComfyUI, a visual tool used to build and run AI image and video generation workflows, aimed specifically at people running AMD graphics cards rather than the more common Nvidia setup. It adds support for loading AI diffusion models that have been shrunk down using a technique called INT4 quantization, which compresses a model's numbers into a smaller format so it runs faster and uses less memory, at some cost to precision. The plugin is built and tested around one specific AMD graphics chip, the Radeon 780M, running on Windows with a particular combination of AMD's HIP software stack, a compatibility layer called ZLUDA that lets AMD cards run CUDA style code, and a custom build of a compiler called Triton that understands this compressed number format natively. It does not include the actual compressed model files or the special Triton build itself. Those must be obtained separately from a companion repository by the same author, along with a version of ComfyUI that already supports this quantization method. Once installed, the plugin adds two new nodes inside ComfyUI: one to load a compressed diffusion model and one to save a model in this compressed format. It also supports combining the compressed model with LoRA files, which are small add on files that adjust how a model behaves, offering three different ways of blending them in depending on whether someone wants the fastest performance or the most accurate preservation of small adjustments. The README documents that it has been tested successfully with two specific published models, one for still images and one for video generation, and it includes diagnostic messages a user can check in the ComfyUI console to confirm the plugin is actually using the fast AMD specific path rather than falling back to a slower general purpose one. The project builds on two earlier open source projects for INT8 and INT4 quantization support in ComfyUI, and its license terms are kept in the repository's LICENSE file, with credit to those earlier projects preserved.