excosy/comfyui_bnb_nf4_fp4_loaders

This project is a small add-on for ComfyUI, a popular node-based interface for running AI image generation workflows such as Stable Diffusion and Flux. It adds two new nodes that let ComfyUI load model files that have been compressed, or quantized, into the NF4 or FP4 formats, which shrink a model's file size and memory use at some cost to precision. The two nodes it adds are one for loading a full compressed checkpoint file and one for loading just the UNET or diffusion portion of a model on its own. Once installed, you simply swap these nodes into an existing Flux image generation workflow in place of the regular model loading nodes, pointing them at compressed model files you have downloaded separately from sites like Hugging Face or Civitai, where NF4-compressed Flux checkpoints and UNET files are commonly shared. Installation is straightforward through ComfyUI's own extension manager, though the README notes you need to select the developer channel there, or alternatively install it directly from its git repository URL. It also requires the bitsandbytes Python library to be installed and an up-to-date version of ComfyUI itself. The code for the loaders was adapted from an existing implementation found in the Forge project, another Stable Diffusion interface. The README is upfront that this project may now be less useful than it once was, noting that a different compression format called GGUF tends to give better results for most users, though it mentions that combining this project's NF4 loader for the main model with a GGUF loader from a separate project for the T5XXL text encoder can still give some users a speed benefit.