ComfyUI-nunchaku

Why we included this project

Consumer GPUs hit a hard VRAM ceiling when running flagship text-to-image models like FLUX or Qwen-Image. This plugin drops Nunchaku's 4-bit inference into ComfyUI so those models run quantized as ordinary custom nodes you can drop into existing workflows, and it extends to Z-Image and image-editing checkpoints as well. LoRA, ControlNet, and PuLID keep working alongside the quantized weights, which usually matters more than raw model support since add-ons are the first thing people refuse to give up. For a 20-series card or any setup with limited memory, this is a practical way to run current open-weight diffusion models at interactive speeds without renting a cloud GPU. Pre-quantized model files and bundled example workflows spare you the work of compiling an inference engine yourself.