Quantize diffusion models to W4A8_INT8, NVFP4, MXFP8, INT8 ConvRot or INT4 ConvRot using ComfyUI's internal kitchen. Per-architecture profiles keep sensitive layers in bf16. Supports Minimax H3, Krea 2, Z-Image, Flux 1/2, Qwen-Image, Wan, LTX-2, Chroma, Ideogram 4, ERNIE, Anima, ACE-Step, SeedVR and Boogu.
https://github.com/tritant/ComfyUI_Kitchen_nvfp4_Converter
git clone https://github.com/tritant/ComfyUI_Kitchen_nvfp4_Converter
comfy node install ComfyUI-Kitchen-nvfp4-converter