Krea 2 NVFP4 Matmul Patch
A community Python script that copies a ComfyUI NVFP4 safetensors file while clearing full-precision matrix-multiplication flags in its quantization metadata.
What happened lately
Krea 2 NVFP4 Matmul Patch timeline
This is a local Python utility for an existing ComfyUI model file, not a model checkpoint or a hosted generator. It rewrites the safetensors JSON header and copies the original tensor bytes into a separate output file. Its repository contains no model weights and labels the project Apache 2.0; the separately obtained Krea weights retain their own license.
The author reports a warm-render improvement on an RTX 5080 with ComfyUI 0.42.3, while explicitly leaving other GPUs, other ComfyUI builds and broader image-quality testing unmeasured. Users need the existing model file and free disk space for a second copy. Removing the metadata flag changes the computation path, so unchanged tensor bytes do not establish unchanged output quality.