Krea 2 NVFP4 patch claims a 2.5x speedup on RTX 5080
A community Python script clears a metadata flag in ComfyUI's Krea 2 Turbo NVFP4 file. Its author reports 11.6-second warm renders falling to 4.6 seconds, with quality tradeoffs still open.
A small metadata change, rather than new weights or fewer sampling steps, is behind a community claim of roughly 2.5 times faster Krea 2 Turbo generation. Hugging Face user sjoe1244 has published a Python patch and a timing comparison for ComfyUI’s NVFP4 model file on an RTX 5080.
The November 2 repository contains findings and patchflag.py, not a downloadable model. It is a community experiment, not a Krea or Comfy-Org recommendation. AITrending read the script and English documentation but did not reproduce the GPU measurements.
What the flag changes
The author’s finding is that krea2_turbo_nvfp4.safetensors from Comfy-Org includes full_precision_matrix_mult: true in the quantization metadata for all 256 quantized layers. The card says that setting causes repeated conversion to bf16 and bf16 matrix multiplication, rather than the GPU’s FP4 computation path.
The script removes that flag from layer entries in the JSON header, writes a new header and copies the original tensor bytes into a separate file. That supports the narrow description of a metadata patch. It does not prove that the resulting computation or generated picture is identical.
The reported 11.6-to-4.6-second comparison
The author reports warm renders of 11.6 seconds for the downloaded file and 4.6 seconds after patching, versus 4.5 seconds for an older community NVFP4 file without the flag. The stated setup uses an RTX 5080 with 16 GB, ComfyUI 0.42.3, PyTorch 2.13.0 with CUDA 13.0 and SageAttention enabled. Images were 864 by 1536 at eight steps and CFG 1, with the same seed and prompt.
Those are one author’s measurements on a specified setup, not a general speed guarantee. The card does not establish results on other GPUs or other ComfyUI builds.
A local workflow with an unresolved quality question
Users run python patchflag.py input.safetensors output.safetensors against their existing model file. The author estimates about 7.6 GB of free space for the copy and says no text-encoder change is needed. The script itself uses Python’s standard library. Comfy-Org’s English card places the original diffusion model under ComfyUI’s models/diffusion_models directory; the patch does not supply the remaining generator components.
The key uncertainty is why the original flag was set. The author explicitly leaves open whether it was intentional, such as a safeguard for quality, and says clearing it also sends activations through FP4. Their quality check covers only one seed: similar style, but a somewhat different composition. There is no broader quality benchmark or blind comparison in the card.
That makes the patch a concrete experiment for an existing local setup, with a measurable claim to investigate. It is separate from reducing Krea’s sampling steps with a LoRA, and the repository’s Apache 2.0 label does not replace the Krea weights’ own community license.
