rockerBOO/krea2-raw-nvfp4

Krea 2 Raw — NVFP4 quant for ComfyUI

Quantized version of Krea 2 Raw for ComfyUI, using NVFP4 for the main transformer blocks with ComfyUI-compatible metadata.

Files

File	Size	Status
krea2_raw_nvfp4_convrot_int8.safetensors	12 GB	Best quality
krea2_raw_nvfp4.safetensors	8.3 GB	Tested

Requirements

Blackwell GPU (SM >= 10.0/12.0) required for inference — RTX 50-series, B100/B200.

BF16 vs NVFP4 comparison

Same seed/workflow/LoRA, only the base model swapped. Generated in ComfyUI.

BF16	NVFP4 (this repo)

License

Krea 2 Community License, inherited from krea/Krea-2-Raw.

---

Quantization method

Quantized with silveroxides/convert_to_quant (ctq):

ctq -i krea2_raw_bf16.safetensors -o krea2_raw_nvfp4.safetensors \
    --nvfp4 --comfy_quant --save-quant-metadata --low-memory --krea2

--krea2 keeps the text-fusion branch (txtfusion.*) and a handful of small first/last layers (first, last.linear, tmlp.*, txtmlp.*, tproj.1) at full precision; all 28 main transformer blocks (attention + MLP, 224 of 264 weight tensors, ~95% of weight parameters) are quantized to NVFP4 with learned-rounding optimization (SVD-based, the tool's default — not RTN/--simple).

An earlier build of this file used --simple (RTN, no learned rounding); this one replaces it.

NVFP4 + ConvRot INT8 variant

krea2_raw_nvfp4_convrot_int8.safetensors is a mixed-format build on top of the NVFP4 file above: attention layers in the 24 interior transformer blocks use ConvRot INT8 instead of NVFP4, MLP layers stay NVFP4, and the first 2 and last 2 transformer blocks are kept at full bf16. Same Blackwell GPU requirement as the NVFP4 file.