Krea2 FinalCut

FP16

FP16 version has 95% reduction in wet skin and improved overall. For those who can use it, this is the best overall choice. Up to 2440x1920 native latent size.

NVFP4

This version will run 20-30% faster on most cards but has less detail then all other versions. However the "Wet skin" issue is more evident in this version more so when compared to FP16

Universal v2

Universal FP8 is my custom blockwise quantization that requires no metadata and should work for most users if Convrot is causing you issues. Convrot still has slightly better performance and should be used if possible.

V2 vs V1

While only marginally larger 13.8GB vs 12.2GB this training sees a massive seed to seed error reduction. Gradient Energy, Sharpness, Edge Density near equal. However a amazing 12.2% more unique RGB colors, in median average test.

• INT8 Convrot is recommended for most users.

• Official Comfy FP8 or INT8 Convrot scaling/quantization if not using universal

• 100's of hours training.

SFW Image Comparison

NSFW Image Comparison