Creator: https://civitai.red/user/OneRing

About this version

Surprisingly, the new release is about 2.5 × more accurate, largely thanks to the updated Rouwei architecture. I enriched the dataset with complex poses and non-standard proportions to boost the model’s diversity. After a few targeted tweaks to the settings, the color palette became richer, and I achieved details that were simply unattainable in the previous version.

Fidelity to the prompt has improved as well: depending on the scene, the model now renders fine-grained details about 20–50 % more accurately than before.

I also toned down the excessive contrast. Picture the softness of the epsilon version combined with all the advantages of the v-pred release — in short, the artwork now looks noticeably smoother.