This is the result of merging a large number of NoobAI and Illustrious V-Pred models with various styles and intentions, using my own iterative merging method which involves merging the same two models twice, once forward with adjusted block ratios and LoRas applied to model B and then again in reverse using DARE.

The goal is to make a multi-purpose checkpoint that knows a lot of subjects, characters and styles that can be influenced easily with loras and prompts.

This checkpoint's weaknesses are mainly that it's not the greatest at hands, and that some prompt & seed combinations are oversaturated.

Usage Tips:

Don't forget to enable V-Pred sampling and Zero SNR/ZSNR. Generations won't look right without it. With ComfyUI this is done with the ModelSamplingDiscrete node

Start with 4.0 to 5.0 CFG and 25-35 steps, and adjust as needed.

Example images all use hi-res fix with 0.1 to 0.2 denoise and so the actual total step count is double what is listed. It seems you need lower denoise than EPS models for the same effect.

I usually begin my positive prompts with:

(very awa, high quality:1.1),

Followed by style and artist names such as:

traditional media, oil painting \(medium\), absurdres,

And then whatever characters, poses, objects, viewpoints etc.

Use a mixture of danbooru & e621 tags, and very simple sentences.