I have not enough VRAM to even upscale examples

Even SFW prompts will contain NSFW content usually.

Use small weights for lora (0.4-0.8) and keywords (0.9-1.2) for best results.

• Model works better on 512*512 images

• Fast Negative is a good negative embedding for this model (use small embeding weight like 0.9 - it works but not so intensive in ComfyUI). Any embedding including this can reduce hentai quality (it looks more generic). As example, fastnegative makes multiple girls very similar to each other, also it hates colored hair. So try to use keywords instead of negative embeddings (worst quality, bad quality, sketch, etc.)

• Good performance on monochrome (hentai) manga images.

• More tags = more unique result (you can copy tags from rule34 or booru images and it will create something similar)

• Preview examples are without any embeddings, so quality can be better

• Image quality is not so good compared to other popular models, but hentai looks more natural.

• Also it works good enough for "unusual" (for me) types of hentai like yaoi ("2boys" tag is enough), futanari, "traps", etc.

The main core for this model was only on huggingface for short time. I can't find it anymore. Quality was horrible and sometimes results looks like nonesense abstraction, but successful results were not so bad.

I tried to mix it with other models and honestly it was a fail at begining. I have used small 20% merges and then merged the best results, also mixed it with other models sometimes.