[FLUX.1 dev] Yuzusoft Style

This FLUX.1 [dev] LoRA helps generate anime-style images with stylistic characteristics of the Yuzusoft eroge brand.

Recommended usage:

• Weight 1

• Include that the image should be anime style in your prompt.

  • I usually throw in "anime wallpaper," somewhere.

  • FLUX seems somewhat resistant to forgetting realistic and non-anime illustration styles, so you may get unexpected results if you don't specifically prompt for anime even with a LoRA applied.

• Long descriptive natural language prompts work well with FLUX.1, but you can get away with specifying just a few details/tags and letting it figure something out.

• Euler simple or Euler beta 30-40 steps seems good for most generations. More complex generations may need more steps than simpler ones

• Distilled CFG Scale from 3.5 to 5 seems to give good results

Dataset info:

• Select images from Yuzusoft titles from Senren Banka and onward were included

• High quality fanarts drawn in the modern Yuzusoft style were included

• I haven't really figured out a good NL captioning flow yet so I just used danbooru tag files that I had already. Please don't expect it to respond significantly to any tokens from the training data captions though.

Other notes:

• Sometimes it may start to draw nipples on cleavage. I think this is because of the dataset, and that Flux doesn't really understand nipples well enough to figure out where it shouldn't be drawing them. If you get this please inpaint them out or alter your seed.