>>104452843Yeah way better prompt adherence and natural language/danbooru tag support, all the images there were made with nlp. Picrel (age old benchmark lol) is the prompt
"Create a anime style image of #shigure_ui_\(vtuber\) (grey serafuku, beret, pom pom hair ornament, loli) holding a red sphere while standing on top of a blue box with a green cylinder floating on the top right, chibi #oozora_subaru in the background in bottom left,
white background, "
It also uses flux vae so theoretically training LoRA for styles/character could be done faster and with much better quality with a lower dataset. Though the vram requirements right now is over what I can do (16gb vram) I am hoping it's possible to bring this down if it gets better support in some other better training UI.
