(The NEW label is misleading, it's an update.)
Trained with the goal of enabling multi-subject capability but didn't turn out well. At least there's a positive probability of multi-char working. Still need to figure out why it fails.
In addition to the character tags, custom style tags were added:
fanart
: fan arts from Danbooru (optional, neutral style)
u149ani
: screenshots from the airing U149 anime
apuri
: Cinderella Girls and Starlight Stage mobile game art (somehow dominated by early CG style, will probably retag in the future)
As the outfits were not manually tagged, it should not be expected they can be replicated.
Training data can be found in my data repo. Training details will be properly documented when I have time. Would appreciate suggestions or any sort of feedback.