Trained over ~500 images, all cropped and labeled for each individual character/costume. It ended up with a strong "anime screencap" style though, so I recommend using it with a low strength value, like 0.4.
The hope was for a more generalized model (without needing a single model per outfit), but the resulting accuracy varies wildly, specially for the characters that had the least references, like Chitose. Kyouko's Tomato also ends up with eyes most of the time.
However, with luck and some coaxing, you can still get nice results.Its use as a clothing model for other characters can also be explored.
