The dataset is made of anime screencaps.

I'm pretty happy with the outcome, being able to depict all the outfits i can remember. Her dresses are a little off, but close most enough of the time.

Using tags "anime screencap", or just "screencap" might help with matching the likeliness, but will impact quality to some extent.

I'm always open for feedback, so I can improve my work in the future.