"Wait a minute! Haven't you already made this model!?" I hear you ask. Yes, but also no.

Whereas my other model was trained on illustrations of characters in leotards, this one is trained on 99 images of gymnasts. This means that leotards, poses, body shapes, hairstyles, etc, specifically for gymnastics are more accounted for. As such, I thought it was worthy of its own model, rather than just versions in my other leotards model.

v3
In my ongoing adventures of not knowing what I'm doing, I tried doing a different keyword structure than I normally do. Use ds_gymnastics_leotard in the positive to activate the lora, and use photorealistic in the negative more an anime look.

v2
This version came about from my studying into better training techniques, as well as learning to train on my own computer. As such, I was trying many different techniques, and stumbled across this version. There's no realistic tag, and I even removed several other superfluous tags (based on the article I read) such as 1girl.

It's not as good at holding style, sometimes wavering between a realistic and an anime look depending on how it feels in a given session, but when it hits, it's a home run, in my opinion.

v1 Anime Version

This version was trained with an activation word and the realistic tag, so it seems to hold anime styles much better than the semi-realistic version.

v1 Semi-Realistic Version

This version was trained without an activation keyword and without the realistic tag. The realistic style bleeds through to images more and distorts anime pictures more, but I think it has plusses and minuses compared to the anime version.

I like both versions a lot.