I created this model because I wanted to generate an OC for a story, and found that most checkpoints used heavily stylized leather or vinyl leotards, when prompted. This LoRA attempts to recreate gymnastics or ballet cloth leotards.

I am still very much a novice when it comes to training models, and this does have a few limitations. Regardless...

Recommended to use at 0.7 weight.

Keywords I tried to train:

• leotard (you can specify various colors, though some colors have no representation in the data set)

• two-tone leotard (specify both colors alongside, such as two-tone leotard, red leotard, white leotard)

• long sleeves, short sleeves and sleeveless variations of leotard

• intricate details will assist with getting interest leotard designs

• white pantyhose to include tights in the outfit, or bare legs to ensure they're removed

• leg lift pose

• standing on one leg, arm up, arms outstretched, etc. will return more gymnastics-like poses

• stretching and split poses have some representation, but are still hit or miss

• holding, ribbon baton will assist with rhythmic gymnastics pictures, but I think they're too detailed to effectively replicate well

Known Limitations:

• Some colors have no representation. For instance, I could not generate a purple leotard at all. Some basic editing and img2img could possibly get around this.

• Colors often bleed from hair or eye prompts, especially on lesser-studied colors

• Poses, especially the split, will often return bizarre results, but good results do come through occasionally!

Only one example image has been inpainted or upscaled at all. The other images are just as they were generated initially with no ControlNet.

All in all, I'm quite happy with this model! It does what I set out to do, plays somewhat nicely with character LoRAs, and produces results that can be upscaled pretty well. I even tried it on a more realistic model (once), and was not unhappy with the results.

I hope you enjoy it as well!