My first serious attempt at lora training. There's probably a lot that can be improved but I'm happy with the current results.

There doesn't seem to be as much loras/embeddings for characters from western animation so I decided to make some of my own.

NOTE: I use two negative embeddings in the generated images uploaded here, bad-artist and bad-hands-5. You can find them just by googling `bad-artist embedding`. Similar results are achievable with standard negative prompts.

I don't have a great understanding of lora training/concepts, but I trained a few different concepts in this lora. The two main concepts were images of catra in her military outfits with her red mask on, and the prompts had 'catra' only with other descriptors. Another concept was catra in her red dress suit, for those I used 'catra in a suit' prompt when training. It seems to trigger these correctly often enough. There's also a 'catra without mask' concept, although the mask disappears anyway if you start prompting hair/face details.

The Lora was trained on NAI, with clip skip set to 2. You'll get best results on NAI-like models, but I did get it to work on some realistic models.

For generation I mostly use DPM++ 2m karras or Euler a. Most of the photos I've created were pure txt2img with latent hires upscale. For a few with more complex poses, I used ControlNet with some catwoman comics images.