Famous and talented cosplayer Danica Rockwood. She does a wide range of styles for her cosplay and modeling so I thought it would be neat to have her be able to work with some stable diffusions.
Tested with a good number of models with fairly consistent results at getting an accurate (but not always perfect) representation. Tested on mostly realistic styled models, but it may work on some anime ones too.
Please let me know what you think and if there's any errors you see, it will help me learn and improve if it needs any fixes.
Trained at 5000 steps with a variable learning rate. Tested lesser steps and they all looked decent, but figured I'd keep the 5000 one.
I am semi new with creating Textual Inversions, but I do think the result here is good. I was aiming for versatility so this is trained on 1.5, as I heard that's the best way to get your TIs to work with a wide range of models.
All of the outputs were with Restore Faces enabled. I did have some okay results with it off with some early tests, but I enabled it at a certain point and forgot to alternate.
I used a range of 5-12 CFG for testing, it started to get a decent representation around 9-10, but it didn't look awful at around 7.
Models used for test images (should be in order unless CivitAI shuffles them):
https://civarchive.com/models/4729/bastard-ai-impossible-fantasy-sex
https://civarchive.com/models/5935/liberty
https://civarchive.com/models/5952/experience
https://civarchive.com/models/4384/dreamshaper
https://civarchive.com/models/12862/stablydiffuseds-magnum-opus
Description
Trained at 5000 steps on 20 images with Automatic1111 using a variable learning rate. 5 vectors per token.