"Fine, I'll do it myself."
I wanted to add Sable to the roster of characters that can be queued in with Anima. Note that I have done some testing here, but admittedly rather limited testing. The tags I list seem to be necessary to ensure her hair tips and tattoo under her eye are kept. Interestingly enough, I also learned that style Loras may influence her appearance. For example, using this Lora I'm particularly fond of and would definitely recommend using (https://civarchive.com/models/2051219/wils-style?modelVersionId=2967302) some of my generations gave Sable blackish purple hair as well as heart tattoos elsewhere on her body. I also removed my tag of "purple hair tips" or "purple hair" to make sure it wasn't Anima taking those tags too seriously; see the three example images of Sable lounging on the bed. The one where she has all white hair with pinkish-purple tips is just using the checkpoint (highly recommend: https://civarchive.com/models/2692236/unholy-desire-mix-dark-serenity-anima-2b-or-anima-29b?modelVersionId=3045898), the one where she has black hair and purple tips is using "purple tips" as a tag with the style Lora, and the one where she has almost half white and half black hair does not use the tag but does use the style Lora.
I think that's pretty neat to be honest, because her identity didn't totally shift; it's just an alternative appearance of the same character :)
Regarding training, the dataset was only 23 images and honestly I think four are too close together the more I look at it. Also, interestingly enough, fan depictions of the character vary in terms of if her hair tips are more purple or more pink as well as whether she has a black heart tattoo on one cheek, both cheeks, or neither. Some of my dataset choices came down to liking the depiction of her physique while also wanting to ensure varied art styles so no one style was burned in.
Any and all feedback is welcome. If you have advice on how to make v2 better or run into any difficulties, let me know.
Description
First version. Trained for 2,000 steps at 0.00002 learning rate (32 linear rank) on a 3090TI.






