My first crack at a customized Anima merge. Trained on a style LoRA created with about a hundred images generated with several of my Illustrious models, this is essentially that LoRA trained into Anima Base V1, (hopefully) capturing the aesthetic style I've been gradually tuning over the last two years.
Description
Launch edition.
Recommendations: CFG scale of 1.5+ (2 seems to be a sweet spot), 15+ steps, simple negative prompting.
Observations: Seems to struggle a bit more on wider aspect ratios, which will skew faces and hands.
Future plans: Further training on my aesthetic style.
FAQ
Comments (5)
is the base model correct? ;)
Apparently, selecting Anima in the setup process did not set it up in the model info panel. Should be fixed now.
@AI_nstein all good. :D
This is a good checkpoint for solo gals. Vertical aspect ratios work much better, for some reason. Multiple characters it gets a little confused, 2 characters it is fine.... could be my prompting skills, or lack thereof. Plays well with character Loras, adds a nice style to my eye.
CFG 2 / Steps 15-20 / Euler A/ Simple or Beta
I'm honestly proud that it performs as well as it does with how inexperienced I am at A: Using Anima and B: Baking my own LORAs for training. More testing with this version has revealed that it will do better at details like foggy fingers or faces with a few more steps; I keep it at 24+ now. I'm actively working on training up a second version, which (hopefully) will address some of the things V1 lacks.
I anticipate this being a very long, ongoing project, and I'm looking forward to providing the community with a useful, enjoyable tool in the long run, if it works out.