This is my first time training a LoRA on local setup, so of course I’ve burned it. Still, it appears to be somewhat stable and flexible to some degree.

Because of overtraining and some mistakes at tagging dataset - some concepts may be hard to get, especially the ones which changes body proportions or form (like turning character into a character doll or a plushie). It also may severely weaken artist styles (some are working fine, some are not so), as well as injecting other characters into a prompt as it forces unwanted characteristics to them.

But, at the very least, main features and likeness of a character are baked very truthfully to original sprites.

*Images are generated in comfy-ui with a range of 17-30 steps/3.5-4.0 cfg - using Euler_a/Er_sde samplers. Most of the images generated with base model - https://civarchive.com/models/2458426/anima-official