Another attempt at learning LoRA making, had to run several iterations before was satisfied with the result. There is several hard-hitting problems with model - mainly with hands and tail, as well as inconsistencies with limbs and body color, most of them can be fixed with adjusting negative and positive prompts.

But the main problem with the model is stability, while it can produce pretty "good" results - it's requires nothing less but a total brute force through many seeds and careful prompt tuning, although simpler concepts are more stable overall.

Can do both SFW and NSFW, somewhat good with styles. Model strength can be pushed down as far as 0.6 without losing in character likeness, which should help with consistency and flexibility in situations with more complex concepts, I've found 0.8-0.9 is good enough for most runs.

Trained on hand-picked and edited shots from IDW comics.

*Images were generated in comfy-ui with a range of 17-30 steps/3.5-4.5 cfg - using Euler_a/Er_sde samplers. Images were generated with base model without any editing (one image was upscaled and passed through second Ksampler (can be spotted by blurry line in the bottom of the picture)- https://civarchive.com/models/2458426/anima-official