Trained on images from gelbooru, including tags.

This is still not quite what I wanted, but I'd say it gives positive results overall.

If you don't like the eyes add "round eyes" and maybe "anime" as well to the positive prompt, it helps a bit.

If you get unwanted speech bubbles, text or background characters, just add that to the negative and it should work fine.

Keep in mind that some models already know the tag nat the lich, but using the lora will still give different results.

I plan to train a v2 (if I at least make back the buzz I spent to train this one lol) with a different activation prompt and some more images removed, mostly stuff with multiple views/dialogue that I missed the first time.