Its not exactly the best, but It works for me. I found the avatars incredibly cute so its LoRA time. Do note that this is my first actual LoRA so I have no clue what I'm doing.

• Trigger word is supposedly "BaTAvatar", but it seems to work without it, so idk. If you're having issues, just try adding it and see what happens.

• Weight of 1 should work fine, but I opt to use 0.9 anyway. Its up to you.

• Some additional tags that are unneeded, but just incase you need to give it an additional nudge, is "[[chibi], [pointy ears], [tail]], [thick arms, thick legs]" but the LoRA itself should be fairly sufficient (at least from my testing).

I might come back later and try to refine this further, but that's unlikely.

Note that I'm an idiot and didnt bother adding any character tags, so this LoRA techincally can't generate canon-characters on its own. You'll have to create them through generic tags alone (ie like, "pink hair, blue eyes, long hair" if you want to attempt to generate Mizuki Himeji's avatar)

One last thing, the model i'm using is simply just Anything v3, it works fine for me, I don't plan on upgrading any time soon (I am aware Anything v4 has been out for ages). In theory it should work on any anime-based model, but I haven't tried it.

It was Trained on animefull-final-pruned-fp16.safetensors, whatever that is. Refer to the link down below

All hail this guy because now I don't need to train textural inversions at a snails pace on my old laptop, I can just force google to train LoRAs for me and get it done in less than an hour instead of a week + now its a LoRA and not an Embed