>>2771585I've heard of people being able to train with as little as 16, but that's probably gonna result in a limited set of poses.
I've used the crap out of this mina ashido lora because it's pretty high quality. The anon that trained it was kind enough to provide the dataset which contains 267 images. I think this is a good benchmark for something generalized.
https://mega.nz/folder/s8UXSJoZ#2Beh1O4aroLaRbjx2YuAPg/folder/sl0mlKiQ