/vt/ - Virtual Youtubers

>>47513368
yeah, this was an issue I had.
when I was training that version of the model I just threw in every variant of that image (there were like 15?), and I was also over training with too high of a learning rate, so almost every image it spit had out a similar composition and background. Even after dropping the learning rate, until I pruned out most of the variants from my dataset, they all had a similar composition of sitting in a japanese style room. Here's an example of the problem. This is the first LoRA I've made, so I'm making a lot of mistakes along the way.