>>48180727thank you!
>>48148897>>48153039https://litter.catbox.moe/pe5hb1.jpgJust a different version. It was supposed to be like a replacement fo 3.1. But it doesn't really work.
3.1 had about 200 vtubers with enough images and was overfit as hell. But it did it's job.
New version has ~500 vtubers, and tagging/filtering setup is better, base resolution is higher, etc. But it just doesn't work. As it learns chuuba X it starts forgetting chuuba Y. Maybe 500 is just too much.
I tried different settings, but validation graph is almost flat or goes up every time. Also better tagging actually makes prompting harder.
So fuck HLL. Instead I'll go label images for my scorer and wait for SDXL. Maybe final version of it won't suck too much and someone makes an anime finetune for it.
