>>39916721It's based on the first version. I've added almost everything with virtual youtuber tag from danbooru, added a lot of artists, added more niche nsfw stuff.
Unfortunately, now i realize i forgot to blacklist some stupid tags like *(cosplay) and 'alternate costume' for vtubers. Also nsfw stuff have too many old images and that negatively affects art style. And i forgot to disable text encoder training at the end, so it will overfit.
Ideally, i should write a script to clean dataset and also balance classes, but i'll probably just finish training as it is, and do it properly some time later using different base model. (WD1.4 or ACertainty maybe?)