>>47850781Oh, I finished my Yukino LoRA. Most of my dataset came from pixiv so I used an interrogator, to tag the images. I think it was one of the wd 1.4 ones. After that, I spent a bunch of time training, testing, pruning the dataset, and training again. Danbooru tags are usually high quality, but the interrogators usually do a good enough job too. If you have images that have multiple characters in them, you probably want to either get rid of them, or edit them so that they only contain the character you're training. also If she has an asymetrical design, you should turn of flip_aug or it will cause more problems than it'll solve.
