>>34747869You can use textual inversion to add your own tags to the dataset by training it on a large batch of images of a single character.
It requires a ton of VRAM though so I can't do it on my shitty 1060.
WD is also trained on Danbooru, I think their dataset is 600k images which isn't really that big. Probably why almost all the images end up looking like Touhou.
>>34747744I tried to generate Pomu by doing
>blonde hair, red eyes, elf ears, fairy wingsbut it just gives me 2hus lol
