>>2895874I'm always have mixed feelings with that with anime characters.
If you train lora with anime screenshot only they end very faithful to the source, but the images look incredibly stiff.
If you train them with fanart its more flexible but you can't pull accurate stuff.
And if you use both its mostly a roulette where sometimes you get something accurate but stiff or good but not accurate, or you get that weird effect where the face is completely 2d but the body looks like an illustration.
I wish you could do in a way where you get the flexibility of fanart trained lora with the style accuracy of anime screenshot trained lora
