/vt/ - Virtual Youtubers

>>43648437
What you can generate is limited by how much of them was in the training set, how easy it is to describe them by features when their name alone won't suffice, and how willing you are to accept the art being a little off

A while back someone here finetuned the leaked NAI model with a bunch of vtuber art as training data which obviously helps a lot. Well, I say "a while back" but I'm pretty sure they're currently working on yet another version of that model.

Picrel is trying to prompt Mio just by name on the original NAI leak (the full one, aka does nsfw but also has worse character recognition) vs a model that's a descendent of the one specialized on vtubers. Obviously you can improve NAI's by actually trying to describe her features as well so the comparison isn't totally fair but I think this is a good example of a model that clearly knows who a certain character is vs a model that doesn't