/vt/ - Virtual Youtubers

>>45535234
Aspect ratio & size doesn't matter. Can be wide, portrait, square. Model variety is good if you don't want to ingrain it into the model, but not if they look too different. Just make them look pretty and as consistent as possible (i.e, not blue eyes in one and yellow eyes in another). A mix of upper body, portrait and full body is the way to go. Retraining Lua because the hands are fucked on 1.0. This is roughly how a dataset looks like. Don't try to make it perfect. We can do it just for fun and see if anything good comes out. If not, your prompting is good enough.