>>34703161The way to solve this (and what I'm working on w/ some people at Eleuther) is using verifiably anatomically correct images (automatically generated posed 3D renders of high quality models from many angles with different lighting) to focus on providing more reliable and curated data. Human art is always going to have biases, errors, mistakes, and most of the current training set is just bad. The problem is there literally isn't enough high quality 2D art in the world for current models to learn effective generalizations. Then it's a matter of training a solid style transfer model to turn the renders into something that people recognize as digital 2D art.
Using rendered images as training data solves a ton of problems and the next generation of models is going to be really impressive (there's a paper coming in January that will probably blow your mind.)
It also helps with the IP issues. You can license or commission a few thousand 3D models and poses and then use them to generate hundreds of millions of images. If there's anyone here with 3D rendering expertise interested in the field I highly recommend you reach out because this is what we need most right now.
