No mention of GANs? As far as I remember, their images felt much more like "sampled from training data" than "averaged from training data". The idea was that you train a "discriminator" that tells generated images apart from real ones, and make the "generator" try to fool it. But somehow it lost to diffusion models, and now all AI imagery looks like slop, when GAN outputs (though imperfect) didn't look nearly as slop. I don't really know what happened though.