Interestingly, I think one reason the newest model out there now, Flux, is a lot better is because it got trained on worse sources, namely smart phone selfies. Part of the challenge so far has been trying to find somewhere between the cartoon/anime versions of things and the super resolution cinematic pictures. They both still have an uncanny aspect to them. But make the picture a bit worse, fuzzy, or basically what people tend to take and it looks a lot more real.