It's a shame that generative image models have pursued realism so relentlessly. Reproducing training data is uninteresting and exploitative.
The imperfections of early models were genuinely interesting, and will end up being the most artistically interesting artifact of all this in the long term
