Something I haven't yet seen mentioned, but that is going through my mind. To me, it doesn't even seem like OpenAI got any better at producing GenAI images. Instead, it seems to me like they now simply removed a whole bunch of guardrails. Guardrails that, for example, made AI images shitty on purpose, so to be "safe" and allow people to kind of recognize. Making all of this "safe" was still very en vogue a few months back, but now there was simply a big policy/societal change and they are going with the trends.

This then allows their pictures to look more realistic, but that also now shows very clearly how much they have (presumably always) trained on copyrighted pictures.