I just wanna throw my two shekels in to the mini-AI discussion here, because I have some experience in this matter.
@Hey Johnny Bravo has already realized the difference between LLM-generated images and local AI (or 'Stable Diffusion'), but as other posters have noticed, that domain is highly populated by the 'coomer' set. The reason for this is, the most popular models out there for that sort of thing (Illustrious and Pony) were all trained on the 'SDXL' model. This model did indeed have a lot of porn and 2d waifu stuff in the dataset.
As a result? It's the best at doing anatomy in even non-NSFW requests.
The reason for this is Stable Diffusion is really great at
inference, i.e. "The muscles look like this under the clothes, so this is how I'll draw the clothes." This is the same reason art students take life-drawing classes: to understand the anatomy of the human body.
Later on, the model SD3 was released, and it was trained exclusively on SFW data. The result?
Anatomy was shit.
TL;DR - If you want a quality SFW picture of Johnny Bravo, you still need the coomer stuff underneath the skin-tight shirt.