I haven't used stable diffusion enough to have a strong opinion on it. But my thinking is LLMs have only recently started contributing novel solutions to problems, so maybe there is some threshold above which there's less sloppy remixing of training data and more ability to form novel insights, and image generators haven't crossed this line yet.

Yeah? Those models are creative.

Who did then?