I have the opposite impression. For me the quality is the cake; polish is just the icing. Putting out a 5-min of effort output with an AI as your image, is all icing and no cake: it has all the right proportions, layout, etc, but something is very wrong with it.

Not sure how diffusion models work in detail so consider this wet neural slop, but I feel like the training data had a sizable portion of it of high quality art, photography, design, and the models captured mostly the polish: the crispiness, the geometry, the contrast: they're after all computable features. It's more difficult to capture the meaningful blurs, the deliberate deviations, and the organic rule breaking. They're properties much harder to "featurize" for, and rely on tags/labeling/descriptions, etc of variable quality.

By all means you can get that with AI anyway, but it also requires you to understand the art itself, and describe it in a way the model produces satisfactory results.