[XorNot]

Or it accelerates the endpoint: cutting the pre-made art out of the process entirely.

AI image generators don't need to see art, they just need to know what blobs of pixels "are" in relation to words. That sort of data can be extracted from just photography of the real world - it just turns out there's a lot less of that easily available and properly tagged then art collections right now.

There's more then enough public-domain examples of "style" to do the rest (and style-transfer was one of the original AI image manipulation applications).