> If an AI art engine outputs a frame of solid blue, is it infringing the copyright of Yves Klein's solid blue "IKB 79"?
Probably not, though even this may be debatable given the specific prompt and specific similarities (for example, if it generated the exact color and exact aspect ratio for a prompt like "Yves Klein IKB 79", I could see an argument for infringement; if it generated the same thing for a prompt like "filled-in 16:5 rectangle with color #0000FF", it would be arguable that it isn't).
> I think that some artists' styles can be accurately replicated without training on any of their work: because the artists' style is generic enough that it can be exhaustively encoded via the works of others.
The important question in copyright is, theoretically, if you actually copied the specific piece or if you happened to create a similar-looking piece by accident. The difficulty of proving one or the other varies with the specific circumstances. For example, it's rather hard to claim you happened to paint a picture almost identical with Picasso's Guernica. It's rather easy to claim that signing an empty piece of canvas was entirely your idea and you weren't copying Dali's signed empty canvases.
Similarly for AI, the copyright discussion will come down to how much of its output is identifiable pieces of other's artworks, especially when prompted for such. I personally haven't played with it, but if it's possible to get SD or other similar models to produce identical or very similar copies of (pieces of) some artist's works using relatively simple prompts*, it should be a pretty slam-dunk case of copyright infringement.
* by this I mean prompts that don't themselves encode the information, like I showed in my "16:7 rectangle with solid color #0000FF" example.