Same reason that self driving cars have plateaued. The things that are really important for that last mile are also really hard and we're not really sure how to get there.

The output between all the major AI image synthesis tools, while impressive, is quite clearly very high dimensional interpolation, and it shares the same limitations in that space as it does the 2d space. We've already thrown insane amounts of data and compute resources at the problem and it's made the things that were done well in the past even better but doesn't move us towards solving the hard parts.