That's an astute observation, and there actually is a similar mechanism at play for both. In dreams you're putting together mental pictures in a different way than when it's coming off your optic nerve, there's no "truth data" to correct from. In diffusion it's generating pictures according to a classifier, starting from random noise. So it can get anchored incorrectly as the noise level goes down, and there's no real anchor to correct it. So you get the start of fingers, fingers at wrong angles in an area where a hand is forming. And by the time it's done, they are botched. Same for words.