Here's a blog post I liked that explains a connection: https://sander.ai/2024/09/02/spectral-autoregression.html

They call diffusion a form of "spectral autoregression", because it tends to first predict lower frequency features, and later predict higher frequency features.