Anyone got any idea how this would work at a high level? I get that for static images some neural nets are “reducing” noise from random/base images based on prompts, but is this based on the same idea?

Absolutely magical.

Bleh. It’s so good. I hate it, but it’s amazing.