Noise In, Noise Out

Diffusion's premise sounds absurd: take a clean image, add a little Gaussian noise, add a little more, and keep going for ~1,000 steps until nothing remains but static.

Then train a network to reverse each step — given a slightly noisy image, predict a slightly less noisy one. If the network can walk backward down the corruption staircase, you can start it at the bottom — pure noise — and let it climb to a clean image that never existed. That's the whole algorithm.