Diffusion models

Generating by un-noising

Language models generate one token at a time. Image generators mostly work completely differently, with an idea that sounds backwards: learn to remove noise.

1. Forward process (no learning): take real data and add a little random noise, again and again, until nothing but noise is left.
2. Train a network to look at a noisy example and predict the noise that was added.
3. Generate: start from pure noise and repeatedly subtract the network’s noise prediction. Structure emerges, step by step.