The wave of AI image generators that arrived in 2022, DALL-E, Midjourney, Stable Diffusion, all rest on a counter-intuitive idea set out in a 2020 paper from Berkeley called "Denoising Diffusion Probabilistic Models." The recipe sounds almost perverse. Take a photograph and destroy it, adding a little random noise at a time until nothing remains but static. Then train a neural network to run the destruction backwards. A model that can reliably undo noise can, started from pure noise, produce an image that was never there.