A new study reveals how the diffusion patterns used by creative robots are based on math, not magic.

Image generators, such as DALL-E, Stable Diffusion and Imagen, are set in motion by diffusion models designed to reproduce training data. Instead of acting as copiers, they generate completely new images, combining ideas in ways that seem imaginative. The process is described as a paradox by Giulio Biroli, a physicist and AI researcher at the École Normale Supérieure. Diffusion models turn an image into pure noise, then reconstruct it step by step, resulting in something different but still meaningful.