Seminar, Hosted by ICTP AI for Scientific Discovery Colloquia Series
Generative AI and Diffusion Models: A Statistical Physics Perspective
École Normale Supérieure, Paris
Abstract
Diffusion models, the state of the art for generating images, video and sound, start from white noise and evolve it through a time-reversed Langevin process into complex outputs, which ties them closely to stochastic thermodynamics. Giulio Biroli shows how statistical physics supplies principles and methods to characterise this generation process: the emergence of features can be understood through symmetry breaking and dynamical phase transitions, while creativity, or its absence in the form of memorisation, can be explained with theories of disordered systems and of dynamics in complex energy landscapes. He closes by discussing why diffusion models outperform earlier generative methods, highlighting a deep connection between their multi-scale generation mechanism and renormalisation-group analysis of physical systems.