49 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Distributional Diffusion Models
Google DeepMind
Recording
Abstract
Standard diffusion training estimates the mean of clean data conditioned on a noisy sample. This talk instead learns the entire conditional distribution using a scoring-rule objective. Sampling that richer posterior permits larger reverse-process steps and fewer inference iterations. Experiments with images and robot trajectories assess the resulting speed–quality trade-off and compare it with conventional diffusion models.