Abstract: Traditional diffusion models are typically trained to predict only the mean of the denoised distribution given a noisy sample. But what if we go beyond the mean? This talk explores how incorporating additional information—such as predicting the covariance of the denoised distribution—can significantly accelerate sampling and improve density estimation. We’ll dive into different techniques for covariance prediction, their theoretical connection, and practical benefits for more efficient and expressive generative modeling.

Slides from the talk given at University of Cambridge Engineering Department.