The Hidden Engine of Vision with Peyman Milanfar (Google)

Peyman Milanfar, who leads Google's Computational Imaging team, argues that denoising, long seen as routine cleanup, is one of the most fundamental operations in modern machine learning: knowing how to remove noise amounts to having a map of the manifold on which signals lie, linking classical inverse problems to diffusion models. The conversation traces early patch-based denoisers, his 2010 paper asking whether denoising is dead, and Regularization by Denoising, which resembles a diffusion process when unrolled. It also covers a one-step diffusion model shipped for zoom on Pixel phones and the trade-off between perception, distortion and authenticity in generative imaging.