The Hidden Engine of Vision with Peyman Milanfar (Google)

Apr 10, 2026

Peyman Milanfar, a Distinguished Scientist at Google who helped build the Pixel camera pipeline, explains why denoising is a fundamental engine of modern vision. He links denoising to manifolds, RED and diffusion-like processes. Conversations cover mobile photography, BM3D to neural improvements, one-step diffusion on Pixel, perception-distortion tradeoffs, and why visual intelligence may be next.

Ask episode

AI Snips

Chapters

Transcript

Episode notes

INSIGHT

RED Turns Denoisers Into Priors

Regularization By Denoising (RED) repurposes a denoiser as an energy term: use the denoiser residual as a prior in inverse problems.
Unrolled RED iterations resemble diffusion: alternate denoiser and linear operators—effectively projecting toward the image manifold.

INSIGHT

Denoising As Manifold Projection

Denoising projects noisy observations back toward the clean-data manifold, often landing on a tangent space rather than the exact point.
This projection gives an implicit map of the manifold that guides reconstruction and generative sampling.

INSIGHT

Geometry Of Noise Explains Diffusion Behavior

There's a bridge between energy-based models and diffusion via a marginal energy integrating noise levels, but direct gradient descent fails due to singularities.
Flow and direct-signal models act like natural gradient descent, avoiding singular wells that naive energy descent would hit.

Get the Snipd Podcast app to discover more snips from this episode

Get the app