Posterior Sampling by Combining Diffusion Models with Annealed Langevin Dynamics

Zhiyang Xun, Shivam Gupta, Eric Price · 2025

Read Posterior Sampling by Combining Diffusion Models with Annealed Langevin Dynamics by Zhiyang Xun, Shivam Gupta, Eric Price (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-26324.txt.

Description

Zhiyang Xun, Shivam Gupta, and Eric Price ask a sharp question about working backward from noisy data. A sensor gives a blurred, noisy view of some hidden signal, and a good model of the signal exists. When can a computer draw fair samples of what the signal might have been, given that view? Such sampling helps with tasks like filling in missing parts of a photo or cleaning up scans.

Xun and his two co-authors focus on a class of smooth, bowl-shaped distributions where older methods can work but break easily when scores are slightly off. They prove that mixing diffusion models, which learn to undo noise step by step, with a heated form of Langevin dynamics works in fast time. Their method also needs only a weaker bound on score error.

Catalog record

CREATORZhiyang Xun, Shivam Gupta, Eric Price
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE146 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-26324

Browse the full library · About the Tower of Babel