Unsupervised Detection of Distribution Shift in Inverse Problems using Diffusion Models

Shirin Shoushtari, Edward P. Chandler, Yuanhao Wang, Ulugbek S. Kamilov · 2025

Read Unsupervised Detection of Distribution Shift in Inverse Problems using Diffusion Models by Shirin Shoushtari, Edward P. Chandler, Yuanhao Wang, Ulugbek S. Kamilov (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-2505-11482.txt.

Description

Shirin Shoushtari and co-authors present in 2025 an unsupervised way to detect when image data have shifted away from what a model was trained on. Diffusion models are used as priors to rebuild images from corrupted measures, but their work suffers when test images differ from training images. Old shift tests need clean test images, which are rarely at hand during a rebuild.

Their score uses only the corrupted measures and the model's own score functions. Theory shows it tracks a standard measure of gap between the two image sets, and tests confirm the match. Aligning the out of group score with the in group score also improved rebuilt image quality across several tasks.

Catalog record

CREATORShirin Shoushtari, Edward P. Chandler, Yuanhao Wang, Ulugbek S. Kamilov
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE75 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-11482

Browse the full library · About the Tower of Babel