KLIP: localized distribution shift detection via KL-divergence with diffusion priors in inverse problems

Researchers introduced KLIP, a new OOD detection metric for inverse problems that compares a diffusion prior with the posterior using KL divergence.

KLIP works without calibration data or prior knowledge of the shifted class, and it can detect both full-image shifts and localized anomalous patches.

The method performed strongly across multiple models, datasets, and inverse problems, including subtle semantic changes in liver CT scans such as healthy versus tumor-bearing cases.

This could improve medical and other computational imaging workflows by flagging distribution shifts more reliably.