Can diffusion models spot when medical images look wrong?

Medical imaging and other inverse problems often encounter out-of-distribution data—like CT scans with tumors when trained only on healthy livers. Existing anomaly detection requires examples of the abnormality or operates only on full images. This work uses the KL divergence between a diffusion model's prior and posterior to flag both whole-image and localized anomalies without calibration data. The method detected tumor-bearing livers and generalizes across diffusion architectures, datasets, and imaging modalities.