NeurIPS

Unsupervised anomaly detection from the mismatch between a flow matching model’s learned velocity and the direct velocity toward a test image, giving pixel-wise heatmaps and image-level scores.