By framing anomaly detection as recovering normal features through conditional flow matching, Flownomaly achieves strong few-shot defect detection with a single inference step and no need for anomalous examples during tuning.
Flownomaly detects manufacturing defects from just a few normal images by learning to recover clean features from potentially anomalous ones. It uses a flow-matching model in frozen feature space to transform noisy tokens toward normal patterns, then flags deviations as anomalies.