Zero-Shot Anomaly Detection

Zero-shot anomaly detection asks whether a model can flag, and often localize, patterns that depart from normality in a new domain without target-specific anomaly labels or training. Common approaches reuse frozen vision-language or vision foundation models with prompts, normal-reference comparisons, and patch or feature distances; others use geometric or frequency cues, lightweight adapters, or multimodal reasoning across images, video, 3D scans, time series, and graphs. Evaluation spans industrial defect benchmarks such as MVTec AD and VisA, surveillance video such as UCF-Crime, and medical, graph, and sensor data, using detection and localization metrics. A central tension is transferring to genuinely unseen anomaly types while balancing auxiliary data or prompt tuning against precise localization, calibrated recall, and practical inference cost.