AD-DINOv3: Enhancing DINOv3 for Zero-Shot Anomaly Detection with Anomaly-Aware Calibration

Zero-Shot Anomaly Detection (ZSAD) aims to detect anomalies in unseen categories without requiring annotations, offering a scalable and efficient solution. While most prior works rely on CLIP for vision–language alignment, we present AD-DINOv3, the first framework to adapt DINOv3 for ZSAD. This adaptation faces two challenges: domain bias between pretraining data and anomaly detection tasks, and the tendency of pretrained representations to confuse anomalies with normal foreground semantics. To address them, we introduce lightweight adapters for vision and text modalities to recalibrate features, and propose an Anomaly-Aware Calibration Module (AACM) that enhances CLS–patch interaction for anomaly-focused Image DINOv3 Ours representations. Anomalies are then localized by measuring similarities between adapted visual tokens and prompt embeddings. Experiments on eight industrial and medical benchmarks show that AD-DINOv3 consistently matches or surpasses state-of-the-art methods, confirming its effectiveness and generality for ZSAD.