SynAD: Synergistic Dual-View Adaptation for Generalized Anomaly Detection

Zero-shot anomaly detection (ZSAD) aims to identify and localise structural defects without target-domain supervision. While vision–language models such as CLIP have gained traction via prompt engineering, cross-modal alignment fundamentally struggles with subtle, non-semantic physical flaws that lack explicit linguistic descriptors. This work introduces SynAD(Synergistic Dual-View Adaptation for Generalized Anomaly Detection), a purely visual adaptation framework built upon frozen vision foundation models (VFMs). Rather than deploying separate view-specific experts or brittle cross-scale attention, SynAD conceptualizes macroscopic context and native localized details as complementary observation distributions of a single unified detector. By injecting shared Weight-Decomposed Low-Rank Adaptation (DoRA) modules into selected self-attention projections, both global and local views are forwarded through identical adapted parameters and decoded by a shared lightweight anomaly head via independent forward passes. During inference, overlapping local predictions are seamlessly stitched at native image resolution and harmonized with global logits using an evidence-balancing coefficient calibrated strictly on the corresponding held-out source validation split. Comprehensive evaluations across 15 industrial and medical benchmarks demonstrate that SynAD achieves highly competitive anomaly detection alongside robust, fine-grained localization, delivering pronounced gains on precision-sensitive metrics on challenging benchmarks (e.g. MVTec AD 2). Furthermore, we verify architectural versatility across representative visual backbones and substantiate practical real-world utility by introducing BoltAD, our field-acquired industrial fastener benchmark characterized by intricate non-semantic geometric flaws.