The practical deployment of Visual Anomaly Detection (VAD) systems is strictly limited by their sensitivity to the complex, non-linear interplay between viewpoint and illumination, which governs the optical visibility of structural defects. Current benchmarks, however, largely overlook this physical reality, evaluating algorithms in sanitized environments that decouple geometry from photometry. To bridge this critical gap, we introduce Multi-View Multi-Illumination Anomaly Detection ( M 2 AD ), a large-scale benchmark designed to rigorously probe VAD robustness within a dense photometric-geometric configuration space. M 2 AD comprises 119,880 ultra-high-resolution images (3,648 × 5,472), systematically capturing 999 specimens across 10 diverse material categories under 120 synchronized imaging states. Leveraging this structured data, we establish two distinct evaluation protocols: M 2 AD-Synergy evaluates the capacity to synthesize information across diverse configurations (fusion), while M 2 AD-Invariant assesses model invariance to stochastic imaging changes (robustness). Our extensive benchmarking demonstrates that State-of-the-Art methods struggle significantly on M 2 AD , often failing to exploit the synergy of multi-view data. This work underscores the necessity for next-generation VAD systems capable of reasoning about imaging physics. The full dataset and test suite are released at https://hustcyq.github.io/M2AD to facilitate future research.
