<p>Image Anomaly Detection (IAD) has achieved remarkable progress and has become a popular computer vision task, particularly in industrial environments. It plays a crucial role in assisting human workers within manufacturing units (i.e., Industry 4.0) by reducing the effort required to perceive and detect anomalies. However, the rapid growth of proposed methods has made it increasingly difficult to assess their relative strengths, limitations, and deployment readiness. To address this gap, this paper provides a comprehensive survey and deployment-aware benchmarking study of industrial anomaly detection. We systematically evaluate four recent state-of-the-art models—GLASS, DDAD, DiffusionAD, and Dinomaly—across six widely used industrial datasets (MVTec-AD, VisA, BTAD, MPDD, WFDD, and VAD), covering a broad range of defect types and acquisition conditions. Our evaluation spans detection accuracy, computational efficiency (parameters, FLOPs, latency, and memory), robustness to noisy training data, and generalization under domain shift. Quantitative cross-domain experiments reveal a pronounced reality gap: models trained on MVTec-AD experience image-level detection performance drops of 15–26% when evaluated on VisA, while pixel-level localization frequently fails altogether. Statistical significance testing and failure-mode analysis further show that no single model consistently dominates across datasets, metrics, and deployment constraints. These findings highlight the limitations of current benchmark practices and motivate a more realistic evaluation paradigm. To this end, we introduce the first deployment-aware benchmarking protocol for IAD, bridging the gap between academic performance and practical industrial requirements. Reproducibility code for all experiments is available on <a href="https://github.com/elachkarcharbel/Supervised-Anomaly-Detection-Survey">Github</a></p>