OPD-IAD: From Language Judgment to Industrial Anomaly Detection via On-Policy Self-Distillation

OPD-IAD introduces an evidence-privileged dense on-policy self-distillation framework for industrial anomaly detection with large vision-language models. It distills privileged defect evidence into the model’s own judgment trajectory, so the generated judgment receives dense supervision instead of being treated only as a textual answer. Its Language-guided Visual Anchoring module re-encodes the image and question under the final judgment, forms semantic anchors, and contrasts them with dense visual features to produce anomaly heatmaps. Language supplies semantic guidance, while dense visual features remain the basis for pixel-level scoring. The abstract reports leading overall performance among LVLM-based IAD methods.