MOPDA: Mixed-Trajectory On-Policy Distillation for Language-Guided Industrial Anomaly Detection

Name: MOPDA: Mixed-Trajectory On-Policy Distillation for Language-Guided Industrial Anomaly Detection
Uploaded: 2026-10-03T15:00:08.882365+00:00
Duration: 528 s
Description: MOPDA bridges language understanding and pixel-level anomaly detection in vision-language models for industrial settings. It uses a self-distillation framework to learn robust judgments from deployment contexts using defect evidence, then employs Language-guided Visual Anchoring to convert these judgments into precise, dense anomaly maps grounded in visual features.

MOPDA bridges language understanding and pixel-level anomaly detection in vision-language models for industrial settings. It uses a self-distillation framework to learn robust judgments from deployment contexts using defect evidence, then employs Language-guided Visual Anchoring to convert these judgments into precise, dense anomaly maps grounded in visual features.

Open the full episode