Xi Jiang, Yue Guo, Jian Li, Yong Liu, Bin-Bin Gao, Hanqiu Deng, Jun Liu, Heng Zhao, Chengjie Wang, Feng Zheng

Researchers have developed AD-Copilot, a vision-language model specialized for industrial anomaly detection that outperforms general-purpose multimodal AI systems on manufacturing inspection tasks. The system uses a novel Comparison Encoder that directly compares paired images at the visual feature level rather than only in language space, addressing a core weakness of standard large language models on detecting subtle defects. The team created Chat-AD, a large-scale industrial dataset mined from sparsely labeled factory images, and demonstrated 82.3% accuracy on their MMAD benchmark while exceeding human expert performance on several inspection tasks.