Industrial anomalies include missing parts, wrong assembly, relationships and incorrect content—not only scratches or color shifts.

Beyond differences: understand the standard

Conventional visual anomaly detection is good at finding regions that differ from normal references, but many field failures are relational. A bolt can be normal in one hole and wrong in another; an intact nameplate can still conflict with the work order. Multimodal systems combine images, written standards, reference samples and object context so the system can connect a location to a rule.

Separate localization, rules and explanation

A general-purpose model alone is not a dependable inspector. Research shows that subtle defects, pixel-level localization and unfamiliar industrial objects remain challenging. A practical architecture separates visual grounding, rule reasoning and structured output, then gives the inspector a chance to confirm the standard rather than allowing unrestricted model narration.

Validate in the field, not by a borrowed score

Every capability still needs scenario validation. Lighting, angle, resolution, defect size and normal variation change performance. What matters is whether evidence can be reviewed, conclusions map to explicit standards, and uncertain cases reach a person. That is how a multimodal model becomes an inspection process rather than an image demo.

Academic sources

AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language ModelsAAAI 2024 / arXiv
Combines large vision-language models with industrial anomaly localization and dialogue.
Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language ModelsarXiv
Studies zero-shot anomaly detection and reasoning, including limits of generalist multimodal models.