Answers are grounded in this paper and cite their sources.
Despite rapid advances in multimodal large language models (MLLMs), effectively utilizing multimodal normal information for industrial anomaly detection remains challenging. Unlike conventional recognition tasks, anomaly detection fundamentally depends on modeling normality, with abnormal samples identified by their deviations from normal patterns. However, normal visual appearances and their textual semantics are represented in heterogeneous feature spaces, limiting consistent normality modeling by MLLMs. To address this issue, we propose a Normality-Mode-Aware Image-Text Alignment framework. The method jointly models normal reference images and normality-class textual descriptions through modality-specific branches. Normality mode predictors estimate probabilistic affiliations to different normality modes, while an explicit alignment objective promotes semantic consistency between visual and textual representations of the same class. The aligned cross-modal class information is then fused into a multimodal normality-class prior and injected into a frozen MLLM for anomaly reasoning. Experiments on the MMAD benchmark across multiple industrial datasets and open-source MLLM backbones show improved anomaly discrimination while largely preserving defect-level semantic reasoning. Further analysis across model backbones, scales, reference-set sizes, and defect categories demonstrates the effectiveness and generality of the proposed framework.
No takes yet. Share an insight, caveat, or question.
Yu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: