In this paper, we present Explicit Dual Distribution (ExDD), a novel framework that explicitly models dual feature distributions to overcome the limitations of industrial defect detection systems that assume uniform outlier distributions and face the data shortage problem in real manufacturing environments. ExDD solves the fundamental flaw of uniform outlier assumptions by utilizing parallel memory banks that capture the unique statistical characteristics of normal and abnormal patterns. In addition, to overcome the data shortage problem, we use a domain-specific text-conditional latent diffusion model to generate synthetic faults within distributions that maintain the context of industrial environments. Finally, we efficiently fuse complementary distance measures through a neighbor-aware ratio score mechanism to amplify signals in regions that show both deviations from normal and similarities to known fault patterns. Experimental results on the KSDD2 dataset show excellent performance of 94.2% I-AUROC and 97.7% P-AUROC, achieving optimal augmentation on 100 synthetic samples.
