CCDM-UAD: Application of Condition-Controlled Diffusion Models in Unsupervised Industrial Anomaly Detection

Anomaly detection is a critical task in industrial manufacturing, aiming to precisely localize defective regions in products. Traditional methods face challenges in acquiring sufficient training data due to annotation costs and privacy constraints. Moreover, existing approaches require category-specific model customization, leading to escalating training costs and limited cross-category generalization. To address these issues in unsupervised multi-class anomaly detection, we propose a Condition-Controlled Diffusion Model framework (CCDM-UAD) that enhances the generation process of Stable Diffusion (SD) models through advanced control mechanisms. The framework incorporates dynamic gating mechanisms to filter features input to the denoising network, effectively capturing anomaly representations to optimize reconstruction accuracy, thereby achieving unsupervised reconstruction of multi-category anomalous regions into normal patterns. Finally, multi-scale features from input and reconstructed images are compared through similarity measurement, with pixel-level anomaly localization heatmaps generated through a multi-step post-sampling mean fusion strategy for defect identification. Experimental results on MVTec-AD and VisA datasets demonstrate that our method significantly outperforms traditional approaches and existing diffusion-based anomaly detection methods.