Abstract: This paper proposes a novel multi-class unsupervised anomaly detection method, DiffAD, to address the challenges of label scarcity and insufficient accuracy in visual anomaly detection within complex industrial scenarios. The method employs a progressive feature reconstruction strategy, with its core being the design of a feature reconstruc-tion component named DADE. This component innovatively divides the reconstruction process into three stages: chaotic, pre-refinement, and strong refinement, effectively enhancing reconstruction quality and process stability. The outstanding innovation of DADE lies in the integration of differential convolutional attention and a detail enhance-ment mechanism. By combining differential convolutions with multi-head self-attention mechanisms and leveraging residual dense connections, it significantly improves the ability to capture subtle variations and high-frequency information in images, thereby enhancing the precision of anomaly localization. Extensive experiments on four repre-sentative datasets, MVTec-AD, VisA, MVTec-3D, and Uni-Medical, demonstrate that DiffAD outperforms existing mainstream methods in both image-level and pixel-level anomaly detection metrics, comprehensively demonstrating its practical application value and potential in the field of unsupervised visual detection.