MFDAD: Multi-View Fusion-Guided Diffusion for anomaly detection

Abstract Single-view industrial anomaly detection often suffers from incomplete visual evidence caused by viewpoint variation and occlusion. Although existing multi-view methods exploit complementary observations, they mainly use cross-view information for feature interaction or anomaly representation, leaving its role in generative reconstruction insufficiently explored. To address this limitation, we propose MFDAD, a Multi-View Fusion-Guided Diffusion framework that explicitly separates the target restoration state from the source of conditional guidance. For each target view, the target image provides the latent state to be restored and is excluded from the conditional branch, while the remaining views are hierarchically organized into a cross-view condition that guides latent diffusion reconstruction through the U-Net cross-attention mechanism. We further introduce a timestep-coupled abnormal-to-normal training formulation in which the retained synthetic anomaly-induced latent deviation is progressively attenuated with the sampled diffusion timestep before Gaussian corruption is applied using the native latent-diffusion scheduler. The corresponding endpoint-consistent ϵ-parameterized target keeps the defect-free target latent as the desired restoration endpoint. On Real-IAD, MFDAD achieves an overall score of 80.84 ± 0.25 across three independent runs, approximately 2.7 points above MVAD, with strong detection and localization performance. Controlled ablations further verify the effectiveness of the target-excluded condition source, hierarchical multi-view condition construction, and timestep-coupled diffusion training.