Prototype-Guided Diffusion Model for Multi-Class Unsupervised Anomaly Detection

Abstract

Unsupervised Anomaly detection is important in industrial inspection and automation, where defects are rare, stochastic, and costly to annotate, while nominal data are abundant. Diffusion models have shown strong potential for unsupervised anomaly detection, where only normal data are available for training. However, standard class-conditional diffusion models rely on a single embedding per class, which insufficiently captures the multi-modal nature of normal distributions in unified multi-class settings. This coarse conditioning can yield over-generalised reconstructions that obscure anomalies or fail to preserve fine-grained normal structures. We propose ProtoDiffAD, a prototype-guided diffusion framework that represents each class using a dictionary of normal prototypes learned via a pretrained variational autoencoder. We further introduce Multi-Prototype Guidance (MPG), which dynamically aggregates relevant prototypes through attention during denoising. This mechanism enforces reconstruction fidelity to the normal manifold, enhancing anomaly localisation. Experiments on standard benchmarks (MVTec-AD and VisA) demonstrate consistent improvements over existing unified methods. The source code will be publicly available.