DPNR: ADual-Prototype Noise Repository based on Prototype Learning for Anomaly Detection

ICLR 2026 Conference Submission18974 Authors

Keywords: Diffusion models Anomaly Detection Prototype Learning

Abstract: Reconstruction-based methods have demonstrated remarkable success in the domain of industrial anomaly detection.Recently, Diffusion Models have been widely applied to industrial anomaly detection, driven by their powerful reconstruction capabilities. However, existing methods predominantly rely on an idealized Gaussian noise assumption. This creates a significant discrepancy with the complex and structured characteristics of anomalies in real-world industrial settings, leading to issues such as unpredictable model behavior and high false alarm rates. To address the aforementioned challenges, we are the first to introduce the concept of prototype learning into the domain of industrial anomaly detection. We formulate the Dual-Prototype Noise Repository (DPNR), a framework designed to guide the generation of realistic, structured noise and thereby replace the simplistic Gaussian noise prior. Specifically, DPNR guides a Dual-Prototype Guided Structured Noise Injection (DP-SNI) mechanism, enabling a dynamic and content-aware noise generation process. To address the limitations of traditional loss functions, we design the Region-adaptive Hybrid Noise Loss (RHN-Loss). By leveraging dynamic blending and adaptive weighting schemes, RHN-Loss provides robust and end-to-end optimizable guidance. Extensive experiments are conducted on three commonly used anomaly detection datasets(MVTecAD, MPDD, and VisA), showing the effectiveness of the formulated method.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 18974