MTDiff: Visual anomaly detection with multi-scale diffusion models

Xubin Wang, Wenju Li, Xiangjian He

Abstract: Highlights•Multi-scale diffusion models broaden the anomaly patterns coverage.•Shortened Markov chain improves efficiency and quality of reconstruction.•Equipped with a concise training and the optimized iterations in diffusion.•A simple but effective anomaly detector working at the feature level.•Robust, accurate MTDiff outperforms in diverse anomaly detection scenarios.