Reproduce “Hierarchical Gaussian Mixture Normalizing Flow Modeling for Unified Anomaly Detection” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, Chongyang Zhang

HGAD adapts normalizing-flow anomaly detection to unified industrial inspection by replacing a single Gaussian latent prior with a hierarchical mixture that captures differences between product classes and variation within each class. Across MVTecAD, BTAD, MVTec3D-RGB, VisA, and a 40-class Union benchmark, it reaches up to 98.4% image-level and 98.9% pixel-level AUROC, while using one model trained only on normal images.

Key insights

  • A unimodal latent prior is poorly matched to multi-class normal data; explicitly modeling several class modes reduces the homogeneous-mapping problem in flow-based detectors.
  • Separating inter-class and intra-class modeling is important: mutual-information objectives distinguish categories, while multiple centers per category represent distinct normal patterns.
  • Combining likelihood with assignment entropy improves sensitivity to features that may appear plausible under the global mixture but are unlikely for the relevant class.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.