Zero-shot anomaly detection aims to spot defects in classes never seen during training, using auxiliary images and vision-language models. Existing distributional prompting methods model prompt variability but treat normal and anomalous semantics independently, which limits their ability to capture how anomalies relate to normality across different granularities.

HiDist addresses this by modeling each prompt token as a diagonal Gaussian and sampling stochastic variations via Monte Carlo. Instead of learning normal and anomalous distributions separately, it parameterizes anomalous distributions as deviations from learned normal references. The resulting distributions are then split into global and local branches, with an adapted global visual representation guiding hierarchical alignment.

Experiments across seven industrial and medical datasets show strong cross-dataset performance on both anomaly detection and localization tasks. The authors also report ablation and representation analyses that support the effectiveness of the distributional modeling and hierarchy-specific learning.