Research map / Label-efficient and robust learning / OOD and anomaly detection
AnomalyMoE: Towards a Language-free Generalist Model for Unified Visual Anomaly Detection
Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Wei Ge, Ming Tang, Jinqiao Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add component-level and global-level experts to patch-level experts (ablation) | AnomalyMoE · MVTec LOCO | image-level AUC (%): 78.2 -> 87.5 | Improved |
| add component-level and global-level experts to patch-level experts (ablation) | AnomalyMoE · MVTec-AD | image-level AUC (%): 98.2 -> 99.5 | Improved |
| add the third expert group to a two-group configuration (ablation) | AnomalyMoE · MVTec LOCO | image-level AUC (%): 86.8 -> 87.5 | Improved |
| add Expert Selection Balancing (ESB) loss (ablation) | AnomalyMoE · MVTec LOCO | image-level AUC (%): 84.2 -> 87.5 | Improved |
| add Expert Information Repulsion (EIR) loss (ablation) | AnomalyMoE · MVTec LOCO | image-level AUC (%): 85.6 -> 87.5 | Improved |
| add ESB loss (ablation) | AnomalyMoE · MVTec-AD | image-level AUC (%): 99.2 -> 99.5 | Improved |
| add EIR loss (ablation) | AnomalyMoE · MVTec-AD | image-level AUC (%): 99.0 -> 99.5 | Improved |
| add ESB loss (ablation) | AnomalyMoE · MVTec LOCO | pixel-level AUC (%): 81.1 -> 82.2 | No clear effect |
| add EIR loss (ablation) | AnomalyMoE · MVTec LOCO | pixel-level AUC (%): 92.0 -> 82.2 | Hurt |
| use language-free experts instead of vision-language models (UniVAD) | AnomalyMoE · Inference speed | inference time (ms): 658.5 -> 58.5 | Improved |
| use AnomalyMoE instead of specialist reconstruction method Dinomaly | MVTec LOCO · MVTec LOCO | image-level AUC (%): 78.2 -> 87.5 | Improved |
| use AnomalyMoE instead of ReContrast | RESC · RESC | image-level AUC (%): 94.6 -> 92.2 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 3 on most shared results.
- Dinomaly — Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection (CVPR 2025) · better on 4 of 5 shared results
- UniVAD — UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025) · better on 3 of 3 shared results
- ReContrast — ReContrast: Domain-Specific Anomaly Detection via Contrastive Reconstruction (NeurIPS 2023) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.88)
- UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · similarity 0.88)
- LogicAD: Explainable Anomaly Detection via VLM-based Text Feature Extraction (AAAI 2025 · similarity 0.88)
- Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.87)
- Unseen Visual Anomaly Generation (CVPR 2025 · similarity 0.87)
- Towards Training-free Anomaly Detection with Vision and Language Foundation Models (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
