Research map / Label-efficient and robust learning / OOD and anomaly detection
CNC: Cross-modal Normality Constraint for Unsupervised Multi-class Anomaly Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CNC, a cross-modal (CLIP text-prompt) normality constraint plus a gated mixture-of-experts module on top of reverse distillation to suppress decoder over-generalization in unsupervised multi-class anomaly detection.
Problem
Reverse-distillation anomaly detectors trained on normal samples from multiple classes let the decoder reconstruct abnormal patch features ("over-generalization"), degrading unsupervised multi-class anomaly detection on industrial images.
Approach
Class-agnostic learnable CLIP text prompts define a textual "normal" anchor that guides decoded features toward normality (plus a feature-level cross-modal control coefficient), combined with a top-K gated mixture-of-experts that routes patch patterns to specialized MLP experts to reduce inter-class interference.
Key results
- On MVTec AD (multi-class), CND reaches 98.6 I-AUROC, 98.0 P-AUROC, 93.0 AUPRO, 99.3 I-mAP, 56.4 P-mAP, with mTotal 89.0 vs. DiAD 87.3.
- On VisA, CND achieves 93.2 I-AUROC and 98.5 P-AUROC, mTotal 82.7 vs. RD4AD 82.5 and UniAD 72.7.
- Ablation on MVTec AD: CNC alone raises the base model from 95.4 to 96.7 I-AUROC; MLF+CNC+MoE gives 98.6 I-AUROC, and MoE with T=5, K=2 is best.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CND instead of DiAD | Multi-class anomaly detection · MVTec AD | mTotal: +1.7 | Helped |
| Use CND instead of RD4AD | Multi-class anomaly detection · VisA | mTotal: 82.5 -> 82.7 | Helped |
| Add CNC to the base model (ablation) | CND · MVTec AD | I-AUROC: +1.3 | Helped |
| Add CNC to the base model (ablation) | CND · MVTec AD | P-AUROC: +1.4 | Helped |
| Add CNC to the base model (ablation) | CND · MVTec AD | AUPRO: +1.7 | Helped |
| Add CNC with MLF (ablation) | CND · MVTec AD | I-AUROC: +2.1 | Helped |
| Add CNC with MLF (ablation) | CND · MVTec AD | P-AUROC: +1.4 | Helped |
| Add CNC with MLF (ablation) | CND · MVTec AD | AUPRO: +1.7 | Helped |
| Add MoE to the MLF model (ablation) | CND · MVTec AD | I-AUROC: +0.5 | Helped |
| Add MoE to the MLF model (ablation) | CND · MVTec AD | P-AUROC: +1.0 | Helped |
| Add MLF to the base model (ablation) | CND · MVTec AD | I-AUROC: +0.6 | Helped |
| Add MLF to the base model (ablation) | CND · MVTec AD | P-AUROC: +0.7 | Helped |
Surprising findings
- On VisA, the RD4AD baseline has a higher mTotal-relevant P-mAP (38.0) than CND (37.8), and RD4AD's AUPRO (91.8) exceeds CND's (91.4).
Limitations
- Performance degrades noticeably with weaker CLIP backbones (ViT-B/32, ViT-B/16) and with resolutions other than 224 or 256.
- Results on some VisA categories (e.g., pcb1 P-mAP, macaroni1) are not uniformly best; RD4AD and DiAD lead in several per-class metrics.
Experimental setup
Datasets: MVTec AD VisA
Benchmarks: MVTec AD VisA
Metrics: I-AUROC P-AUROC AUPRO I-mAP P-mAP mTotal
Baselines: RD4AD UniAD DiAD
Builds on: reverse distillation CoOp CLIP UniAD
Closest papers by meaning
- AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios (CVPR 2025 · similarity 0.88)
- ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.88)
- Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.87)
- UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · similarity 0.87)
- Unlocking the Potential of Reverse Distillation for Anomaly Detection (AAAI 2025 · similarity 0.87)
- Normal-Abnormal Guided Generalist Anomaly Detection (NeurIPS 2025 · similarity 0.86)
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.