Research map / Label-efficient and robust learning / OOD and anomaly detection
Vague Prototype-Oriented Diffusion Model for 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 VPDM, a conditional diffusion model for multi-class unsupervised anomaly detection that conditions on learned vague prototypes (low-pass features matched via optimal transport) to avoid the identical shortcut problem.
Problem
Multi-class unsupervised anomaly detection with reconstruction models suffers from the "identical shortcut" problem, where anomalous information in the condition lets the model reconstruct anomalies.
Approach
Use low-pass-filtered features matched via entropic optimal transport to a set of learnable vague prototypes as the diffusion condition, and build a diffusion process that starts from N(ŷ, I) and incrementally adds details, so anomalous detail is excluded at the input.
Key results
- On MVTec-AD multi-class setting, VPDM reaches 98.4 image AUROC and 97.8 pixel AUROC on average, above HVQ-Trans (98.0 / 97.3) and DiAD (97.2 / 96.8).
- On VisA, VPDM averages 94.2 AUROC detection, outperforming HVQ-Trans (93.2) and DiAD (86.8).
- Ablation on MVTec-AD: replacing the proposed diffusion model with DDPM drops performance from 98.2 to 76.7 AUROC; removing the vague prototype/VPOT module drops it to 68.4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use VPDM instead of HVQ-Trans | MVTec-AD multi-class anomaly detection · MVTec-AD | image-level AUROC: 98.0 -> 98.4±0.04 | Helped |
| use VPDM instead of HVQ-Trans | MVTec-AD anomaly localization · MVTec-AD | pixel-wise AUROC: 97.3 -> 97.8±0.02 | Helped |
| use VPDM instead of HVQ-Trans | VisA multi-class anomaly detection · VisA | image-level AUROC: 93.2 -> 94.2±0.09 | Helped |
| use VPDM instead of HVQ-Trans | VisA anomaly localization · VisA | pixel-wise AUROC: 98.7 -> 98.9±0.02 | Helped |
| add OT prototype matching (ablation) | Vague-only model on MVTec-AD · MVTec-AD | image-level AUROC: 68.4 -> 93.5 | Helped |
| add OT prototype matching (ablation) | Vague-only model on MVTec-AD · MVTec-AD | pixel-wise AUROC: 68.8 -> 94.1 | Helped |
| add the OT loss (ablation) | Vague + OT model on MVTec-AD · MVTec-AD | image-level AUROC: 93.5 -> 97.1 | Helped |
| add the OT loss (ablation) | Vague + OT model on MVTec-AD · MVTec-AD | pixel-wise AUROC: 94.1 -> 96.9 | Helped |
| replace VPDM's proposed diffusion process with standard N(0, I) initialization (ablation) | VPDM on MVTec-AD · MVTec-AD | image-level AUROC: 98.4 -> 76.7 | Hurt |
| replace VPDM's proposed diffusion process with standard N(0, I) initialization (ablation) | VPDM on MVTec-AD · MVTec-AD | pixel-wise AUROC: 97.8 -> 80.1 | Hurt |
| remove vague filtering (ablation) | VPDM on MVTec-AD · MVTec-AD | image-level AUROC: 98.4 -> 96.5 | Hurt |
| remove vague filtering (ablation) | VPDM on MVTec-AD · MVTec-AD | pixel-wise AUROC: 97.8 -> 96.8 | Hurt |
Surprising findings
- Replacing the proposed diffusion model with a standard DDPM causes a large drop (98.2 to 76.7), suggesting a generic diffusion model struggles to generate from a vague condition.
Limitations
- Ablation of the vague (low-pass) operation shows a non-trivial drop when removed, so gains depend on the hand-chosen filtering design.
- Evaluation limited to three industrial datasets (MVTec-AD, VisA, MPDD).
Experimental setup
Datasets: MVTec-AD VisA MPDD
Benchmarks: MVTec-AD VisA MPDD
Metrics: AUROC
Baselines: UniAD DiAD HVQ-Trans DRAEM RD4AD PaDiM MKD PSVDD US JNLD OmniAL
Builds on: DDPM Sinkhorn optimal transport DiAD HVQ-Trans
Closest papers by meaning
- A Diffusion-Based Framework for Multi-Class Anomaly Detection (AAAI 2024 · similarity 0.88)
- Correcting Deviations from Normality: A Reformulated Diffusion Model for Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.88)
- One-for-More: Continual Diffusion Model for Anomaly Detection (CVPR 2025 · similarity 0.86)
- Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.85)
- Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection (CVPR 2025 · similarity 0.85)
- On Diffusion Modeling for Anomaly Detection (ICLR 2024 · similarity 0.85)
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.