Research map / Image generation / Autoregressive and latent generation
Correcting Deviations from Normality: A Reformulated Diffusion Model for Multi-Class Unsupervised 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 DeCo-Diff, a latent-space diffusion model that masks random patches during forward noising and learns to predict the direction of deviation from normality, enabling selective correction of anomalous regions for multi-class unsupervised anomaly detection.
Problem
Standard diffusion-based reconstruction for unsupervised anomaly detection degrades normal regions and struggles to selectively alter anomalous areas, especially in multi-class settings.
Approach
Reformulates the diffusion forward/reverse process so that anomalies act as noise in latent space: random masking leaves normal patches untouched, and the network predicts the deviation direction (zero for normal patches), with DDIM-style deterministic correction applied directly to the target image without a forward pass.
Key results
- On MVTec-AD multi-class, pixel-level AUPRC reaches 69.7 and f1max 74.9, improving over the second-best by 9.0 and 14.1 points (pixel-level AUPRC/f1max vs. MambaAD/GLAD).
- On VisA, image-level AUROC 96.4 vs. 94.3 for MambaAD (+2.1).
- Image-level AUROC on MVTec-AD reaches 99.3, a +0.7 gain over MambaAD's 98.6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DeCo-Diff instead of MambaAD | Multi-class anomaly detection · MVTec-AD | pixel-level AUPRC: 56.3 -> 74.9 | Helped |
| Use DeCo-Diff instead of MambaAD | Multi-class anomaly detection · VisA | pixel-level AUPRC: 39.4 -> 51.3 | Helped |
| Use DeCo-Diff instead of MoEAD | Multi-class anomaly detection · VisA | pixel-level AUROC: 98.7 -> 98.5 | Hurt |
| Use direct ˜z0 correction instead of ˜zt−1 correction (ablation) | DeCo-Diff · MVTec-AD | image-level AUROC: 99.3 -> 99.2 | Hurt |
| Use direct ˜z0 correction instead of ˜zt−1 correction (ablation) | DeCo-Diff · MVTec-AD | pixel-level AUPRC: 75.1 / 75.0 | No clear effect |
| Use 2 reverse steps instead of 5 (ablation) | DeCo-Diff with direct ˜z0 correction · MVTec-AD | image-level AUROC: 99.2 -> 99.2 | No clear effect |
| Use 2 reverse steps instead of 5 (ablation) | DeCo-Diff with ˜zt−1 correction · MVTec-AD | pixel-level AUPRC: 75.0 -> 74.9 | No clear effect |
| Use pixel- and latent-space discrepancies with geometric mean instead of pixel discrepancy only (ablation) | DeCo-Diff anomaly maps · MVTec-AD | pixel-level AUPRC: 68.0 -> 74.9 | Helped |
| Use pixel- and latent-space discrepancies with geometric mean instead of latent discrepancy only (ablation) | DeCo-Diff anomaly maps · MVTec-AD | pixel-level AUPRC: 70.5 -> 74.9 | Helped |
| Combine pixel- and latent-space discrepancies with geometric mean instead of arithmetic mean (ablation) | DeCo-Diff anomaly maps · MVTec-AD | pixel-level AUPRC: 73.5 -> 74.9 | Helped |
| Use geometric pixel/latent discrepancy instead of ResNet50 feature discrepancy (ablation) | DeCo-Diff anomaly maps · MVTec-AD | pixel-level AUPRC: 66.7 -> 74.9 | Helped |
Limitations
- Ablations reported only on MVTec-AD
- Pixel-level AUROC on VisA is second-best (98.5 vs MoEAD 98.7)
- Thresholds for latent and pixel discrepancy are hand-set hyperparameters
Experimental setup
Datasets: MVTec-AD VisA
Benchmarks: MVTec-AD VisA
Metrics: AUROC AUPRC f1max AUPRO
Baselines: RD4AD UniAD SimpleNet DeSTSeg DiAD MoEAD GLAD MambaAD
Builds on: DDPM DDIM Latent Diffusion Model DiAD
Closest papers by meaning
- A Diffusion-Based Framework for Multi-Class Anomaly Detection (AAAI 2024 · similarity 0.93)
- AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model (AAAI 2024 · similarity 0.90)
- Dual-Interrelated Diffusion Model for Few-Shot Anomaly Image Generation (CVPR 2025 · similarity 0.90)
- DZAD: Diffusion-based Zero-shot Anomaly Detection (AAAI 2025 · similarity 0.90)
- A Unified Latent Schrodinger Bridge Diffusion Model for Unsupervised Anomaly Detection and Localization (CVPR 2025 · similarity 0.89)
- Video Anomaly Detection with Motion and Appearance Guided Patch Diffusion Model (AAAI 2025 · similarity 0.88)
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.