Research map / Image generation / Fast diffusion sampling
OmiAD: One-Step Adaptive Masked Diffusion Model for Multi-class Anomaly Detection via Adversarial Distillation
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 OmiAD, a one-step masked diffusion model for multi-class industrial anomaly detection that combines an adaptive masking diffusion teacher with adversarial score distillation (SiD/SiDA-based) using a shared-weight encoder discriminator.
Problem
Diffusion-based anomaly detection is slow due to iterative denoising, and reconstruction models suffer from the "identical shortcut" problem where anomalies are reconstructed along with normal content, limiting real-time industrial deployment.
Approach
An adaptive masking strategy whose masking probability grows with diffusion timestep forces the diffusion model to use broader context, and adversarial score distillation compresses it into a one-step generator, with the student's encoder reused as a discriminator to avoid extra parameters.
Key results
- On MVTec-AD multi-class, OmiAD reaches 98.8 image AU-ROC and 97.7 pixel AU-ROC, the best among compared methods.
- OmiAD inference on MVTec-AD takes 0.0254 s per batch of 64, about 200x faster than DDAD (6.0067 s) and far faster than DiAD (15.662 s).
- In ablation on MVTec-AD, adversarial score distillation with adaptive masking gives 98.8/97.7 AU-ROC, above ADD's 94.5/95.2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace DiAD with OmiAD | Multi-class anomaly detection · MVTec-AD | mAD: 84.0 -> 85.3 | Helped |
| replace RD4AD with OmiAD | Multi-class anomaly detection · VisA | mAD: 77.8 -> 79.3 | Helped |
| replace RD4AD with OmiAD | Multi-class anomaly detection · MPDD | mAD: 73.8 -> 78.9 | Helped |
| replace HVQ-Trans with OmiAD | Multi-class anomaly detection · Real-IAD | mAD: 71.4 -> 76.3 | Helped |
| use adaptive masking instead of no masking (ablation) | DDPM on MVTec-AD · MVTec-AD (image-level) | AU-ROC: 90.2 -> 98.4 | Helped |
| use adaptive masking instead of fixed masking (ablation) | DDPM on MVTec-AD · MVTec-AD (image-level) | AU-ROC: 97.7 -> 98.4 | Helped |
| use ADD instead of adaptive-mask score distillation (ablation) | Adaptive masked diffusion model on MVTec-AD · MVTec-AD (image-level) | AU-ROC: 98.4 -> 94.5 | Hurt |
| add score distillation without adversarial distillation (ablation) | Adaptive masked diffusion model on MVTec-AD · MVTec-AD (image-level) | AU-ROC: 98.4 -> 98.3 | No clear effect |
| add adversarial score distillation to score distillation (ablation) | Adaptive masked diffusion model on MVTec-AD · MVTec-AD (image-level) | AU-ROC: 98.3 -> 98.8 | Helped |
| use one-step OmiAD instead of HVQ-Trans | Anomaly detection inference, batch size 64 · MVTec-AD | inference time (seconds): 0.1795 -> 0.0254 | Helped |
Surprising findings
- The one-step distilled student with adversarial distillation reportedly surpasses its multi-step teacher (AMDM) in detection performance.
Limitations
- Evaluated only on industrial anomaly detection datasets; generality to other domains is untested.
- The discriminator design relies on channel-averaged feature maps whose justification is largely empirical.
Experimental setup
Datasets: MVTec-AD VisA MPDD Real-IAD
Benchmarks: MVTec-AD VisA MPDD Real-IAD
Metrics: AU-ROC AP F1-max AU-PRO mAD inference time
Baselines: RD4AD UniAD SimpleNet DeSTSeg DiAD HVQ-Trans ADD DDAD ReContrast TransFusion
Builds on: SiD (Score identity Distillation) SiDA Diffusion GAN DDPM DiAD
Closest papers by meaning
- On Diffusion Modeling for Anomaly Detection (ICLR 2024 · similarity 0.88)
- Dual-Interrelated Diffusion Model for Few-Shot Anomaly Image Generation (CVPR 2025 · similarity 0.87)
- AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model (AAAI 2024 · similarity 0.87)
- A Diffusion-Based Framework for Multi-Class Anomaly Detection (AAAI 2024 · similarity 0.87)
- One-for-More: Continual Diffusion Model for Anomaly Detection (CVPR 2025 · similarity 0.86)
- Correcting Deviations from Normality: A Reformulated Diffusion Model for Multi-Class Unsupervised Anomaly Detection (CVPR 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.