Research map / Image generation / Autoregressive and latent generation
DZAD: Diffusion-based Zero-shot 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 DZAD, a prompt-free zero-shot anomaly detection method that extracts multi-timestep denoising noise features from a latent diffusion model (Stable Diffusion) with a semantic-guided network, plus a two-branch multi-scale anomaly map and a learned noise weight function.
Problem
Zero-shot anomaly detection and localization on unseen classes of industrial images; existing CLIP/SAM-based methods depend heavily on text prompt quality and quantity.
Approach
Use noise (ELBO-style denoising error) features from a latent diffusion model's denoising process, aggregated over multiple timesteps and weighted by a custom function, together with an SG network that injects semantic information, to detect anomalies without any text prompts.
Key results
- Achieves 0.935 mean image-level AUROC on MVTec AD (Table 1), versus 0.911 for AnomalyCLIP and 0.694 for MAEDAY among no-prompt-free comparisons in the table.
- Image AUROC 0.902 on VisA and 0.928 on BTAD; pixel AUROC 0.867 on MVTec AD and 0.920 on VisA.
- Ablation shows the learned weight function 1+exp(-3(10-n)) improves MVTec AUROC from 0.930 (uniform weight) to 0.935.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DZAD instead of WinCLIP | Zero-shot anomaly detection · MVTec AD | Image AUC: 0.918 -> 0.935 | Helped |
| Use DZAD instead of AdaCLIP | Zero-shot anomaly detection · MVTec AD | Pixel AUC: 0.942 -> 0.867 | Hurt |
| Use DZAD instead of AdaCLIP | Zero-shot anomaly detection · VisA | Image AUC: 0.858 -> 0.902 | Helped |
| Use DZAD instead of SAA | Zero-shot anomaly detection · VisA | Pixel AUC: 0.955 -> 0.920 | Hurt |
| Use DZAD instead of AdaCLIP | Zero-shot anomaly detection · BTAD | Image AUC: 0.886 -> 0.928 | Helped |
| Use DZAD instead of AdaCLIP | Zero-shot anomaly detection · BTAD | Pixel AUC: 0.921 -> 0.923 | Helped |
| Add the SG network intermediate block to the diffusion model (ablation) | DZAD · MVTec AD | Image AUC (cls): 0.812 -> 0.834 | Helped |
| Add multi-timestep feature extraction (ablation) | DZAD · MVTec AD | Image AUC (cls): 0.834 -> 0.930 | Helped |
| Add multi-timestep feature extraction (ablation) | DZAD · MVTec AD | Pixel AUC (seg): 0.752 -> 0.823 | Helped |
| Add the two-branch feature extractor (ablation) | DZAD · MVTec AD | Pixel AUC (seg): 0.752 -> 0.864 | Helped |
| Add the two-branch feature extractor to SGSD + Mt (ablation) | DZAD · MVTec AD | Pixel AUC (seg): 0.823 -> 0.865 | Helped |
| Use weight function 1 + exp(−3 · (10 − n)) instead of SNR (ablation) | DZAD · MVTec AD | AUC: 0.718 -> 0.935 | Helped |
Surprising findings
- Classical diffusion weighting functions (SNR, exp(-6t), exp(-7t)) degrade anomaly detection, with SNR giving only 0.718 MVTec AUROC.
- Cable class is hard for all methods, which the authors attribute to heterogeneous anomaly types.
Limitations
- Pixel-level reconstruction and semantic embedding are weaker in zero-shot setting, as acknowledged by authors.
- Cable class performance is poor due to heterogeneous anomaly types.
- Some tables mix categories and metrics inconsistently, making comparisons hard to verify.
Experimental setup
Datasets: MVTec AD VisA
Benchmarks: MVTec AD VisA BTAD
Metrics: image-level AUROC pixel-level AUROC
Baselines: SAM WinCLIP SAA April-GAN AdaCLIP AnomalyCLIP MAEDAY
Builds on: LDM DiAD DDIM Variational Diffusion Model AnomalyCLIP
Closest papers by meaning
- A Diffusion-Based Framework for Multi-Class Anomaly Detection (AAAI 2024 · similarity 0.90)
- Correcting Deviations from Normality: A Reformulated Diffusion Model for Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.90)
- Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.89)
- AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model (AAAI 2024 · similarity 0.88)
- Dual-Interrelated Diffusion Model for Few-Shot Anomaly Image Generation (CVPR 2025 · similarity 0.88)
- AnomalyCLIP: Object-agnostic Prompt Learning for Zero-shot Anomaly Detection (ICLR 2024 · 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.