Research map / Label-efficient and robust learning / OOD and anomaly detection
CADiff: Context-Aware Diffusion for Controllable Anomaly Generation in Anomaly Detection
Xuan Tong, Yuxuan Lin, Junxiong Lin, Xinji Mai, Haoran Wang, Zeng Tao, Yang Yao, Ruofan Wang, Wenqiang Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add multi-token learning with regularization (CTP) (ablation) | CADiff · MVTec AD | pixel AUROC: 92.8 -> 98.6 | Improved |
| add multi-token learning (ablation) | CADiff · MVTec AD | pixel AUROC: 92.8 -> 97.9 | Improved |
| add regularization terms (localizing and contrastive loss) (ablation) | CADiff with MTL · MVTec AD | pixel AP: 78.1 -> 80.9 | Improved |
| add SAIM self-adaptive interaction module (ablation) | CADiff · MVTec AD | pixel AP: 80.9 -> 84.6 | Improved |
| add masked focus attention (ablation) | CADiff IAR · MVTec AD | IC-LPIPS: 0.31 -> 0.35 | Improved |
| add delayed focus scheduler (ablation) | CADiff IAR · MVTec AD | pixel AP: 82.7 -> 84.6 | Improved |
| sample intensity w uniformly in [0.5, 1.5] instead of w = 1 (ablation) | CADiff · MVTec AD | pixel AP: 81.1 -> 84.6 | Improved |
| use only w = 0.5 generated images (ablation) | CADiff · MVTec AD | pixel AP: 81.1 -> 79.9 | Hurt |
| use only amplified anomalies w = 1.5 (ablation) | CADiff · MVTec AD | pixel AP: 81.1 -> 80.2 | Hurt |
| use CADiff instead of AnoDiff to generate training data | U-Net detector · MVTec AD | pixel AP: 81.4 -> 84.6 | Improved |
| use CADiff instead of AnoDiff | U-Net detector · VisA | image AUROC: 99.7 -> 99.8 | Improved |
| use CADiff instead of AnoDiff | anomaly generation · MVTec AD | IC-LPIPS: 0.32 -> 0.35 | Improved |
Closest papers by meaning
- One-to-More: High-Fidelity Training-Free Anomaly Generation with Attention Control (CVPR 2026 · similarity 0.89)
- AnoStyler: Text-Driven Localized Anomaly Generation via Lightweight Style Transfer (AAAI 2026 · similarity 0.89)
- Unseen Visual Anomaly Generation (CVPR 2025 · similarity 0.88)
- One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · similarity 0.86)
- SetAD: Semi-Supervised Anomaly Learning in Contextual Sets (NeurIPS 2026 · similarity 0.86)
- AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model (AAAI 2024 · 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.
