Research map / Vision-language and multimodal / CLIP and zero-shot
Defect-aware Hybrid Prompt Optimization for Zero-Shot Multi-type Anomaly Detection and Segmentation
Nadeem Nazer, Hongkuan Zhou, Lavdim Halilaj, Ylli Sadikaj, Steffen Staab
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 shared learnable context tokens to the defect-name anchor prompt (ablation) | DAPO · MPDD image-level | AUROC: 53.8 -> 81.2 | Improved |
| add the defect-name anchor to soft-only learnable prompts (ablation) | DAPO · MPDD image-level | AUROC: 73.7 -> 81.2 | Improved |
| share learnable tokens across defect types instead of per-defect tokens (ablation) | DAPO · MPDD image-level | AUROC: 70.9 -> 81.2 | Improved |
| add progressive injection of learnable tokens (ablation) | DAPO · MPDD image-level | AUROC: 79.7 -> 81.2 | Improved |
| replace linear adapters with 2-layer adapters (ablation) | DAPO · MPDD image-level | AUROC: 81.2 -> 80.5 | No clear effect |
| DAPO prompt learning versus handcrafted MultiADS prompts | DAPO · MPDD multi-type segmentation | pixel F1: 42.8 -> 50.0 | Improved |
| DAPO versus MultiADS | DAPO · MAD-Real multi-type segmentation | pixel F1: 52.5 -> 58.5 | Improved |
| DAPO versus MultiADS | DAPO · VisA multi-type segmentation | pixel AUROC: 93.6 -> 88.9 | Hurt |
| DAPO versus MultiADS | DAPO · Real-IAD binary detection | image AUROC: 78.7 -> 84.3 | Improved |
| DAPO versus MultiADS | DAPO · Internal ASIC binary detection | image AP: 61.7 -> 79.1 | Improved |
| DAPO versus MultiADS | DAPO · VisA binary segmentation | AUPRO: 89.7 -> 84 | Hurt |
| DAPO versus MultiADS on unseen stains defect | DAPO · MAD-Real per-defect segmentation | pixel F1: 6.1 -> 40.4 | Improved |
Closest papers by meaning
- AA-CLIP: Enhancing Zero-Shot Anomaly Detection via Anomaly-Aware CLIP (CVPR 2025 · similarity 0.89)
- PromptAD: Learning Prompts with only Normal Samples for Few-Shot Anomaly Detection (CVPR 2024 · similarity 0.89)
- Bayesian Prompt Flow Learning for Zero-Shot Anomaly Detection (CVPR 2025 · similarity 0.88)
- Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.88)
- MetaUAS: Universal Anomaly Segmentation with One-Prompt Meta-Learning (NeurIPS 2024 · 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.
