Research map / Vision-language and multimodal / CLIP and zero-shot
Kernel-Aware Graph Prompt Learning for Few-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 KAG-prompt, a few-shot anomaly detection method that builds a kernel-aware hierarchical graph over multi-layer CLIP-style/ImageBind features with message passing, plus a top-k image-level scoring fusion, achieving SOTA on MVTecAD and VisA.
Problem
Few-shot anomaly detection (FSAD) must detect unseen anomalies from very few normal support images; existing vision-language methods rely on complex text prompts and ignore cross-layer contextual relations among visual features.
Approach
Treat multi-kernel-convolved features from each encoder layer as nodes of a fully connected hierarchical graph, perform gated message passing with ConvGRU updates to capture cross-layer context, and score images by averaging the top-k pixel anomaly values fused with a global cls-based score.
Key results
- 1-shot MVTecAD image AUROC 95.8 and VisA AUROC 91.6, versus 94.6 and 91.2 for best prior (PromptAD / APRIL-GAN).
- Averaged over metrics, 1-shot average 95.2 versus 93.6 for APRIL-GAN, the next best.
- Kernel-aware hierarchical graph gives about +1.6 AUROC over the AnomalyGPT-based baseline on VisA 1-shot (89.7 to 91.3).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace PromptAD with KAG-prompt | Few-shot anomaly detection · MVTecAD 1-shot | AUROC: 94.6 -> 95.8 | Helped |
| replace APRIL-GAN with KAG-prompt | Few-shot anomaly detection · VisA 1-shot | AUROC: 91.2 -> 91.6 | Helped |
| add kernel-aware hierarchical graph (KAHG) (ablation) | KAG-prompt scoring baseline · VisA 1-shot | AUROC: 89.7 -> 91.3 | Helped |
| add kernel-aware hierarchical graph (KAHG) (ablation) | KAG-prompt scoring baseline · VisA 1-shot | pAUROC: 96.5 -> 97.0 | Helped |
| add global cls score s1 while retaining Max scoring (ablation) | KAHG model · VisA 1-shot | AUROC: 91.3 -> 91.4 | Helped |
| replace Max scoring with Top-k Fusion (ablation) | KAHG + global s1 · VisA 1-shot | AUROC: 91.4 -> 91.6 | Helped |
| replace Max scoring with Top-k Fusion (ablation) | KAHG + global s1 · VisA 1-shot | pAUROC: 97.0 -> 97.0 | No clear effect |
| add graph without multi-kernel convolution (ablation) | KAG-prompt · VisA 1-shot | AUROC: 90.5 -> 91.5 | Helped |
| add multi-kernel convolution to graph (ablation) | KAG-prompt · VisA 1-shot | pAUROC: 96.8 -> 97.0 | Helped |
Surprising findings
- APRIL-GAN's reported VisA results are said to benefit from supervised training on a test set, which the authors use to explain its strong pixel-level results.
Limitations
- Gains over the strongest baselines on some subsets and metrics are small (e.g., 1-shot VisA AUPR is below APRIL-GAN), and some subsets are worse than baselines (e.g., transistor pAUROC).
- Relies on synthetic anomalies (NSA) for training, leaving a gap from real anomalies.
Experimental setup
Datasets: MVTecAD VisA NSA synthetic anomalies
Benchmarks: MVTecAD VisA
Metrics: AUROC pAUROC AUPR PRO
Baselines: SPADE PatchCore WinCLIP APRIL-GAN AnomalyGPT PromptAD
Builds on: AnomalyGPT FiLo ConvGRU PromptAD
Closest papers by meaning
- PromptAD: Learning Prompts with only Normal Samples for Few-Shot Anomaly Detection (CVPR 2024 · similarity 0.88)
- UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · similarity 0.87)
- Bayesian Prompt Flow Learning for Zero-Shot Anomaly Detection (CVPR 2025 · similarity 0.86)
- One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · similarity 0.86)
- Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.85)
- Open-Set Graph Anomaly Detection via Normal Structure Regularisation (ICLR 2025 · similarity 0.85)
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.