Research map / Label-efficient and robust learning / OOD and anomaly detection
One-to-Normal: Anomaly Personalization 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 anomaly personalization for few-shot anomaly detection: a DreamBooth-customized anomaly-free Stable Diffusion model transforms each query toward the normal manifold, and a triplet contrastive inference fuses personalized, generated normal, and text-prompt scores.
Problem
Few-shot anomaly detection with pre-trained VLMs relies on direct feature comparison between a query and a handful of normal reference images, which is imprecise and unstable due to non-anomalous variation between them.
Approach
Perform a one-to-normal transformation of each query image via SDEdit-style partial noising and denoising with an object-specific anomaly-free DreamBooth model, then compare the query with its personalized counterpart, a generated anomaly-free memory bank, and CLIP text prompts, combining three scores.
Key results
- 8-shot AUROC of 96.2 on MVTec-AD and 89.9 on VisA, outperforming InCTRL (94.5 and 88.7 at 8-shot on MVTec and VisA per table).
- On RESC, the method surpasses the second-best method by 3.9, 5.9, and 4.6 AUROC points in 2-, 4-, and 8-shot settings.
- Adding generated anomaly-free samples boosts RegAD AUC by 6.8, 8.7, and 10.2 points in 2/4/8-shot settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply One-to-Normal instead of the strongest baseline | Few-shot anomaly detection · MVTec-AD, VisA, KSDD, AFID, and ELPV 8-shot | AUROC (%) | Helped |
| apply One-to-Normal instead of the strongest baseline | Few-shot anomaly detection · OCT2017, HeadCT, and RESC 8-shot | AUROC (%) | Helped |
| apply One-to-Normal instead of the strongest baseline | Few-shot anomaly detection · BrainMRI 8-shot | AUROC (%) | Helped |
| apply One-to-Normal instead of the strongest baseline | Few-shot anomaly detection · MNIST and CIFAR-10 8-shot | AUROC (%) | Helped |
| add the anomaly-free sample comparison branch (ablation) | One-to-Normal triplet inference · MVTec-AD 8-shot | AUROC (%): 91.8 -> 93.1 | Helped |
| add the personalized-image comparison branch (ablation) | One-to-Normal triplet inference · MVTec-AD 8-shot | AUROC (%): 91.8 -> 93.7 | Helped |
| combine personalized-image, anomaly-free-sample, and text-prompt branches (ablation) | One-to-Normal triplet inference · Most datasets across industrial, medical, and semantic domains | AUROC (%) | Mixed |
| add refined text prompts to one-to-normal personalization (ablation) | One-to-Normal · MVTec-AD 8-shot | AUROC (%): 95.6 -> 96.2 | Helped |
| add refined text prompts to one-to-normal personalization (ablation) | One-to-Normal · VisA, KSDD, ELPV, OCT2017, BrainMRI, HeadCT, RESC, MNIST, and CIFAR-10 8-shot | AUROC (%) | Helped |
| augment existing methods with generated anomaly-free samples (ablation) | PatchCore, RegAD, WinCLIP, and InCTRL · 2-shot, 4-shot, and 8-shot anomaly detection | AUROC (%): +6.8%, +8.7%, and +10.2% for RegAD | Helped |
| increase the number of personalization prompts from one to ten (ablation) | One-to-Normal personalization · Few-shot anomaly detection | performance | Helped |
Surprising findings
- Text-prompt-only scoring performs very poorly on medical datasets such as OCT2017 (45.3 AUROC) and RESC (39.6), indicating VLMs lack medical knowledge without reference images.
Limitations
- Requires few-shot normal reference images per category, so not applicable to zero-shot settings.
- Inference is slower than some baselines due to diffusion-based generation.
- Combining all three branches does not always give the best result (e.g., KSDD).
Experimental setup
Datasets: MVTec-AD VisA KSDD AFID ELPV OCT2017 BrainMRI HeadCT RESC MNIST CIFAR-10
Benchmarks: MVTec-AD VisA KSDD AFID ELPV OCT2017 BrainMRI HeadCT RESC MNIST CIFAR-10
Metrics: AUROC AUPRC
Baselines: PaDiM PatchCore RegAD CoOp WinCLIP InCTRL
Builds on: DreamBooth Stable Diffusion WinCLIP InCTRL PatchCore
Closest papers by meaning
- Unseen Visual Anomaly Generation (CVPR 2025 · similarity 0.90)
- Beyond Normal References: Discriminative Few-Shot Anomaly Detection (NeurIPS 2026 · similarity 0.89)
- AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model (AAAI 2024 · similarity 0.88)
- UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · similarity 0.88)
- Generating and Reweighting Dense Contrastive Patterns for Unsupervised Anomaly Detection (AAAI 2024 · similarity 0.88)
- Normal-Abnormal Guided Generalist Anomaly Detection (NeurIPS 2025 · 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.