Research map / Image generation / Autoregressive and latent generation
Unseen Visual Anomaly Generation
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 AnomalyAny, a training-free framework that uses pretrained Stable Diffusion with a single normal image, attention-guided latent optimization, and GPT-4-derived detailed prompts to generate unseen visual anomalies for industrial AD.
Problem
Visual anomaly detection suffers from scarce anomalous training samples; existing anomaly synthesis methods either lack realism or require training on normal/anomalous data, and fail to generate unseen anomaly types.
Approach
Instead of fine-tuning Stable Diffusion, condition generation on a single test-time normal image (starting denoising from a noised normal latent), then optimize latents to maximize cross-attention on anomaly tokens and refine with CLIP loss on GPT-4-written detailed anomaly descriptions.
Key results
- In 1-shot anomaly detection on MVTec AD, AnomalyAny reaches 94.9 I-AUC and 95.4 P-AUC, outperforming PromptAD (94.6) and WinCLIP+ (93.1) and other generation baselines.
- On VisA 1-shot, AnomalyAny achieves 89.7 I-AUC versus 87.4 for AnomalyGPT and 86.9 for PromptAD.
- Generation quality averages IS 2.02 and IC-LPIPS 0.33, the highest IS and diversity among compared generators (NSA, RealNet, AnomalyDiffusion).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AnomalyAny-generated anomalous samples for 1-shot anomaly detection | Anomaly detection models · MVTec AD | I-AUC: 94.6±1.7 -> 94.9±0.4 | Helped |
| Use AnomalyAny-generated anomalous samples for 1-shot anomaly detection | Anomaly detection models · VisA | I-AUC: 87.4±0.8 -> 89.7±0.8 | Helped |
| Use AnomalyAny rather than the strongest alternative generation method | Anomaly generation · MVTec AD average | IS: 1.80 -> 2.02 | Helped |
| Add test-time normal sample conditioning (ablation) | AnomalyAny · Anomaly generation | generation quality | Helped |
| Add attention-guided anomaly optimization (ablation) | AnomalyAny · Anomaly generation | anomaly-description fidelity | Helped |
| Add prompt-guided anomaly refinement (ablation) | AnomalyAny · Anomaly generation | generation realism | Helped |
| Add localization-aware scheduling for attention updates (ablation) | AnomalyAny · Anomaly generation | artifact reduction | Helped |
| Add detailed prompt description c′ alongside anomaly type c (ablation) | AnomalyAny · Anomaly generation | generation realism | Helped |
| Add attention loss Latt (ablation) | AnomalyAny prompt-guided optimization · Anomaly generation | anomaly-region focus | Helped |
| Add prompt-based loss terms (ablation) | AnomalyAny prompt-guided optimization · Anomaly generation | generation quality | Helped |
| Add prompt guidance when anomaly type is unambiguous (ablation) | AnomalyAny · Anomaly generation | generation quality | No clear effect |
Surprising findings
- A training-free generator with no anomalous training data matches AnomalyDiffusion, which was trained on test-set anomalies, on detection performance.
Limitations
- Relies on text prompts, which CLIP may not interpret well for long or specialized descriptions
- Precision for complex anomalies is limited without an anomalous image input
- Comparisons of detection use a 1-shot setting with 100 generated samples per category
Experimental setup
Datasets: MVTec AD VisA web images
Benchmarks: MVTec AD VisA
Metrics: Inception Score IC-LPIPS I-AUC I-F1 P-AUC P-F1 PRO
Baselines: DRAEM NSA RealNet AnomalyDiffusion PaDiM PatchCore WinCLIP+ AnomalyGPT PromptAD
Builds on: Stable Diffusion AnomalyDiffusion CLIP DDPM
Closest papers by meaning
- AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model (AAAI 2024 · similarity 0.90)
- One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · similarity 0.90)
- Dual-Interrelated Diffusion Model for Few-Shot Anomaly Image Generation (CVPR 2025 · similarity 0.88)
- Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.88)
- Generating and Reweighting Dense Contrastive Patterns for Unsupervised Anomaly Detection (AAAI 2024 · similarity 0.88)
- Odd-One-Out: Anomaly Detection by Comparing with Neighbors (CVPR 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.