Research map / Image generation / Autoregressive and latent generation
Dual-Interrelated Diffusion Model for Few-Shot Anomaly Image 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 DualAnoDiff, a dual-branch LoRA-adapted Stable Diffusion that jointly generates a full anomaly image and its anomaly-part image with shared attention plus a background compensation module, for few-shot anomaly image-mask generation on MVTec AD.
Problem
Few-shot anomaly image generation for industrial inspection suffers from limited diversity, unnatural blending of anomalies into backgrounds, and masks misaligned with generated anomalies.
Approach
Two interrelated diffusion branches (global image and anomaly part), each with its own LoRA on frozen Stable Diffusion, exchange information through a self-attention interaction module, with nested prompts and a background compensation module that injects background features via self-attention to preserve object shape; the mask comes from the anomaly-part branch segmented by SAM/U2-Net.
Key results
- Pixel-level anomaly localization on MVTec AD reaches 99.1 AUROC and 84.5 AP, versus 81.4 AP for AnomalyDiffusion (downstream U-Net trained on 1000 generated pairs).
- Generation IS average 1.93 vs 1.80 for AnomalyDiffusion and IC-LPIPS 0.38 vs 0.32 on MVTec.
- Downstream ResNet-18 classification accuracy averages 79.67% vs 66.09% for AnomalyDiffusion.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DualAnoDiff to generate anomaly data instead of AnomalyDiffusion | MVTec anomaly generation · MVTec | IS: 1.80 -> 1.93 | Helped |
| use DualAnoDiff to generate anomaly data instead of AnomalyDiffusion | MVTec anomaly generation · MVTec | IC-LPIPS: 0.32 -> 0.38 | Helped |
| train the U-Net on DualAnoDiff-generated data instead of AnomalyDiffusion-generated data | U-Net anomaly inspection · MVTec | pixel-level AP: 81.4 -> 84.5 | Helped |
| train the classifier on DualAnoDiff-generated data instead of AnomalyDiffusion-generated data | ResNet-18 anomaly classification · MVTec | classification accuracy: 66.09 -> 79.67 | Helped |
| add the Background Compensation Module (ablation) | DualAnoDiff · MVTec bottle, pill, toothbrush | pixel-level AUROC: 98.8 -> 99.1 | Helped |
| add the Background Compensation Module (ablation) | DualAnoDiff · MVTec bottle, pill, toothbrush | pixel-level AP: 83.0 -> 84.5 | Helped |
| add the Background Compensation Module (ablation) | DualAnoDiff · MVTec bottle, pill, toothbrush | pixel-level F1-max: 78.1 -> 78.8 | Helped |
| use the prompt “a vfx with sks” instead of the category and defect names directly (ablation) | DualAnoDiff · MVTec hazelnut-print | generation quality | Helped |
Limitations
- Requires training a separate model per anomaly type
- Mask quality depends on an external segmenter (SAM or U2-Net)
- Background compensation helps only some categories (bottle, pill, toothbrush shown)
Experimental setup
Datasets: MVTec AD
Benchmarks: MVTec AD
Metrics: Inception Score IC-LPIPS AUROC AP F1-max classification accuracy
Baselines: CDC Crop-Paste SDGAN DefectGAN DFMGAN AnomalyDiffusion DRAEM PRN
Builds on: Stable Diffusion AnomalyDiffusion LayerDiffusion DreamBooth LoRA
Closest papers by meaning
- AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model (AAAI 2024 · similarity 0.95)
- Correcting Deviations from Normality: A Reformulated Diffusion Model for Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.90)
- Unseen Visual Anomaly Generation (CVPR 2025 · similarity 0.88)
- DZAD: Diffusion-based Zero-shot Anomaly Detection (AAAI 2025 · similarity 0.88)
- A Diffusion-Based Framework for Multi-Class Anomaly Detection (AAAI 2024 · similarity 0.88)
- OmiAD: One-Step Adaptive Masked Diffusion Model for Multi-class Anomaly Detection via Adversarial Distillation (ICML 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.