One-to-More: High-Fidelity Training-Free Anomaly Generation with Attention Control

A team of researchers, including Haoxiang Rao and Fang Zhao, has introduced O2MAG (One-to-More Anomaly Generation), a novel AI method for creating realistic images of industrial defects without any model training. Published and accepted for CVPR 2026, the system addresses a critical bottleneck in manufacturing AI: the scarcity of anomalous images (like scratches or dents) needed to train robust visual inspection models. Unlike previous few-shot methods that require time-consuming fine-tuning and often produce unrealistic artifacts, O2MAG operates in a training-free manner. It leverages the powerful attention mechanisms within a pre-trained diffusion model, using just one reference image of a real anomaly and a text description to synthesize many new, high-fidelity variations.

O2MAG works by orchestrating three parallel diffusion processes and applying a technique called 'self-attention grafting.' This carefully transfers the visual patterns from the reference anomaly into the generation process of a new, normal image. To ensure precision, the method incorporates an anomaly mask to isolate the defect area and employs 'Dual-Attention Enhancement' to strengthen the model's focus on the masked region, preventing faint or blurry syntheses. Furthermore, an 'Anomaly-Guided Optimization' step aligns the semantic meaning of the text prompt with the true visual characteristics of the target defect. Extensive experiments show O2MAG outperforms prior state-of-the-art methods in generating faithful anomalies, which in turn leads to more effective anomaly detection models when used for data augmentation.