Anomaly-Preference Image Generation

2026-05-04

AI Summary

This work addresses the challenge of generating realistic and diverse anomalies under data scarcity, where existing generative methods often suffer from distributional misalignment and overfitting. The authors propose a novel diffusion-based paradigm that formulates anomaly generation as a preference learning problem. By leveraging real anomalies as implicit positive references and incorporating denoising trajectory deviations as optimization signals, the method achieves effective preference alignment without requiring manual annotations. To further balance structural diversity and fine-grained fidelity during the diffusion process, a time-aware capacity allocation module and a hierarchical sampling strategy are introduced. The proposed approach significantly outperforms current state-of-the-art methods in both realism and diversity metrics, establishing a new benchmark for unsupervised anomaly generation.

Abstract

Synthesizing realistic and diverse anomalous samples from limited data is vital for robust model generalization. However, existing methods struggle to reconcile fidelity and diversity, often hampered by distribution misalignment and overfitting, respectively.To mitigate this, we introduce Anomaly Preference Optimization,a novel paradigm that reformulates anomaly generation as a preference learning problem.Central to our approach is an implicit preference alignment mechanism that leverages real anomalies as positive references, deriving optimization signals directly from denoising trajectory deviations without requiring costly human annotation. Furthermore, we propose a Time-Aware Capacity Allocation module that dynamically distributes model capacity along the diffusion timeline,prioritizing structural diversity during highnoise phases while enhancing fine-grained fidelity in low-noise stages. During inference, a hierarchical sampling strategy modulates the coherencealignment trade-off, enabling precise control over generation. Extensive experiments demonstrate that significantly outperforms existing baselines,achieving state-of-the-art performance in both realism and diversity.

Problem

• anomaly generation
• fidelity-diversity trade-off
• limited data
• distribution misalignment
• overfitting

Innovation

• Anomaly Preference Optimization
• Preference Learning
• Diffusion Models
• Time-Aware Capacity Allocation
• Hierarchical Sampling