Researchers have introduced a new method called Anomaly Preference Optimization for generating diverse and realistic anomalous image samples from limited data. This approach reformulates anomaly generation as a preference learning task, using real anomalies as positive references to derive optimization signals without human annotation. The method also incorporates a Time-Aware Capacity Allocation module to dynamically distribute model capacity across diffusion stages, balancing structural diversity and fine-grained fidelity. Experiments show this technique outperforms existing methods in both realism and diversity. AI
Summary written by gemini-2.5-flash-lite from 2 sources. How we write summaries →
IMPACT Introduces a novel approach to anomaly generation in images, potentially improving model generalization and data augmentation techniques.
RANK_REASON This is a research paper published on arXiv detailing a new method for image generation.