Preference Learning for Anomaly Synthesis: Implicit Alignment and Time-Aware Capacity Allocation
The paper proposes Anomaly Preference Optimization (APO), a paradigm shift that frames anomaly image generation as a preference learning problem. It uses real anomalies as positive references and derives optimization signals from denoising trajectory deviations, eliminating the need for costly human annotation. A Time-Aware Capacity Allocation module dynamically distributes model capacity across diffusion timesteps, prioritizing structural diversity in high-noise phases and fine-grained fidelity in low-noise phases. A hierarchical sampling strategy during inference modulates the coherence-alignment trade-off, achieving state-of-the-art performance in both realism and diversity.