Exposure Bias Can Alleviate Itself via Directional and Frequency Rectification in Flow Matching

2026-06-26

AI Summary

This work addresses the exposure bias between training and inference in Flow Matching, which leads to performance degradation. The authors propose DEFAR, a novel framework that, for the first time, reveals the presence of directional and frequency-adaptive signals within this bias that can be leveraged for dynamic correction. Building on this insight, DEFAR introduces a self-feedback mechanism that operates without external supervision, comprising Anti-Drift Rectification (ADR) to steer states toward their target trajectories and Frequency Compensation (FC) to reinforce missing low-frequency components. Extensive experiments demonstrate that DEFAR significantly outperforms existing baselines on CIFAR-10, CelebA-64, and ImageNet-256/512, exhibiting strong scalability, compatibility with diverse architectures, and robustness during inference.

Abstract

Flow Matching (FM) has achieved remarkable generative performance, yet it suffers from exposure bias due to discrepancies between training and inference. Existing mitigation strategies typically rely on static constraints or external heuristics. In this work, we propose that exposure bias itself inherently contains dynamic signals that can guide its own rectification. To leverage this, we introduce DEFAR (DirEctional-Frequency Adaptive Rectification). This framework simulates the single-step inference process during training to identify exposure bias. It utilizes directional and frequency-adaptive feedback signals from the bias itself to enhance the model's bias tolerance. It consists of two key components: (1) Anti-Drift Rectification (ADR). ADR treats inference-time drift as a signal to learn the direction to steer deviated states back toward the target. ADR endows the model with intrinsic active self-rectification capabilities; (2) Frequency Compensation (FC). Empirically, we observe that accumulated bias often stems from a lack of low-frequency components in high-noise stages, and exposure bias carries the missing frequency. FC leverages the bias itself as a self-feedback weighting factor to reinforce the missing frequency components. Experiments on CIFAR-10, CelebA-64, and ImageNet-256/512 show that DEFAR outperforms prior baselines and further demonstrates favorable scalability, compatibility, and inference robustness.

Problem

• Exposure Bias
• Flow Matching
• Training-Inference Discrepancy
• Generative Modeling

Innovation

• Exposure Bias
• Flow Matching
• Directional Rectification
• Frequency Compensation
• Self-Rectification