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

Imagine you're building a generative model that creates images or text. You want it to perform well during both training and when it's actually generating outputs. However, there's a common issue called exposure bias, where the model performs differently during training than it does during inference. This can lead to poor quality outputs because the model hasn't learned to handle the discrepancies effectively. For instance, it might generate images that look good in training but fall apart when asked to create something new in real-time. This is what's called exposure bias, and it can be a significant hurdle for generative models.

Currently, many solutions to this problem rely on static rules or external signals that don't adapt to the model's learning process. These methods can be limiting because they don't fully utilize the information available during training. They often fail when the model encounters new or unexpected data, leading to subpar performance. This is where the paper's approach comes in.

The authors propose a method called DEFAR, which stands for DirEctional-Frequency Adaptive Rectification. Instead of relying on fixed strategies, DEFAR uses the exposure bias itself as a guide for improvement. It simulates the inference process during training, allowing the model to learn how to correct itself dynamically. The method has two main components: Anti-Drift Rectification (ADR), which helps the model steer back to the target when it drifts during inference, and Frequency Compensation (FC), which addresses missing frequency components in the generated outputs. By leveraging the bias signals, DEFAR enhances the model's ability to generate high-quality outputs even in challenging scenarios.

In practical terms, this means that DEFAR can lead to better generative performance on datasets like CIFAR-10 and ImageNet, outperforming previous methods. For anyone building generative models, this approach offers a promising way to improve robustness and quality without needing extensive retraining or additional data.

The approach introduces a new way to leverage exposure bias for model improvement.

The experiments show solid results across multiple datasets, supporting the claims made.