Guanbo Huang, Jingjia Mao, Fanding Huang, Fengkai Liu, Xiangyang Luo, Yaoyuan Liang, Jiasheng Lu, Xiaoe Wang, Pei Liu, Ruiliu Fu, Shao-Lun Huang
ReflexFlow introduces a new learning objective to reduce exposure bias in Flow Matching, improving generation quality significantly across several datasets.
Flow Matching methods, used in generating images, often face a problem called exposure bias, where the model struggles with differences between training and real-world use. This paper identifies two main causes of this bias: the model's inability to handle biased inputs during training and the lack of capturing essential low-frequency information early on. To address these issues, the authors propose ReflexFlow, a new approach that adjusts the training process to better handle biases and improve image generation quality. Experiments show that ReflexFlow significantly reduces exposure bias, leading to better results on popular image datasets like CIFAR-10 and CelebA.
