Optimizing CycleGAN in the Frequency Domain for Artifact Reduction and Detail Preservation

School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen, Shenzhen 518172, China

^* Corresponding author: 122090328@link.cuhk.edu.cn

Abstract

Unsupervised image transformation models such as CycleGAN have been widely applied. However, they often generate spectral distortions that cause high-frequency artifacts and unrealistic textures. To attempt to address this critical defect, this paper proposes a frequency-aware optimization framework, introducing the spectral consistency loss into the generation process. The key point is the spectral consistency loss, which can minimize the loss of frequency information details and prevent the generation of artifacts. According to the results of experiments, the frequency-aware model reduces the FID of the generated images on the “horse2zebra” benchmark dataset by 15.3% and the KID by 30.1%. In the qualitative comparison, it can be observed that the unnatural artifacts have been significantly suppressed, and spectral analysis also indicates that its frequency distribution is closer to the real image than that of the baseline model. In addition, the human perception survey questionnaire shows that the frequency perception model has won the preference of the majority of people. This work demonstrates that the frequency-domain regularization method is a reliable and effective strategy, which can significantly improve the fidelity and reduce the loss of frequency information for unpaired image transformation, providing a promising approach to mitigating spectral inconsistency in generative models.