Enhanced Chinese Mural Face Generation via FreqSplitAttention and Dual Mask Discriminator

Abstract As an important part of traditional art, Chinese murals embody rich historical and cultural connotations, especially the facial images, which reflect the political, cultural, and religious aspects of ancient society. However, the erosion caused by natural environments and human factors poses significant threats to the protection and inheritance of this cultural heritage. Therefore, the application of digital technology to protect mural faces is particularly urgent. Currently, the research on mural face digital generation faces numerous challenges, including the lack of standardized datasets, inadequate detail retention in generated images, inconsistent styles, and a tendency to overfit. These issues directly impact the effectiveness of mural face digitization. To address these challenges, we propose an improved StyleGAN2 generation architecture based on FreqSplitAttention. This method leverages Fourier transforms to separate high-frequency and low-frequency components of images, achieving simultaneous attention and optimization for both the global structure and local details of mural faces, thereby significantly enhancing the quality and structural consistency of the generated images. Additionally, we introduce a dynamic masking training strategy, employing a Dual Mask Discriminator (DMD) in both spatial and spectral domains to suppress the discriminator's shortcut learning, effectively alleviating overfitting issues during GAN training on limited datasets. Furthermore, we construct a high-quality digital Dunhuang mural face dataset named DMF containing 9552 images with a resolution of 256×256. Experimental results demonstrate that our method excels in generating high-quality and diverse Chinese mural face images. On the DMF dataset, our model achieved the lowest FID score of 27.74, the lowest KID score of 10.22, and the highest IS score of 2.19, all outperforming existing state-of-the-art models. Our method not only preserves the cultural essence of traditional murals but also infuses them with new vitality, providing important theoretical and practical references for the digital preservation and cultural inheritance of murals. Additionally, the DMF dataset we constructed offers a valuable resource for the study of mural art.