RS-GAN: unsupervised running script font generation via disentangled representation learning and contextual transformer

• Published in: Pattern analysis and applications : PAA Vol. 28; no. 2
• Main Authors: Wang, Xuanhong, Li, Cong, Sun, Zengguo, Hui, Luying
• Format: Journal Article
• Language: English
• Published: London Springer London 01.06.2025; Springer Nature B.V
• Subjects: Calligraphy; Computer Science; Deep learning; Feature maps; Generative adversarial networks; Original Article; Pattern Recognition; Representations; Image-to-image translation; Deep learning; Running script font generation; Generative adversarial network
• ISSN: 1433-7541; 1433-755X
• DOI: 10.1007/s10044-025-01468-z

With the rapid development of deep learning, calligraphy font generation has received more and more attention. Existing methods for generating calligraphy fonts usually generate regular script, which has horizontal and vertical strokes and less variation in character shape. However, the difference between running script and regular script is huge, the regular script generation model is not suitable for generating running script fonts. To address this problem, this paper proposes a novel Generative Adversarial Network model for running script font generation, called RS-GAN, which can handle the large geometric changes between different fonts. In RS-GAN, we decoupled the input images into content space and style space to separate accurate content and style representations from the limited authentic work of calligraphers. In addition, by combining the Contextual Transformer (CoT) block with Generative Adversarial Network, the model can fully capitalize on the contextual information between the input keys. The representative ability of the output feature map is enhanced, and the ability to learn complex structural features of the running script (e.g. special simple and continuous stroke techniques) is improved. The proposed RS-GAN is verified on running script data with structural differences. These data come from the “Three Greatest Running Scripts in the World”, namely Wang Xizhi’s “Orchid Pavilion Preface”, Su Shi’s “Cold Food Observance”, and Yan Zhenqing’s “Requiem to My Nephew”. The experimental results show that the proposed RS-GAN successfully generates the running script fonts with different structures and outperforms the existing models, which verifies the effectiveness and robustness of the RS-GAN.