ST3GAN: A Chinese Character Font Generation Model Combining Structural-Stroke Consistency Loss and Style Perception Loss
Chi Zhou, Kaili Wang · 2025
With the rapid development of deep learning technologies, Chinese character font generation based on artificial intelligence has become a hot research topic. However, the core challenge in Chinese character font generation lies in how to simultaneously maintain structural rationality and style consistency during the generation process. Currently, many studies focus primarily on improving the overall quality of the generated images, with less attention given to the accurate preservation of the intrinsic structural features and writing style of Chinese characters. To address this challenge, a novel Generative Adversarial Network (GAN) model, ST3GAN, is proposed. This model introduces structural stroke consistency loss and style perception loss, aiming to enhance the generator’s deep understanding and expression of Chinese character structure and style. Specifically, the structural stroke consistency loss combines structural encoding and stroke encoding to ensure the accuracy of stroke count and layout in the generated characters, while the style perception loss uses the Gram matrix to measure the similarity between the generated image and the target style, thereby enhancing style consistency. Experimental results show that ST3GAN significantly improves the quality of generated images, especially in terms of structural rationality and style consistency, demonstrating a strong advantage over existing methods in generation performance.