CA-Font: A Chinese Font Generation Model Based on Font Style Transfer
Zhaojun Wan, Lianqiang Niu · 2025
In the field of Chinese character generation, the structural complexity, large number, and diverse styles of Chinese characters make generating high-quality images of them a tough task. Current models can mimic various styles, but they still have problems with details, such as stroke omission, image blurriness, and low quality.To address these issues, this study proposes a Chinese character generation model based on font style transfer. The model can convert the font in a content image provided by the user to a target style based on a target style image. First, the Convolutional Block Attention Module (CBAM) was introduced in the encoder and decoder parts of the generator to enhance the model’s feature extraction ability, reduce stroke omission, and better preserve the details of the font image. Second, to solve the problems of blurry and low-quality generated Chinese characters, the Attentional Feature Fusion (AFF) module was introduced during the multi-scale feature fusion stage to effectively integrate font features at different resolutions and generate more accurate and reliable font images.This study used a dataset with 300 fonts and 6,500 characters per font to evaluate the model’s performance. The results showed that the proposed algorithm outperforms existing methods in font generation. Compared to the current best methods, it reduced the pixel-level L1 metric by 6.3%, the root mean square error by 4.3%, increased the structural similarity by 4.01%, and reduced the learned perceptual image patch similarity by 5.5%