DRA-font: Cross-Lingual Font Synthesis by Diffusion-Regularized Adversarial Learning with an Enhanced Attention Module
Xiangrong WU, Dazhong Mu, Xinxin GENG, Baosheng CAO, Xitao XU · IEICE Transactions on Information and Systems · 2026
Few-shot cross-lingual font style transfer aims to accurately synthesize target-language glyphs across different languages. The synthesized glyphs are required to preserve strict character structure and legibility, while inheriting style from a few reference glyphs in another script. This capability is important for alleviating the scarcity of low-resource fonts, enabling multilingual logo design, and facilitating cultural communication. Most existing approaches are GAN-based, but traditional GAN models are often unstable and provide limited capacity to represent fine-grained styles under large cross-script structural discrepancies. To address these challenges, we propose DRA-font, a diffusion-regularized adversarial framework that stabilizes training by integrating a forward diffusion process with time-conditioned discriminators. We further introduce an adaptive diffusion-step scheduling strategy to balance structural fidelity and style expressiveness. Additionally, to improve transferable style modeling, an enhanced attention module is employed, combining multi-head self-attention, local-global attention, and cross-scale attention to capture global dependencies, stroke-level details, and multi-resolution features. Experiments demonstrate that compared with existing cross-lingual font synthesis models, DRA-font achieves superior performance in both style preservation accuracy and structural consistency of generated characters.