DAA-Font: Cross-Lingual Font Generation With Dual-Branch Encoding and Multi-Level Axial Attention
Xiaoyue Yan, Dazhong Mu, Yang Zhang, Shiyuan Cao · IEEE Access · 2026
Cross-lingual font generation aims to transfer font styles across different scripts while preserving character readability. However, existing methods often suffer from inconsistent style representation and the loss of critical structural details. To address these challenges, we propose DAA-Font, a Dual-branch Axial Attention font generation network that integrates dual-branch feature encoding with multi-level axial attention. The network consists of two encoders: a style encoder equipped with a structural orientation module and a direction-guided cross-attention mechanism to capture edge-aware and directional features from English characters; and a content encoder designed to model the global and local structural patterns of Chinese glyphs through convolutional and attention-based layers. A multi-level axial attention mechanism is further applied to style features across multiple scales, enabling fine-grained modeling of structural variations along horizontal and vertical axes. In addition, a style injection module dynamically modulates the decoding process to enhance style consistency in the generated fonts. Extensive experiments on few-shot and cross-lingual font generation tasks demonstrate that DAA-Font produces structurally complete characters to high style fidelity, achieving superior results to state-of-the-art methods across quantitative and qualitative benchmarks.