MA-Font: Few-Shot Font Generation by Multi-Adaptation Method

Abstract

Few-shot font generation (FFG) stands as a pivotal technique in Chinese character generation, enabling the creation of new fonts by leveraging a limited set of available font images. Despite the remarkable success of existing cross-language font generation methods, they tend to ignore some domain-specific characteristics. In addition, they cannot achieve one-to-many language conversion, that is, they cannot give the same style to texts in different languages. Therefore, this paper introduces a novel end-to-end method for Chinese character generation that aims to achieve cross-language font generation. This method incorporates three pivotal modules: the content self-adaptation module, the multi-Head attention module, and the co-adaptation module. The content self-adaptation module preserves the semantic structure of the content image by capturing spatial similarities in arbitrary positions in the content feature map. The multi-head attention module is used to capture local and global features of the style reference images. Finally, the co-adaptation module reorganizes the captured style features based on the semantic structure of the content image to generate new features. In comparative experiments, our model demonstrates superior overall performance compared to existing cross-lingual font generation methods.

Keywords

• Font generation
• self-attention
• multi-adaptation
• few-shot font generation
• style transfer
• image-to-image translation