Abstract

This study introduces a cutting-edge method for few-shot font generation, capturing the complexity and subtlety of font styles with minimal reference style images. Motivated by the time-consuming and labor-intensive process of traditional font design, particularly for languages with extensive glyph sets, such as Chinese or Korean, our approach streamlines font creation by utilizing both global and local style elements through a patch-based attention mechanism and a multi-task encoder. This innovation not only addresses the high production costs and manual effort associated with conventional font design but also overcomes the limitations of prior techniques that rely on comprehensive component definitions or multi-stage training processes. By extracting global style codes for common font family characteristics and local style codes from detailed patches, our model emphasizes critical stylistic features such as serifs, stroke shapes, and spacing. The inclusion of triplet loss and style fidelity loss further refines the model's accuracy, ensuring the generated fonts faithfully replicate the desired styles. Demonstrated through experiments on Korean and Chinese characters, our method achieves better efficiency, quality, and fidelity, outpacing current state-of-the-art solutions and highlighting the potential of an attention-based patch encoding for font generation for different languages.