Chinese Character Font Generation Based on Diffusion Model

Automatic font generation is a task of imitation, aiming to create a font library that mimics the style of reference images while preserving the content of character images. This paper introduces the Multi-Scale Feature Diffusion Font Generation (MS-Font), a novel approach that addresses the limitations of current font generation models, particularly in generating complex characters and managing significant style variations. MS-Font leverages a Multi-Scale Feature Fusion (MSF) module to blend global and local content from various scales, improving the reproduction of complex characters and enabling the generation of unseen Chinese characters with minimal samples. Additionally, a Style Insertion (SI) module is introduced for style representation learning, utilizing a font style extractor to decouple style from content and guiding the diffusion model. Extensive experiments demonstrate that MS-Font exhibits state-of-the-art performance in generating a variety of characters and styles. Compared to previous methods, it excels in handling complex characters and significant style variations.