Kuzushiji characters were used in Japan hundreds of years ago, and many valuable ancient documents are written in Kuzushiji. Research into generating Kuzushiji characters increases the training data for recognizing these characters and enhances people’s interest and knowledge of Kuzushiji, which is crucial for understanding Japan’s past. However, few studies have focused on the Kuzushiji font generation task. A significant challenge in studying Kuzushiji font generation is the increasing number of deteriorated historical documents, leading to some characters being lost or illegible. In recent years, image generation using deep learning has made remarkable progress, and various methods for generating fonts have been developed with great success. However, because Kuzushiji is different from ordinary fonts, with complex structures that have larger deformations and continuous strokes, existing font generation methods do not perform well in generating Kuzushiji font. To address this problem, we propose a Kuzushiji font generation framework based on the diffusion model. Specifically, based on FontDiffuser, we introduce a stroke style module to efficiently extract font style at the stroke level, and we utilize a content-retaining model to ensure the structural integrity of the font. We conducted extensive experiments on a large number of fonts, and the results show that our model is capable of generating high-quality Kuzushiji characters, achieving state-of-the-art performance.