Generating and Evaluating Cursive Chinese Calligraphy by Semi-Classifying Style: A Case Study Using a Diffusion Model

Generative AI offers a promising approach to overcoming the challenges of optical character recognition (OCR) for traditional Chinese cursive calligraphy. By generating text-image data, we can address data insufficiency and imbalance, enhancing the effectiveness of deep learning models in capturing the variability of cursive styles. In this study, we developed a text-to-text-content-image generation model utilizing the Cursive Chinese Calligraphy Dataset and the WordStylist framework. By categorizing images according to style characteristics, our method reduces the learning curve and improves the quality of generated artifacts while tackling common issues in handwritten datasets, such as the absence of author annotations and mixed styles without proper labeling.We trained three models with varying style categories: four-class, eight-class, and sixteen-class, and introduced a semi-annotating method for style categorization. Experimental results indicate that fewer style categories lead to greater variation in cursive characters within each class. The sixteen-class model exhibited more consistent styles, resulting in faster convergence and more stable generation, thereby offering users a broader range of style choices.To evaluate the quality of the generated content, we employed OCR-Embedding with Maximum Mean Discrepancy(MMD), derived from CMMD, as well as expert evaluations. The results from OCR-Embedding MMD were generally consistent with those of expert assessments. According to the experts, the four-class model performed best in zero-shot sample generation. We infer that zero-shot characters benefit from features of other characters within the same style class, as these characters have more training data available, contributing to improved generation of zero-shot characters.