SCI-Font: Enhancing Content-Style Representation for Chinese Calligraphy Generation with Skeleton, Contour and Inexact Paired Data

The generation of Chinese calligraphy is a topic of great interest across various applications. Current methods typically approach it as an image-to-image translation problem. However, existing models encounter challenges in generating high-quality Chinese calligraphic characters due to the lack of effective guidance information for yielding satisfactory content-style representations from unpaired data. In this paper, we propose a novel Chinese calligraphic font generation model called SCI-Font from unpaired data, through utilizing the skeleton and contour information of Chinese characters as well as inexact paired data to enhance the content-style representation, mainly motivated by observations that the skeleton, contour and inexact paired data can provide important content information, style information and overall supervision of Chinese calligraphic characters, respectively. It is worth of pointing out that the skeleton, contour and inexact paired data are all automatically yielded by existing methods without auxiliary networks. Both skeleton and contour consistency losses are further introduced to supervise the generation procedure. The effectiveness of the proposed model is demonstrated by amounts of experiments over 16 Chinese calligraphic datasets. Experimental results show that the proposed model outperforms state-of-the-art models remarkably in terms of Frechet inception distance, and produces more realistic characters than existing models.