Efficient and Scalable Chinese Vector Font Generation via Component Composition

TL;DR

This work tackles scalable Chinese vector font generation by exploiting the compositional structure of Chinese characters. It introduces the Component Affine transformation Regressor (CAR), which regresses per-component affine transformations that are applied to component Bézier curves to synthesize vector glyphs, trained on a large-scale dataset of over $92{,}560$ Unicode characters annotated with components and six layouts NL00–NL05. A Glyphs plugin is provided to integrate the method into professional font design workflows, and a multi-loss objective (pixel, overlap, centroid, inertia) along with fusion strategies and layout-specific strategies enables high-quality generation and zero-shot extension to unseen fonts. Experiments demonstrate that CAR surpasses state-of-the-art vector font methods (e.g., DVF-v2 and GAN baselines) on complex Chinese characters and enables rapid font expansion across styles without fine-tuning. The combination of a large, richly annotated dataset, a regression-based composition model, and practical tooling promises substantial efficiency gains for real-world Chinese font design and broad character coverage.

Abstract

Chinese vector font generation is challenging due to the complex structure and huge amount of Chinese characters. Recent advances remain limited to generating a small set of characters with simple structure. In this work, we first observe that most Chinese characters can be disassembled into frequently-reused components. Therefore, we introduce the first efficient and scalable Chinese vector font generation approach via component composition, allowing generating numerous vector characters from a small set of components. To achieve this, we collect a large-scale dataset that contains over \textit{90K} Chinese characters with their components and layout information. Upon the dataset, we propose a simple yet effective framework based on spatial transformer networks (STN) and multiple losses tailored to font characteristics to learn the affine transformation of the components, which can be directly applied to the Bézier curves, resulting in Chinese characters in vector format. Our qualitative and quantitative experiments have demonstrated that our method significantly surpasses the state-of-the-art vector font generation methods in generating large-scale complex Chinese characters in both font generation and zero-shot font extension.

Figures (23)

• Figure 1: Comparison of control points between Chinese and English fonts.Left:The dark lines represent Chinese fonts, while the warm lines are English fonts. Right: Examples of control points for English character "M" (Top) and Chinese character "媒" (Bottom). The number of control points per character in the Chinese fonts is much larger than that in the English fonts, indicating that Chinese fonts are more complex than English ones.
• Figure 2: 6 main layouts for Chinese characters in our dataset. Note that NL03 has 8 variations and they are categorized as one.
• Figure 3: Our Dataset. Each sample contains ID, the character (Hanzi), Unicode, Radicals, Layout and its components.
• Figure 4: Statistical Information of our dataset.
• Figure 5: #components v.s. #chars they could assemble. Compared to traditional design pipeline, our approach is able to increase the efficiency by over 30 times.