Cangjie: Bridging Standard and Variant Chinese Calligraphy With Unique Character Encoder

Despite generative model advances, synthesizing Chinese calligraphy with accurate glyphs, consistent styles and variant characters support remains challenging. Most existing methods treat GT as the only correct output, neglecting valid variant characters with additional or missing components. To address this limitation, we propose Cangjie, a novel framework grounded in a structure-identity coupled representation paradigm. We introduce a unique character encoder, which fundamentally resolves the mathematical ambiguity of ideographic description sequences by establishing a strict one-to-one mapping constraint, ensuring structural precision. Additionally, a calligraphy style fusion module is incorporated to integrate fine-grained stylistic features, providing better control over the artistic output. Unlike existing methods that rely on pixel-level style transfer, our approach learns the universal grammar of character construction, enabling zero-shot synthesis of variant characters. Furthermore, we propose a new evaluation framework that combines character glyph similarity and character style similarity. The former improves glyph assessment by balancing foreground and background contributions, while the latter utilizes large vision models to measure style consistency. Experimental results demonstrate that Cangjie excels in both structural accuracy and stylistic fidelity. It outperforms existing methods in generating standard and variant characters, setting a new benchmark for controllable, culturally-aware calligraphy synthesis.