CalliGAN: multi-scale gradient and self-attention GAN for ancient calligraphy synthesis

Abstract

Chinese calligraphy synthesis supports digital heritage preservation, but remains challenging due to complex brushstrokes and global structural requirements. We present CalliGAN, a Generative Adversarial Network (GAN) with two variants: CalliGAN-Gen for unconditional character generation and CalliGAN-Trans for style transfer. Both variants address a unified goal—modeling calligraphy at multiple scales and maintaining structural coherence—using shared architectural components. A multi-scale gradient (MSG) generator progressively upscales features from 4 × 4 to 256 × 256, preserving fine brush details. A Spectral Normalized Self-Attention (SNSA) discriminator models long-range dependencies, ensuring structural coherence. Trained on 12,500 grayscale images of Yan Zhenqing and Mi Fu styles, CalliGAN achieves a Fréchet Inception Distance (FID) of 28.7, outperforming StyleGAN2 by 32%. A validated Calligraphy Style Score (CSS) shows 91% style fidelity. Expert evaluation confirms superior visual quality. CalliGAN provides a robust tool for creative cultural heritage recreation, with potential future extensions to restoration applications.