HierWoodNet: HierStyle multi-scale semantic style transfer method for Chinese woodblock painting

Woodblock painting, a traditional Chinese intangible cultural heritage, is known for its vivid colors and distinctive textures. However, modern lifestyles and declining cultural inheritance have led to risks of style loss and discontinuity. AI-based image style transfer offers a promising avenue for preserving and creating woodblock art. However, existing methods struggle with semantic feature extraction, style consistency, and color modeling in this domain. To address these challenges, we propose a hierarchical woodblock style transfer network. Our approach introduces a Content-Aware Positional Encoder based on the Transformer architecture, which enhances multi-scale feature representation and global structural perception. This helps preserve content during style transfer. We further present a HierStyle Encoder that learns rich, hierarchical style features from data, enabling fine-grained control over the transfer process and improving semantic and textural coherence. A progressive decoder integrates multi-stage upsampling with cross-layer connections to preserve semantic structure while injecting detailed style features. Finally, the ChromaNova Color Decoder dynamically optimizes color representation through a multi-dimensional color loss function that considers both color contrast and distribution differences. Quantitative experiments show that our method achieves the best PSNR, FID, LPIPS, and NIMA scores among the evaluated methods, supported by qualitative comparisons on multi-color woodblock style transfer. The source code and trained models will be released after acceptance at https://github.com/LPDLG/HierWoodNet .