Abstract Modeling structured artistic images with complex patterns remains a challenge in style transfer. Thangka art, featuring intricate iconography and long-range spatial composition, serves as an ideal benchmark. We propose Dynamic Adaptive Convolution and Global Dependency Attention (DAG-Style Attention), an end-to-end framework. It features a Dynamic Adaptive Convolution (DAConv) module that dynamically allocates multi-scale kernels to extract features from complex, irregularly sized objects. Concurrently, the Global Dependency Perceptual Attention (GDPAttention) module models long-range semantic correlations to prevent structural distortion and ensure stylistic consistency. Experiments on Thangka and WikiArt datasets demonstrate our framework’s superiority. On 512 × 512 Thangka images, it achieves 14.186 dB PSNR and 0.517 SSIM, outperforming the strongest baseline (ASFNet) while exhibiting superior visual fidelity. Ultimately, this approach provides a robust tool for cultural heritage preservation via precise digital reconstructions.
Yang et al. (Fri,) studied this question.