DART: a dynamic all-directional reconstructive transformer for geometry-preserving Thangka style transfer

Thangka, a representative form of Himalayan art, poses distinctive challenges to neural style transfer because its strict geometric compositions and heterogeneous textures are prone to structural distortion and texture artifacts. We propose DART (Dynamic All-Directional Reconstructive Transformer), an end-to-end framework integrating three components: Multi-Directional Strip Attention (MDSA) to model long-range directional dependencies; a Dynamic Kernel Feed-forward Network (DKFN) to adaptively aggregate identity and multi-scale features for heterogeneous textures; and a Structure-Aware Multi-Scale Edge Loss to preserve geometric contours and fine details. Experiments on a high-quality Thangka dataset show leading overall performance among evaluated general-purpose methods. Cross-domain evaluation on MS-COCO, a blind user study, and qualitative tests on unseen Dunhuang murals and European historical paintings further demonstrate transferability, subjective preference, and generalization without fine-tuning. DART provides a practical approach for the digital preservation and creative reuse of cultural heritage imagery.