DELTA: Directional-Aware Encoding and Local Transformer for Thangka Style Transfer

Thangka is a symbolic traditional art form from the Tibetan region, playing a crucial role in preserving history and culture. However, owing to its intricate craftsmanship and vulnerability to aging, effective digital preservation and innovative dissemination have become increasingly essential. Image style transfer offers a promising solution by integrating modern visual expression with traditional Thangka aesthetics, thereby facilitating both cultural preservation and artistic innovation. Transformer-based approaches have demonstrated remarkable success in style transfer; however, two major challenges persist in their application to Thangka style transfer: (1) geometric distortions resulting from insufficient directional modeling; and (2) detail blurring and artifacts caused by over-stylization. To address these issues, we propose a novel style transfer framework, Directional-aware Encoding and Local Transformer (DELTA). DELTA introduces two key components: Directional-Aware Positional Encoding (DAPE) and Local Attention Transformer (LocalFormer). DAPE, enhanced by Pinwheel Convolution (PConv), improves the Transformer's capacity to capture fine-grained geometric structures, thereby better preserving the structural integrity of Thangka images. LocalFormer, equipped with Efficient Local Attention (ELA), enhances the Transformer's ability to fuse local features effectively, improving style consistency. Extensive experiments demonstrate that DELTA significantly outperforms state-of-the-art style transfer models such as AdaIN, AdaAttN, and StyTr<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>, achieving an average improvement of 13.5% in SSIM, along with reductions of 6.2% and 28.4% in LPIPS and content loss, respectively. Visual results show that DELTA effectively preserves the geometric structure of Thangka while consistently generating visually coherent and natural style transfer outputs.