High-Fidelity Image Style Transfer by Hybrid Transformers1
Zhe-Wei Hsu, Shih‐Hsuan Yang, Bo-Jiun Tung · 2024
Image style transfer is a technique in computer vision by which the artistic style of one image is applied to the content of another while keeping the structural features. Image style transfer finds applications in creating artwork, design and branding, entertainment and media, and many other fields. Current image style transfer methods fail to satisfactorily retain global characteristics and local details simultaneously. This paper proposes a hybrid transformer architecture that incorporates mixed convolutional network modules. By integrating the transformer and the convolutional modules, the global and local features are captured and fused. Experimental results demonstrate that the proposed method achieves more favorable visual fidelity, reducing the combined content and style loss by at least 10% as compared with the state-of-the-art approaches.