High-Resolution Locally Controllable Portrait Style Transfer
<p>Guoquan Jiang<sup>1</sup>, Huan Xu<sup>2</sup>, Zhanqiang Huo<sup>1</sup></p> · Academic Journal of Computing & Information Science · 2025
Most of the current portrait style transfer algorithms focus on the overall portrait style transfer for low-resolution images. However, in real life, users need to have fine control over specific regions of images with different resolutions, and the stylized images generated by existing methods have problems such as structure loss, local contour deformation, and color rendering errors. Therefore, this paper proposes a high-resolution locally controllable portrait style transfer model. By introducing a novel U-block, the method is not only suitable for local portrait style transfer but also can effectively handle the overall portrait style transfer task. By adopting two U-shaped encoders with different structures, this method constructs a unique generator structure, so that the structural features of the content domain and the style domain can be learned more fully, and the problem of structure loss in the stylized image is reduced. In addition, we propose a local portrait style transfer module, which allows users to perform accurate local style transfer based on the segmentation masks of different regions. To further improve the effect of local feature fusion and reduce the distortion of local contour, a Local feature fusion module (LFFM) is designed, which abandons the traditional feature splicing method and improves the quality of local stylized images by using a style attention mechanism. Finally, to reduce artifacts and color rendering errors, a local portrait style loss is introduced as a constraint to ensure that the style transfer region accurately learns the target style, while keeping the original structural features of other regions unchanged by histogram matching. The results of comparative and ablation experiments on four different style datasets show that the proposed method achieves excellent performance in global and local portrait style transfer, which verifies the effectiveness of the proposed method in portrait style transfer.