A Novel Arbitrary Style Transfer Algorithm via Multi-Order Attention
Jiaqi Chang, Dong Sun, Qingwei Gao, Yixiang Lu, Muxi Bao, Yongli Hu · 2025
Arbitrary style transfer aims to transfer style images to content images. Existing solutions either directly fuse the style features with the content features or adaptively normalize the content features according to the style features to match the global statistics of the two. However, little attention is paid to the positional relationship of the content information, and the distributional relationship of the style features is not taken into account, so the results often suffer from local distortion and style detail errors. To improve this phenomenon, we propose a new arbitrary style transfer algorithm to solve the existing problem and obtain content and style balanced results, which consists of a Multi-Order Attention, and a Merge module. This algorithm is named SMA (Style Transfer Algorithm via Multi-Order Attention). SMA obtains the content encoding and inputs it together with the style encoding into the MoA module to obtain the weighted distribution of the style encoding as well as the overall distribution, and then adaptively normalizes and fuses the content features through the distribution relationship; then, the Merge module fuses the features at multiple levels. In addition, a new multi-order style loss function is derived based on MoA, which can enhance the learning of style details. Finally, experiments demonstrate that our method achieves good results on several metrics. Our algorithm can balance content and style well, making the results more appealing and exploring new angles of style transfer algorithms.