In recent years, style transfer has become increasingly prominent in various domains, especially fashion. As a tool for designers, clothing style transfer generates a wide array of styles, enabling rapid experimentation and fostering creative inspiration. However, current methods have poor performance in transferring textures and colors from style images to clothing, and commonly result in blurred boundaries between clothing and background. To address these challenges, an arbitrary style transfer algorithm for clothing is proposed, utilizing attention network and feature fusion for more effective and efficient style application. In this paper, the criss-cross attention network is incorporated to extract comprehensively global features and capture long-range dependencies, thus minimizing artifacts and enhancing texture transfer. Through a novel multi-level feature fusion approach, color transfer becomes more natural and coherent, closely aligning with the color palette of the style image. Additionally, semantic segmentation is employed to separate clothing from the background, preserving the original background and character. The experimental results show that the user preference of this paper’s algorithm is much higher than existing methods, and single 512 \(\,\times \,\) 512 image style transfer takes only 0.314 s with real-time performance.