Artistic style transfer aims to repaint the content image with the learned artistic style. Despite notable progress achieved by convolutional neural network (CNN)-based methods, existing approaches often suffer from complex loss landscapes and limited generalization, particularly under data-scarce settings. To address these challenges, we propose a novel style transfer framework that explicitly integrates domain adaptation principles with Sharpness-Aware Minimization (SAM). Unlike previous methods, our approach aligns high-dimensional feature representations between content and style domains to preserve semantic structure, while enabling flexible stylistic transformations through controlled texture and appearance modulation in lower-dimensional spaces. Furthermore, SAM is employed to smooth the optimization landscape, leading to more stable training dynamics and improved generalization. Extensive experiments conducted on the WikiArt and DIV2k datasets show that the proposed method consistently outperforms state-of-the-art techniques, achieving a 0.0286 improvement over Cascade Style Transfer (CST) in the CLIP Score metric, while producing visually clearer and more coherent stylized results. Notably, our framework shows strong advantages in small-sample style transfer scenarios, where traditional methods typically degrade. These results indicate that the proposed method provides a robust and efficient solution for artistic style transfer with enhanced stability and generalization.
Lv et al. (Mon,) studied this question.