Enhancing media image style transfer with advanced StyleGAN2 architectures

This paper presents a media image style transfer approach built upon an enhanced StyleGAN2 framework, designed to address common issues in traditional style transfer methods such as mode collapse, loss of image details, and style distortion. The proposed method integrates a ResNet-based generator with a PatchGAN discriminator and incorporates the DCL loss function to significantly boost the quality of generated images and stabilize the training process. Experimental evaluations on the Horse2Zebra and Cityscapes datasets demonstrate that this approach produces superior image quality. Qualitative assessments reveal that the technique effectively preserves the original content throughout the style transfer while delivering impressive stylization effects. Compared to established models like CycleGAN, CUT, and DCLGAN, our method achieves clearer images with richer color representation and successfully mitigates problems such as texture deformation and blurred details. Quantitative metrics, including Inception Score (IS) and Fréchet Inception Distance (FID), further confirm the method's advantages over existing solutions. Notably, on the Horse2Zebra dataset, the method achieves a 13% reduction in FID and a 6% increase in IS, highlighting marked improvements in both image fidelity and diversity, as well as robust generalization across datasets. Ablation studies underscore the contribution of the DCL loss function in enhancing edge detail rendering and overall image quality. Moreover, generalization tests validate that the improved StyleGAN2 model not only adapts well across varied datasets but also excels in diverse image style transfer applications.