Research on optimizing digital art style transfer models based on generative adversarial networks

Artificial intelligence-driven generative adversarial networks hold significant application potential in digital art image generation. An improved style transfer model based on the CycleGAN architecture is constructed, incorporating multi-scale residual modules and channel attention mechanisms to enhance feature representation. Combined with composite loss constraints, this approach achieves optimized style consistency. Experiments conducted on WikiArt and COCO datasets demonstrate improved performance compared to CycleGAN and StyleGAN2: FID reduced to 14.82, SSIM increased to 0.742, and LPIPS decreased to 0.276. The generated results exhibit superior texture fidelity and hierarchical representation, validating the model's stability and accuracy across multi-style domain tasks.