Cross-domain Image Generation: Style Transfer and Image-to-image Translation

With the rapid development of deep learning, cross-domain image generation has become an important research area in computer vision, with applications in artistic creation, image editing, and visual content generation. This paper reviews recent advances in two major branches: style transfer and image-to-image translation. Representative methods, including conditional GANs, CycleGAN-based models, attention mechanisms, and diffusion models, are compared in terms of principles, network architectures, and applications. Style transfer focuses on combining image content with artistic style, whereas image-to-image translation emphasizes semantic consistency across different visual domains using paired or unpaired data. Despite significant improvements in generation quality, diversity, and controllability, existing methods still face challenges in computational efficiency, training stability, robustness, generalization, and evaluation. Recent diffusion models have achieved superior fidelity and stability, indicating a shift from adversarial learning toward probabilistic generative modeling. Finally, future research directions, including lightweight architectures, multimodal integration, enhanced controllability, and more reliable evaluation frameworks, are discussed.