Generative Adversarial Networks (GANs), as an important breakthrough in the field of deep learning, have shown great potential in the automatic generation of artistic images. This article systematically studies the application of GAN in artistic creation, with a focus on analyzing the technical principles, architectural characteristics, and performance in artistic image generation of three representative models: DCGAN, CycleGAN, and StyleGAN. Through comparative experiments, this article evaluates the performance differences of different models in artistic style transfer, image generation quality, and controllability. The research results indicate that StyleGAN performs the best in generating high fidelity artistic images, with FID scores reduced by about 20% compared to the baseline model; CycleGAN has unique advantages in unpaired data style transfer tasks; As a fundamental model, DCGAN still holds value in terms of training stability and computational efficiency. This article also explores the challenges faced by GANs in artistic creation, including mode collapse, unstable training, copyright ethics, and looks forward to future development directions. This study provides theoretical basis and technical reference for the practical application of automatic generation technology for artistic images.
Huan Chen (Thu,) studied this question.