Since their introduction in 2014, Generative Adversarial Networks (GANs) have profoundly advanced the field of computer vision. Research on their applications primarily centers on image generation, which is underpinned by the core concept of an adversarial game between generator and discriminator networks. This dynamic process progressively refines the models ability to distinguish real from synthetic data and to generate highly realistic images, ultimately driving the entire system toward a Nash equilibrium. However, the image quality generated by early GAN models was subpar and the structure was prone to incoherent, particularly at high resolutions. Moreover, there were issues such as training instability and model collapse. These challenges, in turn, motivated and propelled further in-depth research into GANS. Grounded in the task paradigms of GANs for image generation, this review systematically summarized the developmental trajectory and ongoing challenges from 2014 to 2025 in critical domains including generator design, loss functions, training stabilization and cross-modal or multimodal synthesis. It also proposes two relatively new research directions: interactive editing (e.g., DragGAN) and adversarial distillation of diffusion models (e.g., Diffusion-GAN), thereby offering references for scholars conducting research on image generation. Furthermore, this paper finds that there are numerous metrics for evaluating generative models, such as FID and IS. How to establish a unified evaluation system and conduct qualitative and quantitative analysis of models is also one of the key research directions.
Zhijie Zhang (Thu,) studied this question.