Abstract
Abstract
Text-to-image (T2I) synthesis has advanced significantly with the development of deep generative models, particularly Generative Adversarial Networks (GANs). This paper presents a PRISMA-aligned systematic review of GAN-based T2I models, covering 19 models evaluated across six datasets and six evaluation metrics. The study introduces a secondary five-dimensional taxonomy that characterises models based on text encoder type, cross-modal alignment mechanism, generator–discriminator design, training stabilisation strategy, and controllability. In addition, five core GAN training objectives are formally analysed to examine their role in improving training stability and generation quality. A comparative analysis across models, datasets, and design choices reveals that encoder capacity and cross-modal alignment strategies are primary drivers of cross- dataset generalisation, while attention mechanisms and discriminator design significantly influence image quality and semantic consistency. Furthermore, a consistent trade-off between controllability and diversity is observed across models, with no existing approach achieving both simultaneously. GAN-based approaches are further contextualised through comparison with diffusion and autoregressive models, highlighting their advantages in inference efficiency and controllability despite limitations in scalability and training stability. Overall, this review provides a structured analytical perspective on the evolution of GAN-based T2I synthesis, identifies key design trade-offs, and outlines open challenges for future research.