Generative Adversarial Networks (GANs) still encounter hurdles such as ensuring consistent training, avoiding mode collapse, and generating high- quality images. In reaction to this, the article presents a novel GAN model called Realism- Enhanced Progressive GAN (REP-GAN). REP-GAN tackles these problems by incorporating a few important improvements: Layered Progressive Growth to incrementally increase model complexity, Dynamic Instance Normalization (DIN) to adjust feature maps as needed during training, and Perceptual Realism Loss (PRL) to evaluate image realism based on human perception rather than pixel-level comparisons. The enhancements increase the stability of training and improve the quality and diversity of images produced by REP-GAN. The article details the organization of REP-GAN, highlighting its algorithms and mathematical equations, showing how these components work together to outperform existing GAN models. Thorough testing and analysis have confirmed that REP-GAN outperforms other advanced methods in producing a wider range of realistic images, indicating a major advancement in GAN-based image generation technology.