This paper presents SRGAN, a generative adversarial network (GAN) for image super-resolution (SR). The main contribution is the introduction of a perceptual loss function that combines an adversarial loss and a content loss. The adversarial loss encourages the model to generate images that are indistinguishable from real images, while the content loss ensures that the generated images retain the structural and perceptual characteristics of the original. The content loss is based on high-level feature maps from the VGG network, which are more invariant to changes in pixel space. The proposed method achieves significant improvements in perceptual quality compared to state-of-the-art methods, as demonstrated by extensive mean opinion score (MOS) tests on public benchmark datasets. The results show that SRGAN outperforms other methods in terms of both PSNR and SSIM, and the generated images are closer to the original high-resolution images in terms of perceptual quality. The method is particularly effective for 4× upscaling factors, where traditional methods often fail to capture fine texture details. The paper also discusses the design of the network architecture, the choice of loss functions, and the training process. The results demonstrate that the proposed method provides a more realistic and visually pleasing image reconstruction compared to traditional methods.This paper presents SRGAN, a generative adversarial network (GAN) for image super-resolution (SR). The main contribution is the introduction of a perceptual loss function that combines an adversarial loss and a content loss. The adversarial loss encourages the model to generate images that are indistinguishable from real images, while the content loss ensures that the generated images retain the structural and perceptual characteristics of the original. The content loss is based on high-level feature maps from the VGG network, which are more invariant to changes in pixel space. The proposed method achieves significant improvements in perceptual quality compared to state-of-the-art methods, as demonstrated by extensive mean opinion score (MOS) tests on public benchmark datasets. The results show that SRGAN outperforms other methods in terms of both PSNR and SSIM, and the generated images are closer to the original high-resolution images in terms of perceptual quality. The method is particularly effective for 4× upscaling factors, where traditional methods often fail to capture fine texture details. The paper also discusses the design of the network architecture, the choice of loss functions, and the training process. The results demonstrate that the proposed method provides a more realistic and visually pleasing image reconstruction compared to traditional methods.