ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks ESRGAN improves upon the SRGAN by enhancing three key components: network architecture, adversarial loss, and perceptual loss. It introduces the Residual-in-Residual Dense Block (RRDB) as the basic network unit, replaces batch normalization with residual scaling, and uses features before activation for perceptual loss. These improvements lead to better visual quality and performance in the PIRM2018-SR Challenge. The paper presents an enhanced SRGAN (ESRGAN) that achieves better visual quality than SRGAN and won the first place in the PIRM2018-SR Challenge. The model uses a deeper network with RRDB blocks, a relativistic discriminator, and improved perceptual loss. It also introduces a network interpolation strategy to balance perceptual quality and PSNR. The paper discusses the challenges of training deep networks for super-resolution, including the use of residual scaling, smaller initialization, and the impact of different datasets and training patch sizes. It also presents results on various benchmark datasets, showing that ESRGAN outperforms other methods in terms of sharpness and details. The paper also discusses the trade-off between perceptual quality and distortion, and presents results on the PIRM-SR Challenge, where ESRGAN achieved the best perceptual index. The model uses a combination of PSNR-oriented and GAN-based methods, and employs network interpolation to balance the two. The paper concludes that ESRGAN achieves consistently better perceptual quality than previous methods and won the first place in the PIRM-SR Challenge. The model uses a novel architecture with RRDB blocks, a relativistic discriminator, and improved perceptual loss. It also introduces a network interpolation strategy to balance perceptual quality and PSNR.ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks ESRGAN improves upon the SRGAN by enhancing three key components: network architecture, adversarial loss, and perceptual loss. It introduces the Residual-in-Residual Dense Block (RRDB) as the basic network unit, replaces batch normalization with residual scaling, and uses features before activation for perceptual loss. These improvements lead to better visual quality and performance in the PIRM2018-SR Challenge. The paper presents an enhanced SRGAN (ESRGAN) that achieves better visual quality than SRGAN and won the first place in the PIRM2018-SR Challenge. The model uses a deeper network with RRDB blocks, a relativistic discriminator, and improved perceptual loss. It also introduces a network interpolation strategy to balance perceptual quality and PSNR. The paper discusses the challenges of training deep networks for super-resolution, including the use of residual scaling, smaller initialization, and the impact of different datasets and training patch sizes. It also presents results on various benchmark datasets, showing that ESRGAN outperforms other methods in terms of sharpness and details. The paper also discusses the trade-off between perceptual quality and distortion, and presents results on the PIRM-SR Challenge, where ESRGAN achieved the best perceptual index. The model uses a combination of PSNR-oriented and GAN-based methods, and employs network interpolation to balance the two. The paper concludes that ESRGAN achieves consistently better perceptual quality than previous methods and won the first place in the PIRM-SR Challenge. The model uses a novel architecture with RRDB blocks, a relativistic discriminator, and improved perceptual loss. It also introduces a network interpolation strategy to balance perceptual quality and PSNR.