ESRGAN

ESRGAN, short for Enhanced Super-Resolution Generative Adversarial Networks, is a deep learning model designed for single-image super-resolution. It aims to reconstruct high-resolution images from low-resolution inputs, delivering more realistic textures and finer details than earlier GAN-based approaches. ESRGAN builds on the principles of SRGAN and is noted for improved perceptual quality.

Architecture and training approach: The generator employs residual-in-residual dense blocks (RRDB) and avoids batch normalization to reduce texture artifacts and enable deeper networks. Upsampling is typically performed with sub-pixel convolution to efficiently increase resolution. The discriminator follows a PatchGAN design, judging realism at the patch level. Training combines a perceptual loss, based on features from a pre-trained VGG network, with an adversarial loss from the discriminator. Many implementations also adopt a relativistic GAN loss (RaGAN) to stabilize training and encourage more realistic outputs.

Data, scope, and usage: ESRGAN is commonly trained on high-resolution datasets such as DIV2K and is frequently configured for a 4x (x4) upscaling, though other factors are possible. Its architectural choices—notably the RRDB blocks and removal of batch normalization—contribute to sharper textures and fewer visual artifacts compared with earlier methods.

Impact and variants: Since its introduction in 2018, ESRGAN has influenced a wide range of image super-resolution tools and research. It spurred subsequent variants and practical tools, including Real-ESRGAN, which address real-world degradations. Limitations include substantial computational requirements and the potential for texture hallucinations in some outputs.