Despite the breakthroughs in accuracy and speed of single image super-resolution using faster and deeper convolutional neural networks, one central problem remains largely unsolved: how do we recover the finer texture details when we super-resolve at large upscaling factors? The behavior of optimization-based super-resolution methods is principally driven by the objective function. Recent work has largely focused on minimizing mean squared error (MSE), which produces "high peak signal-to-noise ratios, but they are often lacking high-frequency details and are perceptually unsatisfying." We propose SRGAN, applying a generative adversarial network (GAN) to super-resolution. We employ a perceptual loss function consisting of an adversarial loss and a content loss. Extensive mean-opinion-score (MOS) testing shows that SRGAN sets a new state of the art for photo-realistic super-resolution.

The ill-posed nature of super-resolution means that for each low-resolution image, multiple plausible high-resolution solutions exist. MSE-based optimization inherently produces the "average of all plausible solutions," which tends to be overly smooth. The key insight is that pixel-wise image differences do not align with human perception of image quality. A super-resolved image with lower PSNR can appear more realistic than one with higher PSNR. The authors propose to replace pixel-space optimization with optimization in a perceptual feature space, using a VGG-based content loss combined with an adversarial loss that encourages solutions "perceptually hard to distinguish from the HR reference images".

Early super-resolution methods using prediction-based filtering (linear, bicubic, Lanczos) are "fast but oversimplify the SISR problem and usually yield solutions with overly smooth textures." Learning-based methods progressed through example-pair approaches, sparse coding, and CNN-based methods. The seminal SRCNN by Dong et al. introduced end-to-end deep networks for SR. Prior work on perceptual loss functions by Johnson et al. demonstrated that "loss functions closer to perceptual similarity recover visually more convincing HR images" than pixel-wise MSE. Key architectural innovations include residual blocks with skip connections and learned upscaling filters shown superior to bicubic pre-upscaling.

The generator consists of 16 residual blocks with identical layout. Each block contains two convolutional layers with 3x3 kernels and 64 feature maps, batch normalization, and ParametricReLU activation. Two sub-pixel convolution layers (following Shi et al.) perform the upsampling. Skip connections from input to output relieve the network from learning the identity mapping, significantly improving training efficiency. When trained with MSE loss alone (without GAN), this architecture — termed SRResNet — already sets a new state of the art on PSNR benchmarks.

The discriminator is trained to "differentiate between the super-resolved images and original photo-realistic images". It employs eight convolutional layers with 3x3 filters, where feature maps increase from 64 to 512, doubling with each strided convolution. LeakyReLU activations (alpha=0.2) replace standard ReLU, and strided convolutions replace max pooling for spatial reduction. Two dense layers followed by sigmoid activation produce the binary classification output. This architecture follows the design principles of DCGAN.

The paper's central innovation is the perceptual loss function: l^SR = l^SR_content + 10^-3 * l^SR_adversarial. The content loss measures distance in VGG19 feature space rather than pixel space: the Euclidean distance between feature representations phi_{i,j}(I^HR) and phi_{i,j}(G(I^LR)) from a specific VGG layer. This is "more invariant to changes in pixel space" and captures high-level semantic content. The adversarial loss -log D(G(I^LR)) pushes super-resolved images toward the natural image manifold. The 10^-3 weighting on adversarial loss prevents competition with content loss, balancing texture hallucination against content fidelity.

On Set14 at 4x upscaling, SRGAN achieves MOS of 3.72 compared to SRResNet's 2.98 and DRCN's 2.84, while original HR images score 4.32. Notably, SRGAN's PSNR (26.02 dB) is actually lower than SRResNet (28.49 dB), confirming the paper's thesis that PSNR and perceptual quality are misaligned. Testing with 26 raters and over 29,328 ratings provides statistical rigor. VGG feature loss from deeper layers (phi_{5,4}) produces better texture detail than shallower layers (phi_{2,2}). The authors acknowledge limitations: approaches that "hallucinate finer detail might be less suited for medical applications or surveillance" and text/structured scenes remain challenging.

The authors conclude with a dual contribution: SRResNet "sets a new state of the art on public benchmark datasets when evaluated with the widely used PSNR measure," while SRGAN "augments the content loss function with an adversarial loss by training a GAN." Using extensive MOS testing, they confirm that SRGAN reconstructions for 4x upscaling are, "by a considerable margin, more photo-realistic than reconstructions obtained with state-of-the-art reference methods." The paradigm shift from pixel fidelity to perceptual fidelity fundamentally redefines the objective of super-resolution research.