Super-Resolution GAN (SRGAN)

• SRGAN is a generative adversarial network that reconstructs high-resolution images from low-resolution inputs using a combination of content and adversarial losses.
• It features an SRResNet-based generator with sub-pixel convolution upsampling and a VGG-style discriminator to enhance perceptual quality.
• The framework shifts focus from traditional pixel-wise metrics to human perceptual quality, influencing numerous variants in both natural and domain-specific image super-resolution.

Searching arXiv for SRGAN and closely related follow-up work to ground the article in published sources. arxiv_search(query="SRGAN Ledig super-resolution generative adversarial network ESRGAN RankSRGAN", max_results=10) arxiv_search(query="Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network", max_results=5) Super-Resolution Generative Adversarial Network (SRGAN) is a generative adversarial network for single-image super-resolution in which a generator reconstructs a high-resolution image from a low-resolution observation and a discriminator is trained to differentiate super-resolved images from original photo-realistic images. Introduced by Ledig et al., SRGAN replaced plain pixel-wise optimization with a perceptual loss composed of a content term and an adversarial term, and was presented as the first framework capable of inferring photo-realistic natural images for $4\times$ upscaling factors (Ledig et al., 2016). Its historical importance lies less in maximizing distortion-oriented metrics than in formalizing perceptual super-resolution as adversarial learning on the natural image manifold (Ledig et al., 2016).

1. Emergence and problem definition

Single-image super-resolution is highly under-determined, especially at large upscaling factors such as $4\times$. Prior CNN-based methods had largely optimized pixel-wise mean squared reconstruction error, equivalently maximizing PSNR. In the SRGAN formulation, this optimization regime was identified as the principal cause of overly smooth “average” reconstructions that lack sharp textures and fine details, even when PSNR and SSIM are high (Ledig et al., 2016).

SRGAN reframed the objective of super-resolution. Rather than treating the task only as pixel recovery, it coupled reconstruction with adversarial discrimination so that generated outputs would be pushed toward the manifold of natural images. The central methodological claim was that human perceptual studies, particularly mean opinion score (MOS) tests, do not align well with MSE-optimized reconstructions, and that adversarial training can recover perceptually plausible high-frequency detail that distortion-minimizing networks suppress (Ledig et al., 2016).

This shift established a durable distinction between distortion-oriented super-resolution and perceptual super-resolution. A plausible implication is that SRGAN should be understood not simply as a particular network topology, but as a change in the optimization target of super-resolution itself.

2. Canonical SRGAN architecture

The original SRGAN generator uses an SRResNet backbone. Its input is $I^{LR}\in\mathbb R^{W\times H\times C}$ and its output is $I^{SR}\in\mathbb R^{rW\times rH\times C}$. The first layer is $\mathrm{Conv}(3\times3,64)+\mathrm{PReLU}$, followed by $B=16$ residual blocks. Each residual block contains $$\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}+\mathrm{PReLU}$$ and then $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$, with a skip connection from block input to block output. After the residual stack, a $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$ layer is combined with a long skip connection from the first convolution. Upsampling is performed by two sub-pixel convolution layers, each of the form $$\mathrm{Conv}(3\times3,256)\rightarrow\mathrm{PixelShuffle}\times2\rightarrow\mathrm{PReLU}$$, giving an overall $4\times$0 scale factor. A final $4\times$1 maps features back to RGB, and the network is fully convolutional, so it can process arbitrary image sizes (Ledig et al., 2016).

The discriminator is a VGG-style classifier with no max-pooling, using strided $4\times$2 convolutions for downsampling. Feature-map sizes double whenever spatial resolution halves, progressing as $4\times$3 (stride $4\times$4), $4\times$5 (stride $4\times$6), $4\times$7 (stride $4\times$8), $4\times$9 (stride $I^{LR}\in\mathbb R^{W\times H\times C}$0), and $I^{LR}\in\mathbb R^{W\times H\times C}$1 (stride $I^{LR}\in\mathbb R^{W\times H\times C}$2). All convolutional layers use LeakyReLU with $I^{LR}\in\mathbb R^{W\times H\times C}$3. After the final convolutional block, two fully connected layers $I^{LR}\in\mathbb R^{W\times H\times C}$4 and a sigmoid output the probability that an input is a real high-resolution image (Ledig et al., 2016).

This generator–discriminator pairing became the reference template for a large number of later systems. Even when later work replaced residual blocks, changed the discriminator, or altered the loss, the SRGAN decomposition into a super-resolution generator, an adversarial critic, and a perceptual objective remained intact.

3. Perceptual loss, optimization, and training

SRGAN’s defining objective is the perceptual loss

$I^{LR}\in\mathbb R^{W\times H\times C}$5

The adversarial term follows the standard minimax GAN objective,

$I^{LR}\in\mathbb R^{W\times H\times C}$6

while in practice the generator minimizes the heuristic form

$I^{LR}\in\mathbb R^{W\times H\times C}$7

Two content losses were studied: a pixelwise MSE term and a VGG19 feature loss based on activations $I^{LR}\in\mathbb R^{W\times H\times C}$8 before the $I^{LR}\in\mathbb R^{W\times H\times C}$9-th max-pool and after the $I^{SR}\in\mathbb R^{rW\times rH\times C}$0-th convolution. The best reported SRGAN results used the deep VGG content loss with $I^{SR}\in\mathbb R^{rW\times rH\times C}$1 (Ledig et al., 2016).

Training proceeded in two stages. SRResNet with MSE only was trained for $I^{SR}\in\mathbb R^{rW\times rH\times C}$2 updates at learning rate $I^{SR}\in\mathbb R^{rW\times rH\times C}$3. SRGAN was initialized from SRResNet-MSE weights and then trained for $I^{SR}\in\mathbb R^{rW\times rH\times C}$4 updates at $I^{SR}\in\mathbb R^{rW\times rH\times C}$5 and another $I^{SR}\in\mathbb R^{rW\times rH\times C}$6 updates at $I^{SR}\in\mathbb R^{rW\times rH\times C}$7, alternating generator and discriminator updates with $I^{SR}\in\mathbb R^{rW\times rH\times C}$8. The training set comprised $I^{SR}\in\mathbb R^{rW\times rH\times C}$9 images from ImageNet. Low-resolution inputs were generated by Gaussian blur and bicubic downsampling with factor $\mathrm{Conv}(3\times3,64)+\mathrm{PReLU}$0. Each update used $\mathrm{Conv}(3\times3,64)+\mathrm{PReLU}$1 random $\mathrm{Conv}(3\times3,64)+\mathrm{PReLU}$2 high-resolution patches; low-resolution inputs were scaled to $\mathrm{Conv}(3\times3,64)+\mathrm{PReLU}$3, high-resolution images to $\mathrm{Conv}(3\times3,64)+\mathrm{PReLU}$4, and VGG feature maps were rescaled by $\mathrm{Conv}(3\times3,64)+\mathrm{PReLU}$5 so that losses were comparable in magnitude. At test time, BatchNorm statistics were fixed so outputs were deterministic (Ledig et al., 2016).

The resulting optimization problem is explicitly non-equivalent to PSNR maximization. This is not an incidental implementation choice but the defining principle of SRGAN.

4. Evaluation and the PSNR–MOS divergence

The original evaluation protocol used Set5, Set14, and BSD100 at $\mathrm{Conv}(3\times3,64)+\mathrm{PReLU}$6 upscaling, with PSNR and SSIM measured on the Y channel after cropping a $\mathrm{Conv}(3\times3,64)+\mathrm{PReLU}$7-pixel border. Perceptual evaluation was performed by MOS over $\mathrm{Conv}(3\times3,64)+\mathrm{PReLU}$8 human raters with scores from $\mathrm{Conv}(3\times3,64)+\mathrm{PReLU}$9 to $B=16$0 (Ledig et al., 2016).

The reported results formalized a recurring pattern in perceptual super-resolution. On Set5, Set14, and BSD100, SRResNet (MSE) obtained $B=16$1 dB with SSIM $B=16$2, whereas SRGAN (VGG5.4+GAN) obtained $B=16$3 dB with SSIM $B=16$4 (Ledig et al., 2016). On BSD100, however, MOS was $B=16$5 for bicubic, approximately $B=16$6 for SRCNN and SelfExSR, approximately $B=16$7 for DRCN and ESPCN, $B=16$8 for SRResNet (MSE), $B=16$9 for SRGAN (VGG5.4+GAN), and $$\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}+\mathrm{PReLU}$$0 for original HR images (Ledig et al., 2016).

A common misconception is that the lower PSNR of adversarial super-resolution indicates an inferior reconstruction model. The data associated with SRGAN contradict that interpretation. The original paper explicitly argued that higher PSNR and SSIM do not imply better perceptual quality, and that adversarial plus VGG losses yield lower PSNR but much higher MOS (Ledig et al., 2016). The same divergence reappeared in prostate MRI: at $$\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}+\mathrm{PReLU}$$1 upscaling, SRCNN achieved PSNR $$\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}+\mathrm{PReLU}$$2 dB and SSIM $$\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}+\mathrm{PReLU}$$3, whereas SRGAN achieved PSNR $$\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}+\mathrm{PReLU}$$4 dB and SSIM $$\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}+\mathrm{PReLU}$$5, yet SRGAN received the highest MOS of $$\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}+\mathrm{PReLU}$$6 and was reported as visually the most similar to the original HR images (Sood et al., 2019).

This pattern suggests that SRGAN should be evaluated in a multi-objective regime. Distortion metrics, perceptual ratings, and downstream task performance quantify different properties of the reconstruction and may rank methods differently.

5. Principal methodological descendants

Subsequent work modified SRGAN along three main axes: generator architecture, adversarial objective, and supervision beyond pixel or VGG feature losses. Several representative descendants illustrate how the original framework became a family of related methods rather than a single frozen design.

Variant	Core modification	Representative reported outcome
ESRGAN (Wang et al., 2018)	RRDB without Batch Normalization, Relativistic average GAN, perceptual loss using features before activation	won first place in the PIRM2018-SR Challenge
RankSRGAN (Zhang et al., 2021)	Siamese Ranker and rank-content loss to mimic perceptual metrics such as NIQE, Ma, or PI	PIRM-Test: NIQE $$\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}+\mathrm{PReLU}$$7, PI $$\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}+\mathrm{PReLU}$$8, PSNR $$\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}+\mathrm{PReLU}$$9 dB
EdgeSRGAN (Angarano et al., 2022)	$\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$0 residual blocks, transpose-convolution upsampling, int8 quantization, knowledge distillation	up to $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$1 fps inference
SRGAN-CKAN (Navaro-Aviña et al., 2 May 2026)	CKAN operators replace linear local mappings in the residual body	LPIPS improved from $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$2 to $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$3 with PSNR-Y $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$4 dB
SRTransGAN (Baghel et al., 2023)	transformer-based encoder–decoder generator and ViT discriminator	average improvement of $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$5 on PSNR and SSIM scores

ESRGAN was the most influential early reformulation. It replaced SRGAN’s residual blocks with Residual-in-Residual Dense Blocks without Batch Normalization, substituted the standard discriminator with a Relativistic average GAN, and computed perceptual loss on VGG features before activation rather than after. The paper reported that these changes produced consistently better visual quality with more realistic and natural textures than SRGAN, and human-rated MOS improvements of $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$6–$\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$7 point over SRGAN (Wang et al., 2018).

RankSRGAN addressed a limitation of classical SRGAN training: perceptual metrics such as NIQE, Ma, and PI are non-differentiable and therefore cannot be optimized directly. It introduced a separate Ranker trained with a pairwise margin-ranking loss and then used a rank-content loss to drive the generator in the direction of better perceptual scores. The reported PIRM-Test results placed RankSRGAN at NIQE $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$8, PI $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$9, and PSNR $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$0 dB, improving upon both SRGAN and ESRGAN in perceptual metrics while matching or slightly exceeding SRGAN’s PSNR (Zhang et al., 2021).

Later descendants broadened SRGAN’s operating envelope rather than only its perceptual fidelity. EdgeSRGAN tailored the architecture for CPU and Edge TPU devices through a smaller generator, post-training full-integer quantization, and knowledge distillation, reporting up to $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$1 fps inference while retaining image quality that remained better than bicubic under quantization (Angarano et al., 2022). SRGAN-CKAN pursued a different direction by increasing the expressivity of local operators through spline-based functional mappings under constrained hardware, while SRTransGAN re-architected both generator and discriminator with transformer modules to exploit global context more effectively (Navaro-Aviña et al., 2 May 2026, Baghel et al., 2023).

6. Domain-specific and scientific variants

SRGAN has been repeatedly specialized for domains in which the degradation model, anisotropy pattern, or task objective differs substantially from natural-image super-resolution. In prostate MRI, a direct adaptation retained the Ledig et al. design with modifications for single-channel input and VGG loss computation by channel replication. On $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$2 patients from PROSTATE-DIAGNOSIS and PROSTATEx, the study compared SRGAN with SRCNN, SRResNet, and Sparse Representation at $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$3 and $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$4, concluding that although SRGAN did not achieve the best PSNR or SSIM, it was visually the most similar to the original HR images according to MOS (Sood et al., 2019).

A related anisotropic MRI formulation modified SRGAN to expand only the degraded axis. In synthetic anisotropic super-resolution, low-resolution patches of size $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$5 were upsampled to $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$6 by a height-only sub-pixel convolution. For anisotropic $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$7 height-only super-resolution, bicubic interpolation reported PSNR $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$8 and SSIM $\mathrm{Conv}(3\times3,64)+\mathrm{BatchNorm}$9, whereas SRGAN reported PSNR $$\mathrm{Conv}(3\times3,256)\rightarrow\mathrm{PixelShuffle}\times2\rightarrow\mathrm{PReLU}$$0 and SSIM $$\mathrm{Conv}(3\times3,256)\rightarrow\mathrm{PixelShuffle}\times2\rightarrow\mathrm{PReLU}$$1 (Sood et al., 2019). This is a notable case in which an SRGAN derivative improved both perceptual appearance and conventional full-reference metrics.

Radiographic imaging motivated a different modification. AID-SRGAN introduced a medical practical degradation model that considers various degradation factors beyond downsampling, described as the first composite degradation model proposed for radiographic images. It coupled an attention-based denoising module with a super-resolution module, used a separate-joint training approach, and reported PSNR $$\mathrm{Conv}(3\times3,256)\rightarrow\mathrm{PixelShuffle}\times2\rightarrow\mathrm{PReLU}$$2 at scale factor $$\mathrm{Conv}(3\times3,256)\rightarrow\mathrm{PixelShuffle}\times2\rightarrow\mathrm{PReLU}$$3, which was $$\mathrm{Conv}(3\times3,256)\rightarrow\mathrm{PixelShuffle}\times2\rightarrow\mathrm{PReLU}$$4 higher than that obtained by SPSR (Huang et al., 2022).

In aerial imagery, SA-SRGAN made super-resolution explicitly semantic. Its generator was based on the Residual-in-Residual Dense Network of ESRGAN but used only $$\mathrm{Conv}(3\times3,256)\rightarrow\mathrm{PixelShuffle}\times2\rightarrow\mathrm{PReLU}$$5 RRDBs rather than $$\mathrm{Conv}(3\times3,256)\rightarrow\mathrm{PixelShuffle}\times2\rightarrow\mathrm{PReLU}$$6; a frozen segmentation network $$\mathrm{Conv}(3\times3,256)\rightarrow\mathrm{PixelShuffle}\times2\rightarrow\mathrm{PReLU}$$7 with $$\mathrm{Conv}(3\times3,256)\rightarrow\mathrm{PixelShuffle}\times2\rightarrow\mathrm{PReLU}$$8 classes supplied a domain-specific semantic feature loss. On the $$\mathrm{Conv}(3\times3,256)\rightarrow\mathrm{PixelShuffle}\times2\rightarrow\mathrm{PReLU}$$9, $4\times$00, $4\times$01, and $4\times$02 tasks, the method improved segmentation mIoU over RRDN-CNN by $4\times$03, $4\times$04, $4\times$05, and $4\times$06, respectively, and on the $4\times$07 task yielded IoU $4\times$08 against $4\times$09 for RRDN-CNN despite lower PSNR and SSIM (Frizza et al., 2022). This demonstrates that an SRGAN-style model can be optimized for downstream task utility rather than distortion fidelity alone.

Scientific simulation introduced another axis of specialization. PC-SRGAN retained the SRGAN backbone but added physical-consistency losses based on the governing PDE, with separate inner-domain and boundary-condition terms. For benchmark transient PDEs, the paper reported PSNR/SSIM of $4\times$10 for PC-SRGAN versus $4\times$11 for SRGAN and $4\times$12 for bicubic, while requiring only approximately $4\times$13 of the high-resolution training snapshots that vanilla SRGAN needed (Hasan et al., 10 May 2025). Here the standard SRGAN objection—that generated details may be visually plausible but physically meaningless—was addressed by moving physics into the loss rather than the architecture.

Other domain-specific variants targeted deployment or modality structure. Modified SRGANs for microscopy removed Batch Normalization, replaced sub-pixel or deconvolution upsampling with interpolation plus convolution, and used a tiling strategy to exploit GPU parallelism, reporting PSNR $4\times$14, SSIM $4\times$15, MOS $4\times$16, and whole-image inference time $4\times$17 s on $4\times$18 images (Bhatia et al., 2020). MLP-SRGAN introduced Residual MLP-Mixer in Residual Dense Blocks for single-dimension slice-direction upsampling in FLAIR MRI and reported $4\times$19 M parameters versus $4\times$20 M for ESRGAN, with training time $4\times$21 h versus $4\times$22 h and evaluation time $4\times$23 s versus $4\times$24 s (Mitha et al., 2023).

7. Limitations, misconceptions, and continuing relevance

The most persistent misconception surrounding SRGAN is that it is a PSNR-improving architecture. The original evidence shows the opposite: SRGAN deliberately trades a small drop in distortion metrics for a large gain in perceptual realism, and later work repeatedly confirmed that adversarial emphasis on textures raises pixel-wise deviations even when visual fidelity improves (Ledig et al., 2016, Sood et al., 2019). In that sense, SRGAN is not a universal replacement for MSE-optimized super-resolution; it is a model family for settings in which perceptual plausibility, edge fidelity, or downstream semantics matter.

A second limitation is that SRGAN does not, by itself, guarantee semantic, physical, or clinical correctness. SA-SRGAN depended on a high-quality, domain-specific segmentation network $4\times$25, and the paper noted that generalizing to new domains would require re-training or fine-tuning $4\times$26 (Frizza et al., 2022). PC-SRGAN required knowledge of the governing PDE, boundary conditions, appropriate time integrator, and additional hyperparameters for the physics losses (Hasan et al., 10 May 2025). MLP-SRGAN explicitly noted that at very high upsampling factors, hallucinated textures may drift away from physical reality if certain classes were under-represented in training (Mitha et al., 2023). Edge deployment likewise exposed a distinct failure mode: post-training int8 quantization reduced LPIPS quality even when inference throughput improved dramatically (Angarano et al., 2022).

These limitations do not diminish SRGAN’s importance. ESRGAN described SRGAN as a seminal work capable of generating realistic textures during single-image super-resolution (Wang et al., 2018), and the subsequent literature shows why. SRGAN established the generator–discriminator–perceptual-loss paradigm for super-resolution; later work either refined that paradigm, constrained it with semantics or physics, or reimplemented it under deployment constraints. This suggests that SRGAN’s enduring contribution is methodological: it defined super-resolution as a perceptual synthesis problem in which fidelity to the target image can be measured in feature space, by human preference, or by task performance, rather than by pixel error alone.