ESRGAN

ESRGAN is a 2018 GAN-based image upscaling model that produces sharper textures than previous super-resolution methods, introducing the RRDB block and relativistic discriminator that still underpin modern AI upscalers.

What It Is

When someone uploads a small or low-resolution image into an AI upscaler — Topaz Gigapixel, Magnific, an Automatic1111 or ComfyUI workflow — the model has to invent pixels that were never captured. Earlier methods (bicubic interpolation, basic CNNs) produced bigger images that still looked blurry: smooth where there should be hair, mushy where there should be brick. ESRGAN was the breakthrough that made AI-upscaled images finally look like photographs instead of soft enlargements.

ESRGAN — Enhanced Super-Resolution Generative Adversarial Networks — was published at the ECCV 2018 Workshops by Xintao Wang and colleagues. According to ESRGAN arXiv, it improved on the earlier SRGAN paper through three architectural changes:

• The Residual-in-Residual Dense Block (RRDB): a deeper feature-extraction unit that stacks residual blocks inside residual blocks, with batch normalization removed. Batch norm was discarded because it was producing visible artifacts in upscaled outputs.
• A relativistic discriminator: instead of asking “is this image real or fake?”, it asks “is this image more realistic than that one?” The training signal becomes comparative, which pushes the generator toward more believable textures.
• A perceptual loss computed before the activation layer: the loss compares features earlier in a deep network, where supervision for fine textures is stronger.

According to ESRGAN GitHub, this combination won 1st place in the PIRM2018-SR Challenge (region 3) on perceptual index. The original repository at xinntao/ESRGAN is now mainly a historical reference. The practical successor is Real-ESRGAN (2021), trained on synthetically degraded images so it survives the messy compression, blur, and noise of real photos and screenshots.

How It’s Used in Practice

Most users meet ESRGAN through the Real-ESRGAN fork rather than the original. It ships as the default GAN upscaler in ComfyUI and Stable Diffusion WebUI workflows, drives the “AI upscale” buttons in tools like Upscayl, and appears as one of the model options inside Topaz Gigapixel and similar consumer apps.

Where it shines: re-sharpening illustrations, anime frames, screenshots, and synthetic AI-generated images where you want clean lines and crisp texture without hallucinated content. A typical workflow runs a 4× Real-ESRGAN pass after image generation to bring a square render up to print-ready resolution. For larger jumps, it gets paired with tiled upscaling so a single GPU can handle big images without running out of memory.

Where it struggles: real-world photos with heavy degradation, faces that need identity-preserving detail, and any scenario where you want truly new content (not just sharper versions of what’s there). For those, diffusion-based upscalers like Magnific or SUPIR usually win.

Pro Tip: Don’t reach for ESRGAN when your input is already clean and high-resolution — you’ll get smoothing artifacts on detail that didn’t need fixing. Use it on the small, blurry, or compression-damaged stuff. And always run Real-ESRGAN, not the 2018 original; the successor handles real-world inputs the older model was never trained for.

When to Use / When Not

Common Misconception

Myth: ESRGAN “adds detail” by hallucinating new content into the image. Reality: ESRGAN reconstructs plausible texture based on patterns it learned during training. It doesn’t invent objects or features that weren’t already implied by the input pixels. That kind of generative reimagining is what diffusion-based upscalers do — and it’s also why ESRGAN outputs look more faithful but less imaginative than Magnific or SUPIR.

One Sentence to Remember

ESRGAN is the 2018 GAN that made AI upscaling genuinely usable, and through its Real-ESRGAN successor, it’s still the default first-pass enhancer in most image generation pipelines.

FAQ

Q: What’s the difference between ESRGAN and Real-ESRGAN? A: Same architecture lineage and lead author. Real-ESRGAN was retrained with synthetic degradations so it handles compressed, noisy, real-world images. The original was trained on clean inputs and is mostly a historical reference now.

Q: Is ESRGAN better than diffusion-based upscalers like Magnific? A: Different tools. ESRGAN is faster, runs locally on modest GPUs, and stays faithful to the input. Diffusion upscalers add more imagined detail and produce richer textures, but cost more compute and can drift from the source.

Q: Can I run ESRGAN on my laptop? A: Yes. Real-ESRGAN runs on consumer GPUs and even in CPU mode. Standalone apps like Upscayl wrap it for one-click use, and ComfyUI workflows let you chain it with other models in a local image generation pipeline.

Sources

• ESRGAN arXiv: ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks - Original Wang et al. paper introducing the RRDB block, relativistic discriminator, and pre-activation perceptual loss.
• Real-ESRGAN GitHub: Real-ESRGAN: Practical Algorithms for General Image/Video Restoration - The successor trained on synthetic degradations, the version most production pipelines actually use.