
What is Real-ESRGAN?
Real-ESRGAN is an algorithm and open-source codebase that uses AI to upscale and restore images and videos that are blurry, noisy, or heavily compressed. It was published in 2021 by Xintao Wang and colleagues at Tencent ARC Lab and the Shenzhen Institute of Advanced Technology. According to the GitHub API (accessed 04/11/2026), the xinntao/Real-ESRGAN repository was created on 19/7/2021, has around 37,300 stars, uses the BSD-3-Clause license, and is written in Python. The most recent code push was on 6/8/2024 and the latest release is v0.3.0 (20/9/2022), so the repository is essentially inactive. This article explains what Real-ESRGAN can do, hidden costs when running it, three ways to get started according to the README, image copyright risks, and when to choose other tools instead.

Real-ESRGAN takes an image or video frame, upscales it using a deep learning model, and returns a sharper image. Diagram: OPCAIOS.
What is Real-ESRGAN and who created it?
Real-ESRGAN is a super-resolution model for realistic images, an extension of ESRGAN and trained entirely on synthetic data. The author group includes Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan, who belong to Tencent ARC Lab and the Shenzhen Institute of Advanced Technology.
Normal upscaling just stretches pixels, so the image becomes blurry; Real-ESRGAN "guesses" missing details based on what the model has learned. What makes Real-ESRGAN different, according to the arXiv paper (2107.10833) and the GitHub README (accessed 04/11/2026), is that the authors simulated many types of real-world image corruption: blur, noise, JPEG compression, repeated downsampling. Thanks to this, the model handles old photos, images downloaded from social media, or animation frames. The paper was presented at the ICCV Workshops conference in 2021. Many applications with graphical interfaces were built on this foundation, for example Upscayl, according to the README.
What are Real-ESRGAN's main features?
Real-ESRGAN has many models for different image types, plus options for processing large images and sharpening faces. The table below summarizes according to the GitHub README (accessed 04/11/2026), restated in English.
| Feature | Brief description |
|---|---|
| General image model | RealESRGAN_x4plus (4x upscale) and RealESRGAN_x2plus (2x upscale) |
| Smaller model for general scenes | realesr-general-x4v3, with -dn noise reduction option to avoid flat images |
| Anime images | RealESRGAN_x4plus_anime_6B, smaller model size |
| Animated video | AnimeVideo-v3 model, processes each frame |
| Face sharpening | Integrated GFPGAN via --face_enhance option |
| Custom ratio | --outscale upscales to any ratio, remainder uses standard interpolation |
| Large images | Tiling to run on low-memory GPUs |
| Training | Has training and fine-tuning code on private data |
One point often misunderstood: the actual ratio of the model is 2x or 4x. When you choose 3.5x, Real-ESRGAN still upscales using the 4x model then downscales again with cheaper interpolation, according to the README.
Is Real-ESRGAN free, and what are the hidden costs?
Real-ESRGAN is free: the code is under the BSD-3-Clause license, allowing commercial use if you keep the copyright notice. The real cost lies in hardware: a machine with a GPU, electricity, or money to rent cloud GPUs when processing large volumes.
| Item | Information (accessed 04/11/2026) |
|---|---|
| Code license | BSD-3-Clause according to GitHub API |
| Software fee | 0 đồng, no official paid version |
| Hardware | The ncnn version runs on Intel, AMD, Nvidia GPUs; Python version integrates with Nvidia GPUs with CUDA |
| Online service | Has demo on Replicate, Colab, Hugging Face Spaces; Replicate charges per its own terms, we have not read the pricing page |
Hardware costs do not include tax and may change. We have not measured processing speed on any specific machine, so we do not give time figures.
How to install and start using Real-ESRGAN?
Real-ESRGAN has three ways to get started according to the README: try it online, download the pre-built ncnn-vulkan executable, or install the Python code. People who don't know code should start with the pre-built executable since it doesn't require installing Python, PyTorch, or CUDA.
The installation method using Python, according to the README on GitHub (read on 04/11/2026), requires Python 3.7 or higher and PyTorch 1.7 or higher. The steps are as follows:
- First, download the source code to your machine:
git clone https://github.com/xinntao/Real-ESRGAN.git
cd Real-ESRGAN
- Next, install the dependencies:
pip install basicsr
pip install facexlib
pip install gfpgan
pip install -r requirements.txt
python setup.py develop
- Download the model file into the
weightsdirectory:
wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth -P weights
- Finally, run inference on the image folder
inputs, with results saved in theresultsfolder:
python inference_realesrgan.py -n RealESRGAN_x4plus -i inputs --face_enhance
With the pre-built executable, you download the compressed file for Windows, Linux, or macOS from the Releases page, extract it, then run the sample command in the README:
./realesrgan-ncnn-vulkan.exe -i input.jpg -o output.png -n model_name
A practical example (hypothetical): a shop owner has 50 old product images that are only 600 pixels wide and wants to post them on a marketplace that requires larger images. They run the ncnn executable with the realesrgan-x4plus model on the entire folder, then review each image and discard the ones where text on the packaging is distorted.
What are Real-ESRGAN's limitations, risks, and legal considerations?
Real-ESRGAN has three groups of limitations: a repository that is rarely maintained, uneven quality with difficult images, and copyright risks regarding the images you process. The model doesn't recover truth; it creates details that look plausible.
The first limitation is maintenance. According to the GitHub API (read on 04/11/2026), the most recent code push was on 6/8/2024 and the repository has around 650 open issues; with newer Python and PyTorch versions, you may encounter compatibility errors that you need to handle yourself. The next limitation is technical: the README notes that the ncnn version divides images into multiple tiles and then combines them, so there may be slight misalignment between tiles, and it doesn't support outscale. With small text or faces, the model might "fabricate" details that don't exist in the original image, so don't use upscaled images as evidence or documentation.
Finally, there's the legal aspect. The BSD-3-Clause license applies only to the code, not to the images or videos you process; upscaling someone else's image doesn't make you the owner. When sharpening real people's faces, consider regulations about personal data. This is reference information, not legal advice; we have not cross-referenced original legal texts.
How is Real-ESRGAN different from other image upscaling tools?
Real-ESRGAN is a free, open-source model platform that runs locally, suitable for people who can handle setup. Other options usually trade simplicity for cost or vice versa. The table below is a general description, not a measurement.
| Tool | Main point | Suitable for |
|---|---|---|
| Real-ESRGAN | Open source, many models, runs locally, rarely updated repository | Command-line users, developers |
| Upscayl | GUI application built on Real-ESRGAN | People who don't know code |
| waifu2x | Long-standing tool for anime images, README has comparison with anime models | Illustration makers |
| Clarity Upscaler | Upscaling that "creatively" adds details using diffusion models | Art images, accepting image changes |
| Node in ComfyUI | Attach upscaling models into image generation workflow | People already using ComfyUI |
When should you use Real-ESRGAN?
Real-ESRGAN is suitable when you need to upscale many images, want to run it on your own machine, and don't want to pay per use. If you only need a few images, an online service or application with a graphical interface will be faster.
| Situation | Suggestion |
|---|---|
| A few random images, don't know code | Try the online demo or Upscayl |
| Hundreds of product images, have a GPU machine | Pre-built ncnn running on folders |
| Need automation in a workflow | Python version, call from a script or n8n |
| Anime or illustration images | RealESRGAN_x4plus_anime_6B model |
| Images for evidence, documentation | Don't use, since the model can create fake details |
Our Experience at OPCAIOS
We downloaded the Real-ESRGAN repository to our work machine to try upscaling image resolution, but we haven't incorporated it into opcaios.org's daily workflow yet. We haven't measured quality or speed numerically. Real-ESRGAN is also listed in the Repo category of the Utilities section on opcaios.org. When considering long-term use, we evaluated four criteria:
- Is the repository still being updated: Real-ESRGAN has been virtually stalled since 2024, so we'd have to handle compatibility issues ourselves.
- License: BSD-3-Clause is easy to work with, but we need to check GFPGAN and accompanying repositories separately.
- Hidden costs: it needs a GPU; if we rent one, we calculate by running hours.
- Can it replace tools we're already using: with a few images each week, an application with a graphical interface might be sufficient.
Read more: What is ComfyUI?, What is Docker?, What is Clarity Upscaler?, What is GitHub?, 5 best AI image sharpening tools 2026.
Frequently Asked Questions
Can Real-ESRGAN run on a machine without an Nvidia card?
Yes. According to the README on GitHub (read 04/11/2026), the ready-to-run build realesrgan-ncnn-vulkan supports Intel, AMD, and Nvidia GPUs, with files for Windows, Linux, and macOS. This build comes with models pre-packaged, so you don't need to install CUDA or PyTorch. The Python version runs best with Nvidia GPUs that have CUDA; running on CPU is possible but much slower. On weaker machines, you can use the tile option to reduce GPU memory usage.
Can Real-ESRGAN upscale video?
Yes, but indirectly. Real-ESRGAN processes frame by frame, so for video you typically extract frames using ffmpeg, upscale each frame, then recombine them with the original audio. The README has a separate AnimeVideo-v3 model for animated video and a Colab demo for anime video. With real-world video, results may flicker between frames; try a short clip first.
Can Real-ESRGAN be used for commercial purposes?
Real-ESRGAN code is under the BSD-3-Clause license, which according to the GitHub API (read 04/11/2026) allows commercial use if you retain copyright notices and license conditions. However, accompanying repositories like GFPGAN, BasicSR, and facexlib have their own licenses, and you need to check each one. The software license also doesn't grant rights to the images you process. This is reference information, not legal advice.
Why does an image upscaled by Real-ESRGAN look artificial?
Real-ESRGAN creates plausible details based on training data; it doesn't recover actual details that have been lost. With human skin, hair, small text, or fabric patterns, the model may make the image look flat like a drawing or create strange artifacts. You can try the realesr-general-x4v3 model with the -dn noise reduction option to balance, according to the README. GFPGAN sharpens faces but may alter their appearance.
Can someone without coding knowledge learn to use Real-ESRGAN?
Yes, if you start with easy steps. First, try the online demo to see if the results fit your needs. Next, download the ready-to-run ncnn-vulkan build and run a sample command from the README with one image. Finally, if you want to automate it for many images, ask an AI assistant like Claude Code to write a script that calls the command for an entire folder. If you want to systematically learn how to integrate AI tools like this into your daily work, check out OPCAIOS's AI application training course.
Source: GitHub xinntao/Real-ESRGAN (README, API, Releases page, read 04/11/2026), arXiv paper 2107.10833; not read/not verified: detailed licenses for GFPGAN, BasicSR, facexlib, Replicate pricing page, speed and quality on specific hardware; star count and pricing subject to change. This experience is OPCAIOS's.
Translated from Vietnamese with AI. Read the original.
Want to apply this in your business?
Book a free 30-minute call (Vietnam time, in English or Vietnamese).


