Real-ESRGAN: Open-Source AI Image and Video Restoration
github.com · checked 2 Oct 2026
Open-source AI tools for restoring and enlarging images and video.
How it plugs in
Signal on Image Upscaling Software8.3#3 of 99
Real-ESRGAN is free software that applies AI to image and video restoration. It offers batch jobs, desktop and self-hosted workflows, portable executables, browser demos, and models spanning general imagery, anime, and video. JPG, PNG, and WEBP inputs are accepted, and its executable produces corresponding image files.

Review
In a line
This open-source project uses AI to restore and enlarge images and video. It runs on Linux, macOS, Windows, the web, or a self-hosted setup.
What it does
Its pretrained options address general imagery, anime pictures, and animation video. The collection includes AnimeVideo-v3, smaller anime-video models, anime-image and general models, plus Real-ESRNet variants.
JPG, PNG, and WEBP files can be processed, including in batches. Inference handles alpha channels, grayscale, and 16-bit images. Portable runs provide 2x, 3x, or 4x scaling, with 4x as the default; Python can choose another output size with --outscale. GFPGAN adds face enhancement.
How it connects
The project is released under the BSD-3-Clause license, whose conditions cover redistribution and use. Training can be finetuned with a user's own data or paired data.
Online inference is available through the Tencent ARC Demo and two Colab demos. Hugging Face Spaces is connected through Gradio, and the project lists NCNN-Android, VapourSynth, and NCNN projects that use it.
Portable NCNN downloads are available for Windows, Linux, and macOS on Intel, AMD, and Nvidia GPUs. Required binaries and models are bundled, so CUDA and PyTorch are unnecessary for that executable.
Good fit
This suits people seeking free software for enlarging or restoring still images and video, including anime content. Desktop use, batch jobs, online demos, and self-hosting support different workflows.
It can also work for users who want to further train with their own or paired data, or who prefer Python for additional control.
Watch for
Python setup needs Python 3.7 or later together with PyTorch 1.7 or later. The portable executable is easier to launch, but lacks some Python inference capabilities, including outscale.
Portable processing is restricted to 2x, 3x, or 4x ratios, so verify that those scales and the supported file formats fit your needs.
Pricing
Real-ESRGAN has a free pricing model, and the project identifies a free plan.
Verdict
Choose Real-ESRGAN if you want open-source restoration that can run locally, through a browser demo, or from a portable executable. Its general, anime, and animation-video models, along with batch processing and custom training, provide a flexible toolkit.
It may not suit people seeking a turnkey service with no setup: Python requires specific software versions, while the portable build omits some Python functions.
Our take
Real-ESRGAN works well for users who value free, open-source restoration and can select among Python, portable, desktop, or online methods. Multiple model families, batch handling, custom training, and common image formats broaden its use. Setup is the key trade-off: Python needs Python 3.7+ and PyTorch 1.7+, while portable builds omit some capabilities. Pick it for local control; a simpler hosted process may be preferable.
- The BSD-3-Clause license makes the software free and open source
- Runs on Linux, macOS, Windows, web, and self-hosted setups
- Batch handling works with common JPG, PNG, and WEBP files
- Models cover general images, anime, and animation video
- Portable builds include binaries and models without CUDA or PyTorch
- Python use requires Python 3.7+ and PyTorch 1.7+
- Portable executables do not support every Python inference feature
- Portable inference is limited to 2x, 3x, or 4x scale ratios
- Pricing details supplied here identify a free plan only
- Pick it ifPeople restoring images or video with free open-source software; Anime image and animation-video upscaling; Users who need local, portable, or self-hosted processing
- Skip it ifYou need every Python feature in a simple portable executable; You want a hosted workflow without choosing a runtime; Your required scale or file format falls outside the stated support
Pricing
Real-ESRGAN uses a free pricing model and has a free plan.
Questions
Is Real-ESRGAN free?
Yes. It uses a free pricing model, and the open-source project is distributed under the BSD-3-Clause license.
Which operating systems can run Real-ESRGAN?
It supports Linux, macOS, and Windows, plus web and self-hosted use.
What files can it process?
It accepts JPG, PNG, and WEBP input. The portable executable can output JPG, PNG, and WebP.
Does Real-ESRGAN support anime video?
Yes. AnimeVideo-v3 and smaller anime-video models are included, alongside general and anime-image models.
Does it need CUDA or PyTorch?
The portable executable contains the needed binaries and models, avoiding a CUDA or PyTorch environment. Python use needs Python 3.7+ and PyTorch 1.7+.
Can I train it with my own data?
Yes. The released training code supports finetuning with your own data or paired data.