Real-ESRGAN
An image and video upscaler trained on synthetic data, with portable NCNN apps and PyTorch face enhancement.
Portable no-CUDA executables or a PyTorch script, with anime models and optional GFPGAN face enhancement.
Real-ESRGAN is a practical image and video restoration and upscaling tool. Photographers and editors who have low-resolution or degraded images can run it from the command line, and the portable executable avoids a CUDA or PyTorch setup.
What it is
It extends ESRGAN to restore real-world degraded images, trained with pure synthetic data. You can run portable NCNN-Vulkan executables for Windows, Linux and macOS, or the PyTorch inference script, with models for general photos, anime images and anime video, tile options for large images and optional GFPGAN face enhancement.
Who it's for
It suits photographers, archivists and video editors who want to enlarge or clean up old images. It is a command-line tool, so it suits people willing to type commands or use one of the community GUIs the README lists.
Key features
Portable NCNN executables
Windows, Linux and macOS builds include binaries and models and need no CUDA or PyTorch.
General and anime models
Includes realesr-general-x4v3 and models tuned for anime images and anime video.
Face enhancement
Integrates GFPGAN to restore faces during upscaling.
Denoise strength control
The -dn option balances noise removal to avoid over-smoothed results.
Tiling and alpha support
Tile options handle big images; alpha channel, grayscale and 16-bit images are supported.
Finetuning and training
Training code and finetuning on your own data are documented.
Up and running in 3 steps
Our honest verdict
What's great
- The portable executables run without installing CUDA or PyTorch.
- Dedicated anime image and anime video models exist.
- Arbitrary output scales are supported through --outscale.
- Community GUIs such as Waifu2x-Extension-GUI are listed for users who dislike command lines.
Where it falls short
- The README shows commands, not a graphical app, so non-technical users need a third-party GUI.
- The PyTorch route needs Python 3.7 or later, PyTorch 1.7 or later, basicsr, facexlib and gfpgan, and downloaded model files.
- Upscaling invents plausible detail; it is not a way to recover real information that was never captured.
- The last release listed on GitHub is v0.3.0 from 2022, so check the repo activity yourself before relying on new features.
Pricing & license
Real-ESRGAN is free under the BSD-3-Clause licence. The README links online demos on Replicate, Colab and Hugging Face; no paid plan from the project itself is shown, and the only cost is your own computer.
Open remote AI-training roles
767 open remote AI-training roles, updated November 14, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
Do I need a GPU?
The portable executables use Vulkan via NCNN and the README says no CUDA or PyTorch environment is needed. Performance depends on your graphics hardware.
Can it fix faces?
Yes, with GFPGAN integration for face enhancement via --face_enhance.
Is it okay for commercial work?
The code is BSD-3-Clause. Check the licence of each model file you download, since the LICENSE file covers the repository code.
