Real-ESRGAN

Overview

Real-ESRGAN is an open-source project for AI-based image and video restoration, with pretrained models for general imagery, anime, and animation. It suits users who want control over the restoration workflow and are comfortable choosing between a portable executable and a Python setup. Its breadth is a strength, but the portable version omits some Python features.

Released under the BSD-3-Clause license, the project permits use and redistribution in source and binary forms subject to the license conditions. Its repository has 36.9k stars and 4.5k forks, a sign of substantial developer interest rather than a guarantee of support or results. Browse Image Upscaling Software, Video Upscaling Software, AI Image Upscalers, and AI Video Upscalers to compare categories.

Key features

Model selection and restoration

Models cover general images, anime images, and anime video, with Real-ESRNet variants, a smaller anime-image model, and AnimeVideo-v3 alongside small anime-video models. GFPGAN integration adds face enhancement. That range makes the project useful for mixed image workflows and animation, rather than only one narrow image style.

Training uses synthetic data, and the released training code supports finetuning on a user's own or paired data. That is valuable for users prepared to customize a model; it is a technical option, not a turnkey benefit for someone seeking a simple hosted service.

Scaling and formats

The portable executable supports 2x, 3x, and 4x scaling, with 4x as the default. Python inference adds arbitrary output scaling through --outscale and includes the RealESRGAN_x2plus model. Because portable inference does not support all Python functions, users needing precise output sizing should favor the Python route.

Inference handles alpha-channel, grayscale, and 16-bit images, as well as JPG, PNG, and WEBP inputs. The executable can save JPG, PNG, or WebP, and batch processing is supported. These capabilities make it more flexible than a single-file, single-format upscaler, though the portable format and scale constraints still matter.

Ways to run it

Portable NCNN executables are available for Windows, Linux, and MacOS for Intel, AMD, and Nvidia GPUs. They bundle required binaries and models and do not require CUDA or a PyTorch environment, reducing setup demands for users who can work within their limits. The Python implementation requires Python 3.7 or newer and PyTorch 1.7 or newer, a better fit for users who need its fuller feature set and can manage those dependencies.

Online inference is available through the Tencent ARC Demo and two Colab demos. Hugging Face Spaces integration uses Gradio; related projects include NCNN-Android and VapourSynth. These routes expand how the models can be accessed, but do not replace the flexibility of running the Python implementation.

Pricing

Real-ESRGAN is free and open source under BSD-3-Clause. There is no paid plan to compare: the choice is between using the supplied portable executables, setting up Python and PyTorch, or trying the online demos. The repository also provides a pricing page URL, but that page returns Page not found.

Platforms

The project supports Linux, macOS, Windows, self-hosted use, and web-based access through online inference. Portable downloads target Windows, Linux, and MacOS across Intel, AMD, and Nvidia GPUs. The Python implementation is the option for users who need features unavailable in the portable executable.

Who it's for

Real-ESRGAN is a strong fit for technically capable users who want free restoration models, batch processing, multiple image types, and choices for general or anime content. It is also worth considering for users who want to finetune on their own data or process animation video.

It is less suitable for someone who wants a single polished application with no setup decisions, or who needs Python-only scaling features without installing Python and PyTorch. Users looking for a bounded hosted allowance or a more straightforward service may prefer an alternative.

Pros and cons

• Pro: General, anime, and anime-video models, plus GFPGAN face enhancement, serve several restoration needs in one project.
• Pro: Portable executables bundle their dependencies and models, avoiding CUDA and PyTorch setup.
• Pro: Batch processing and support for alpha, grayscale, and 16-bit images offer practical flexibility.
• Con: Portable inference lacks some Python functions, including arbitrary output scaling.
• Con: The Python route requires Python 3.7+ and PyTorch 1.7+, which adds setup for users seeking the fullest control.

Alternatives

NextGenUp is another free, open-source option, with a completely free plan and no trial limit or subscription; consider it if those stated terms are your priority.

Bigjpg suits readers who prefer a freemium service across web, desktop, and mobile platforms; its free plan caps use at 20 pictures per month, 5MB uploads, and 4x enlargement, with slow speed and a shared server.

LetsEnhance is a freemium API option whose free plan provides 10 credits, 24-megapixel input, 50MB input size, and 8-megapixel output. Choose it when those stated API limits match the job.

PixelPanda Image Upscaler offers API, extension, and web access, with one-time purchases of 10 credits for $1.99 a month or 50 credits for $10 a month; it may suit readers who prefer credit packs to a subscription.

Upscayl offers a 10-credit free trial and a Pro plan at $24.99 a month; consider it if a trial-based paid plan fits better than using Real-ESRGAN directly.

PicWish Image Upscaler is a freemium API alternative with limited non-HD daily downloads on its free plan and daily HD downloads for most online features on Pro.

SupaRes is a freemium API option whose free plan provides 10 processing credits, 64 megapixels, and one user; consider it when those bounded limits suit your workload.

Upscale.media provides a free plan with three credits and three downloads each month, making it an option for occasional API use within that monthly allowance.

Verdict

Choose Real-ESRGAN if you want a free, open-source restoration project with broad model coverage, batch support, and a portable route that avoids CUDA and PyTorch. Its main trade-off is the split between portability and control: users who need arbitrary scaling or other Python-only features must accept the Python and PyTorch requirements, while users wanting a more managed workflow should look elsewhere.