Real-ESRGAN

Overview

Real-ESRGAN is an open-source AI restoration project for enlarging and repairing images and video, aimed at people comfortable choosing between a desktop executable, Python code, or an online demo.

The project is released under the BSD-3-Clause license, which permits use and redistribution in source and binary forms subject to its conditions. Its repository has 36.9k stars and 4.5k forks, a substantial community footprint for an open-source tool.

For a broader comparison, browse Image Upscaling Software, AI Image Upscalers, AI Video Upscalers, and Video Upscaling Software.

Key features

Models and restoration

Real-ESRGAN offers models for general images, anime images, and animation video, with Real-ESRNet variants, a smaller anime-image model, and AnimeVideo-v3 among the options. GFPGAN integration adds face enhancement. Synthetic-only training is a defining part of the project, while support for finetuning on a user's own or paired data gives technically capable users a way to adapt the released training code.

Inference handles alpha-channel, grayscale, and 16-bit images, plus JPG, PNG, and WebP inputs. Batch processing is supported. These capabilities make it more suitable for varied image collections than a tool limited to a single image type, though the stated input formats remain a practical boundary.

Scaling and ways to run it

The Python implementation supports arbitrary output scaling through the --outscale option and includes an x2 model. Portable inference is simpler to deploy: its executable bundles binaries and models, needs no CUDA or PyTorch environment, and is offered for Windows, Linux, and MacOS across Intel, AMD, and Nvidia GPUs. That convenience comes with a trade-off: the portable version supports fixed 2x, 3x, or 4x ratios, defaults to 4x, and does not expose every Python feature, including arbitrary outscale.

Python requires version 3.7 or newer and PyTorch 1.7 or newer. Online inference is also possible through a Tencent ARC Demo and two Colab demos. Hugging Face Spaces integration uses Gradio, and related projects include NCNN-Android, VapourSynth, and NCNN.

Pricing

Real-ESRGAN is free, with a free plan and no paid tiers described. That makes it a strong fit for users who can run the software themselves or use its online demos; there is no published subscription or per-use price to weigh against commercial services.

The project is distributed under the BSD-3-Clause license, which permits redistribution and use in source and binary forms subject to its conditions. The maintainers invite questions by email at [email protected] or [email protected].

Platforms

Real-ESRGAN is listed for Linux, MacOS, Windows, web, and self-hosted use. Portable executables cover the three desktop operating systems, while online demos provide a browser-accessible route. Python users need Python 3.7 or newer and PyTorch 1.7 or newer, so the code path demands more setup than the bundled executable.

Who it's for

Choose Real-ESRGAN if you want a free, open-source restoration toolkit with model choices spanning general images and anime, batch processing, and a path to custom finetuning. The portable build suits users who want bundled dependencies; Python is the better fit when arbitrary scaling or the full inference feature set matters. Readers seeking a managed workflow rather than executable or code-based options may prefer a dedicated hosted alternative.

Pros and cons

• Pro: General, anime-image, and anime-video models, plus GFPGAN face enhancement, give users distinct restoration paths rather than a single model choice.
• Pro: Bundled portable executables avoid a CUDA or PyTorch setup, lowering the barrier to local inference.
• Pro: Free distribution under BSD-3-Clause and released finetuning code support use and adaptation without a paid plan.
• Con: Portable scaling is limited to 2x, 3x, or 4x; arbitrary output scaling requires the Python implementation.
• Con: The Python route depends on Python 3.7+ and PyTorch 1.7+, which adds installation and compatibility work.

Alternatives

NextGenUp is another free, open-source option for users who want a stated no-subscription, no-watermark offering rather than Real-ESRGAN's model-and-runtime choices.

Bigjpg may fit users who prefer a service with a defined free quota: its free plan allows 20 pictures per month, slow speed, shared server processing, a 5MB upload cap, and up to 4x enlargement.

LetsEnhance is worth considering when its API-based plans and stated input and output limits better suit the workflow: its free tier includes 10 credits, 24-megapixel input, 50MB input size, and 8-megapixel output.

PixelPanda Image Upscaler offers API, extension, and web access with one-time credit purchases, making it an alternative for users who want those access routes and purchase terms.

Upscayl has a free trial with 10 credits and a Pro plan priced at $24.99 a month; consider it if a credit-based trial and paid plan are preferable.

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

Krea Enhancer may suit users who want compute-unit allowances, including 100 units per day on its free plan and 5,000 per month on Basic at $9 a month.

SupaRes has a free plan with a one-time quota of 10 processing credits, 64 megapixels, and one user, a clearer fit when those account and quota limits match the task.

Verdict

Real-ESRGAN is a strong choice for image and animation enthusiasts, developers, and self-hosters who want free restoration models, local execution, and room to customize training. Its main advantage is the range of models and execution paths; look elsewhere if you need a fully managed service or want arbitrary scaling without taking on the Python and PyTorch requirements.