Real-ESRGAN

Real-ESRGAN is an open-source AI project for restoring and upscaling images and video, best suited to users who want model choices and control over how inference runs. Its portable apps lower setup demands, while Python offers more scaling flexibility; choose the version carefully if arbitrary output sizes matter.

Overview

Models cover general images, anime artwork and animation video, with Real-ESRNet variants and GFPGAN face enhancement also included. Training uses synthetic data, and users can fine-tune the released training code on their own or paired data. The BSD-3-Clause license permits redistribution and use in source and binary forms subject to its conditions.

Inference is available through Python, portable NCNN executables and online demos, including Tencent ARC and Colab. Hugging Face Spaces integration uses Gradio, and the project also points to NCNN-Android, VapourSynth and NCNN integrations.

Compare it with other Image Upscaling Software, AI Image Upscalers, AI Video Upscalers and Video Upscaling Software.

Key features

Models and image support

General-image and anime-image models suit still-image restoration, while AnimeVideo-v3 and smaller anime-video models target animation. GFPGAN adds a face-enhancement option. Inference handles alpha channels, grayscale and 16-bit images, alongside JPG, PNG and WebP inputs; the portable executable can output those same formats. Batch processing is supported, useful when processing multiple images rather than one at a time.

Scaling and runtimes

The portable NCNN version offers fixed scale ratios of 2, 3 or 4, with 4 as the default. Its bundled binaries and models avoid the need for CUDA or a PyTorch environment, and downloads are provided for Windows, Linux and macOS on Intel, AMD and Nvidia GPUs. The tradeoff is that portable inference lacks some Python functions, including arbitrary output scaling through --outscale.

Python requires version 3.7 or newer and PyTorch 1.7 or newer. It is the better fit when custom output scaling is important, including use with the RealESRGAN_x2plus model, but brings runtime setup that the portable executable avoids.

Pricing

Real-ESRGAN is free and open source, with no paid tiers described. That makes it a strong option for users comfortable running the software themselves; the tradeoff is choosing and setting up an inference route rather than selecting a commercial service plan.

Platforms

Portable downloads cover Windows, Linux and macOS, and the project also offers online inference through Tencent ARC and Colab demos. Its Python implementation has explicit Python and PyTorch minimum versions. The project is also categorized for web and self-hosted use, though the available demos and portable executables are distinct ways to access it.

Who it's for

Choose Real-ESRGAN if you want free restoration models, batch processing, specialized anime options or the ability to fine-tune on your own data. It also suits users who need alpha, grayscale or 16-bit image support. It is less compelling if you need arbitrary scaling but want to avoid Python setup, or if you prefer a single hosted workflow with a stated per-use allowance.

Pros and cons

Pro: Portable executables bundle binaries and models, avoiding CUDA and PyTorch setup.

Pro: General, anime and video models, plus GFPGAN and fine-tuning support, cover varied restoration needs.

Pro: Batch processing and alpha-channel, grayscale and 16-bit support accommodate workflows beyond ordinary single-image enlargement.

Con: Portable scaling is limited to 2x, 3x or 4x, and it omits Python's arbitrary outscale option.

Con: Python users must meet the Python 3.7+ and PyTorch 1.7+ requirements.

Alternatives

NextGenUp is another free option, with a completely free plan and no trial limit or subscription. Pick it if that stated plan arrangement is a better fit.

Bigjpg offers a free plan capped at 20 pictures per month, slow speed, shared server processing, 5MB uploads and 4x enlargement. Consider it if a web-oriented service with those defined free-plan limits suits your needs better than Real-ESRGAN's selectable runtimes.

SupaRes has a free plan with 10 processing credits, a 64-megapixel limit and one user. Its paid Basic plan costs $9.00 per month and includes 100 processing credits; choose it if a credit-based API or web service better matches your workflow.

Imglarger provides a free plan with 100 credits per month, one image at a time, a 1024px maximum input side and 2x scaling. It may suit users who want those explicit monthly limits and mobile or web access.

PixelPanda Image Upscaler sells one-time credit purchases, including 10 credits for $1.99 per month and 50 credits for $10.00 per month. Consider it if buying credits without a subscription is preferable.

Adima Image Upscaler Pro offers a free plan with 20 free 4x upscales, GPU or CPU support, three high-quality filters, 16x upscale and batch processing. It is worth considering if those stated limits and platform options fit better.

Cutout.Pro offers a subscription with 80 credits per month for $5.00 per month. Choose it if a paid monthly credit allowance is more suitable.

LetsEnhance gives new users 10 free credits, with a 24-megapixel input limit, 50MB input limit and 8-megapixel output limit. Its API platform may suit users who want that credit-based offering.

Verdict

Real-ESRGAN is a strong choice for users who want free, adaptable restoration models and can select a runtime to match their needs. Its main advantage is the breadth of models and image handling; look elsewhere if you need an especially straightforward hosted service or want arbitrary scaling without the Python implementation.