Upscayl

Where Upscayl sits

Upscayl is a free, open-source desktop application that enlarges images with AI and runs entirely on your own machine. That sentence contains its whole positioning: in a category of subscription upscalers that upload your pixels to somebody's server and meter you in credits, Upscayl is the one you download, run offline, and never pay for. It is developed by Nayam Amarshe and TGS963 and hosted on GitHub under the AGPL-3.0 license, where at the time of writing the repository shows just over 50,000 stars and 2,595 forks — one of the more visible open-source image projects around. The first commit dates to July 2022, and the project has shipped 29 releases since, with v2.15 the current release in this pass.

What I look for in this category is simple: does the enlargement hold up at 100% zoom on the hard parts — hair, foliage, fine text — and can I run it on the volume of images I actually have? Upscayl answers the first question with "usually, and you get to choose the model" and the second with "yes, but your GPU is the limit."

What it actually does

Upscayl wraps Real-ESRGAN and a set of community upscaling models in an Electron desktop app for Linux, macOS and Windows (distributed as Flatpak/AppImage and installers, among other formats). You drop in an image or a folder, pick a model and a scale factor, and it renders an enlarged output locally. Out of the box it ships several models aimed at different material — Ultrasharp and Ultramix Balanced for general photography, High Fidelity (the HFA2k model) and Remacri for detail retention, dedicated digital-art models, and the standard Real-ESRGAN 4x general model — and a community repository of additional custom models is available for anyone who wants to experiment. Batch processing is built in, scaling factors are configurable rather than fixed at 4x, there is a GPU-selection setting for multi-GPU machines, and recent versions have added a redesigned interface, better batch handling and an in-app updater. Because inference runs on your GPU (Vulkan via the ncnn backend), nothing is uploaded anywhere and there is no account, no watermark and no per-image cost. Development is funded by donations and sponsors rather than a paid tier.

My testing notes

Two observations from my runs. First, output quality is model-dependent in a way that is genuinely useful rather than confusing. In my tests on a set of low-resolution web images — a portrait at 400px wide, a landscape with dense foliage, and a screenshot containing small UI text — Ultrasharp gave the most consistently natural skin and fabric texture, the foliage landscape came out slightly waxy with the default model but much cleaner with High Fidelity, and the screenshot was the honest failure case: small aliased text gains pixels but not legibility, and no ESRGAN-family model I tried reconstructed readable glyphs from genuinely undersized text. That last point is a category limit, not an Upscayl bug, but the app makes it easy to compare models on the same input, which is how I settled on per-asset defaults for my workflow.

Second, speed is entirely hardware-bound, and that cuts both ways. On a machine with a discrete GPU, a 2x upscale of a 12-megapixel photo took seconds and batch jobs chugged along unattended; on a laptop without Vulkan-capable graphics the app is slow or effectively unusable, and there is no cloud fallback by design. Upscayl's own documentation is upfront about GPU selection and Windows performance-mode quirks, which matched what I saw: letting the OS throttle the GPU doubled render times until I changed the power settings. The other practical note is output management — batch runs write files next to the source or wherever you point them, and the app has gotten better at naming and folder handling over successive releases, but it is still a tool rather than a pipeline: no watched folders, no scripting surface, no API.

Who it is actually for

The clearest way I can frame the audience: Upscayl is the tool for someone with a large personal archive, privacy constraints, or a simple refusal to pay per image. Photographers cleaning up old low-resolution shots, artists prepping portfolio prints from smaller originals, designers who inherited web-sized assets and need print-sized versions — all of these run into Upscayl's sweet spot. I have also seen it do real work for small e-commerce sellers who need catalog images sharpened without uploading proprietary product photos to a third-party server; running locally means the images never leave the machine, which for some businesses is not a preference but a policy.

The corollary is who should skip it. If your machine is a basic laptop with integrated graphics, the compute you save in dollars you pay in time. If your material is genuinely tiny — thumbnails, old forum attachments, camera-phone shots from 2009 — set expectations before you start: upscaling makes small images bigger and cleaner, it does not reconstruct information that was never captured, and in my tests the jump from 2x to 4x on already-poor sources produced larger files more often than better ones. Upscayl makes experimenting with models cheap, which is exactly how to calibrate that expectation on your own images rather than mine.

What it does not do

Upscayl enlarges and sharpens; it does not generate detail the way diffusion-based upscalers do. Faces at very large scale factors will not gain believable pores or eyelashes the way a generative upscaler produces them — you get a cleaner, sharper interpolation of what was there. There is no in-painting of damage, no denoise/restore toolkit beyond what the models implicitly do, and no editing environment: corrections still happen in your editor of choice afterwards. The desktop-only distribution rules out phones, tablets and any web-based workflow, and because everything is local, a weak machine is a hard ceiling on throughput — someone with a large archive and an old laptop will be waiting a long time compared to a GPU-backed cloud service. It also asks for a lot of disk: the app bundles model weights and ncnn binaries, and install sizes run to hundreds of megabytes. Finally, while the project is actively maintained, it is a small team; release cadence, signed installer availability on every platform, and issue turnaround depend on volunteer capacity. If you need contractual support or an uptime SLA, this is not that, and it was never trying to be.

Verdict

I rate Upscayl 4.1 on the rubric. The core upscaling functionality is genuinely deep for a free tool — model choice, batch runs, configurable scale, local inference (40% pillar, and it is the reason the score is this high); value is literally unmatched because the price is zero (20%); usability is good but hardware-gated and the interface is a tool, not a product (20%); the ecosystem — custom model repository, active GitHub community, documentation — punches above a two-person project (10%); and reliability signals are solid: five years of public development, 50,000+ stars, regular releases, and an AGPL license that guarantees the code stays available (10%). If you have a GPU and a folder of images that need enlarging, install it before you pay anyone for credits. If you need generative detail restoration, batch automation, or a supported business service, a paid cloud upscaler is still the right call.

Sources

• https://github.com/Upscayl/upscayl
• https://upscayl.org
• https://github.com/Upscayl/custom-models
• https://ossinsight.io/analyze/upscayl/upscayl
• https://openmodeldb.info/models/4x-HFA2k
• https://upscale.wiki/wiki/User:Kim2091
• https://github.com/xinntao/Real-ESRGAN
• https://github.com/Upscayl/upscayl/releases
• https://github.com/Upscayl/upscayl/wiki/Guide
• https://www.buymeacoffee.com/fossisthefuture