Free AI Image Upscalers for Web Assets (2026): 7 Options, Local First

For a free AI upscaler that keeps your images on your own machine, install Upscayl: it is open source, runs on Windows, macOS and Linux, and does the work locally. If you can't install anything, the in-browser version of waifu2x does the work in the tab. Web services like Bigjpg and Let's Enhance are fine for a few images, but they upload your file and cap the free tier.

Seven tools below, in three groups: desktop and open source, in-browser, and hosted web services. Every limit, license and feature was checked against each vendor's own page or GitHub README on 11 November 2026. Where a page doesn't publish something, we say so.

What an upscaler can and can't do for a website

An AI upscaler guesses the missing pixels. That works well for a photo or product shot taken at low resolution, an old screenshot you need at 2x for high-density screens, or a small hero image that looks soft on a large monitor. It is a guess, though: fine text, thin lines and brand marks are where the guesses show. More on that in when to use vector instead.

For each tool we record five things: the free limit, the maximum scale or output size, whether it runs locally or uploads, the license if it is open source, and the platforms.

Desktop and open source

All three run on your own hardware, with no account, no quota and no watermark. The cost is a download and, for good speed, a reasonable GPU.

Upscayl

Upscayl is a desktop app: pick an image, pick a model, choose a scale. The site lists upscaling "up to 16x", and the download page says the desktop app does all processing on your machine with no internet connection needed. It is AGPL-3.0.

Platforms: Windows 10 and later, macOS 12 and later, Linux (Flathub, AppImage, Snap, AUR, DEB, RPM). On a Mac, brew install --cask upscayl or the Mac App Store.
The catch: it needs a Vulkan-compatible GPU, and the README warns that many integrated GPUs do not work.
Note: the project now also sells Upscayl Cloud; the desktop app stays free and open source.

It suits most people who want the job done without a command line.

Real-ESRGAN

Real-ESRGAN is the model family many upscaling apps are built on, and you can run it directly. It is BSD-3-Clause.

Portable executables for Windows, Linux and macOS use NCNN and Vulkan and work on Intel, AMD and Nvidia GPUs, with no CUDA or PyTorch install. They take a scale of 2, 3 or 4.
The Python version (pip install basicsr plus the repo's requirements) adds --outscale for any output size, --face_enhance (GFPGAN) for faces, and RealESRGAN_x4plus_anime_6B, a smaller model for illustrations.
Tiling (-t) splits big images to fit in memory; the README warns it can leave visible block seams.

It suits a script that upscales a folder of assets in a build step or a batch job.

chaiNNer

chaiNNer began as an upscaling app and is now a node-based image pipeline: load, upscale, resize, convert, save, all wired as a graph. It is GPL-3.0 and runs on Windows, macOS and Linux.

Models: PyTorch (through Spandrel), NCNN, ONNX and TensorRT. You bring the models; the README points to OpenModelDB for community upscaling models.
Hardware: CUDA and TensorRT on Nvidia, ROCm on Linux or NCNN for AMD, MPS on Apple Silicon, NCNN on Intel, and a CPU fallback for every framework.

It suits someone who wants to pick a model per job, say one for photos and one for UI screenshots, and chain the upscale with a resize and export.

In the browser, no upload

Two tools whose own pages say the image is processed in the tab. Each downloads a model on first use, then runs on WebGPU or WebAssembly.

waifu2x (in-browser)

unlimited.waifu2x.net calls itself an experimental in-browser version of waifu2x, and its privacy policy says images are processed in your browser and "never sent to remote servers".

Scale: 1x (denoise only), 2x or 4x.
Models: several art models (including one for scans) and one photo model, with five denoise levels.
Backend: WebGPU or WebAssembly, chosen automatically. No size limit is published.
License: the upstream project, nunif, is MIT.

Despite the anime roots, the art models handle flat graphics and illustrations well, which is close to what a lot of web assets are.

Pixconvert AI Image Upscaler

Pixconvert's upscaler runs Real-ESRGAN in the browser; the page says the image stays on your device and only usage stats are sent. No account, no daily cap.

Scale: 2x or 4x. The first run downloads about 34 MB per scale.
Size limits: JPG, PNG or WebP up to 50 MB, but anything over 800 px on the long side is shrunk to 800 px first, so output tops out at 1600 px (2x) or 3200 px (4x).
Output: always PNG.

The 800 px cap makes it a tool for small assets: thumbnails, avatars, product shots.

Hosted web services with a free tier

These upload your image and process it on their servers. Fine for a public product photo; think twice for an unreleased screenshot or anything under NDA.

Bigjpg

Bigjpg has a free plan of 20 images a month, up to 4x, with a 5 MB upload limit. Free jobs run on a shared, slower server, with no batch mode. The pricing table doesn't mention a watermark. Paid plans start at $6 for two months (500 images a month) and go up to 16x with 50 MB uploads.

Let's Enhance

Let's Enhance gives 10 free credits when you sign up, no card needed. Free results carry a watermark, and the free plan upscales to 8 megapixels. Paid plans start at $9 a month billed annually ($12 monthly), with output up to 256 megapixels on Starter. Ten watermarked images makes it a way to try it, not a free tool.

Side by side

Tool	Free limit	Max scale / size	Local or upload	License	Platforms
Upscayl	Unlimited	Up to 16x	Local	AGPL-3.0	Windows, macOS, Linux
Real-ESRGAN	Unlimited	2–4x (exe); any size (Python)	Local	BSD-3-Clause	Windows, macOS, Linux
chaiNNer	Unlimited	Depends on model	Local	GPL-3.0	Windows, macOS, Linux
waifu2x in-browser	Unlimited	2x or 4x	In browser	MIT (nunif)	Any WebGPU/WASM browser
Pixconvert	No daily cap	2x/4x, max 3200 px	In browser	Not published	Any modern browser
Bigjpg	20 images/month, 5 MB	4x free (16x paid)	Upload	Proprietary	Web
Let's Enhance	10 credits once, watermark	8 MP free	Upload	Proprietary	Web

When to upscale and when to use SVG

Upscaling is for raster content: photos, product shots, screenshots, textures. For anything drawn, a vector beats any upscaler.

Logos and icons: find or ask for the SVG. A vector is sharp at every size, is usually smaller than the PNG, and can be optimized with SVGOMG. If the brand only gave you a PNG, check their press kit first; our free SVG icon and logo resources list where official marks live.
Screenshots of a UI you own: take the screenshot again at a higher device pixel ratio (a browser's device emulation or a 2x display). A real capture beats a reconstruction.
Charts and diagrams: export them from the source again, as SVG if you can.
OG images and social cards: rebuild them at the target size from text and a logo rather than upscaling an old one. The OG image generator does this in the browser.

Text and logo artifacts

Upscalers are trained mostly on photos and illustrations, and text is where they slip. Expect letters that wobble or merge, thin strokes that thicken or vanish, smooth gradients that pick up texture, and brand colours that drift a little. At 4x, small text can turn into shapes that look like letters but aren't.

Three habits help. Use the smallest scale that does the job: 2x for a high-density version of a small asset is safer than 4x. Pick a model meant for the content, an illustration model for flat graphics and a photo model for photos. And zoom to 100% on any text before you ship it; if it reads wrong, redo that part as live HTML text or a vector.

How to choose

Private or unreleased images, regular use: Upscayl, or Real-ESRGAN if you want it in a script.
No GPU that Upscayl supports: chaiNNer with a CPU or Apple Silicon backend, or one of the in-browser tools.
No install allowed: waifu2x in-browser for flat graphics, Pixconvert for small photos.
A few public images, quickly: Bigjpg's free plan, within 20 a month and 5 MB.
After any of them: the output is usually a large PNG. Convert and compress it before it goes on a page; compressing an image to 100 KB explains how format and quality set the size.

FAQ

Is there a free AI image upscaler with no watermark?

Yes. Upscayl, Real-ESRGAN, chaiNNer and the in-browser waifu2x are free with no watermark and no quota. Of the hosted services here, Let's Enhance watermarks free results; Bigjpg's pricing page doesn't mention one.

Which upscalers don't upload my image?

The desktop tools (Upscayl, Real-ESRGAN, chaiNNer) run on your machine. The in-browser waifu2x and Pixconvert say they process in the tab, which you can check in the Network tab of DevTools. Bigjpg and Let's Enhance process on their servers.

Can I upscale a logo with AI?

You can, but expect soft edges and distorted lettering. For a logo, an SVG from the brand is almost always the better answer, and smaller too. Upscale only when no vector exists, and check every letter at 100% zoom.

How much should I upscale for a website?

Usually 2x: enough for high-density screens without inventing too much detail. Then resize to the largest size the layout actually displays and compress. Shipping a 4x image that shows at 600 px only makes the page heavier.

To finish the job, the toolbox has a free image compressor and image converter that run in the browser. For more design resources, browse the design directory.