How to upscale an image without uploading it anywhere
You can enlarge and clean up an image entirely inside your browser, with nothing sent to a server.
Most image upscalers ask you to hand over the file first. That is a problem when the image is a client's artwork, an unreleased illustration, or a photograph of a person who never agreed to be on someone else's server. There is another way: run the model in your own browser and never let the file leave the machine.
Why "free online tool" usually means "upload it"
Upscaling runs a neural network over the pixels. That work has to happen somewhere, and the straightforward place to put it is a server you rent. The file goes up, the result comes down, and the service absorbs the compute cost.
The consequence is easy to miss because nothing about it looks unusual: you get a drop zone, a progress bar, and a download link. But between those steps, a copy of your image existed on hardware you do not control, under a retention policy you did not read.
For most snapshots this is a non-issue. For unreleased work, client material, or anything covered by an agreement you signed, it is the whole issue.
Where each approach puts your file
There are three shapes of upscaler, and the difference that matters is not the model they use. It is where the file ends up.
| Approach | Where the work happens | What leaves your machine |
|---|---|---|
| Server-based web tools | A rented server | The image itself |
| Desktop applications | Your own machine | Nothing |
| In-browser processing | Your own machine | Nothing |
waifu2x is the long-standing example of the first row: the processing happens server-side, so the image has to be uploaded, and there is a cap on how large a file you can send. It is free, and its model is tuned for illustration, which is why it stayed the default recommendation for years.
Upscayl is the clearest example of the second row. It runs the model on your own hardware and sends nothing, which is exactly the right design — the only cost is that you have to install it, keep it updated, and be on a machine where you are allowed to install software at all.
In-browser processing sits in the third row for the same reason a desktop app does, minus the install step. That is the entire difference, and it is worth being precise about it rather than pretending it is a difference in privacy.
The alternative: run the model locally
Browsers can now execute neural networks directly on the GPU through WebGPU. The model file is downloaded once, and from then on the computation happens on your own hardware — the same way a desktop application would work, except there is nothing to install.
That inverts the usual trade-off. You are no longer choosing between convenience and control; the convenient path is also the private one, because the file never has a reason to move.
What this costs you in practice
Running locally is not free of consequences. Two of them are worth stating plainly.
You download the model once. About 11.6MB, cached afterwards. Every later image starts immediately.
Your own hardware does the work. Speed depends on your GPU rather than on a datacenter. On an Apple Silicon laptop, a 1024-pixel-square image at x4 finishes in about 7.1 seconds with the photo model, and about 3.9 seconds with the illustration model. There is an upper bound on input size — 12.6MP — above which the browser cannot reliably hold the intermediate buffers.
You need a browser with WebGPU. Recent Chrome and Edge qualify. Without it, the page tells you so instead of quietly falling back to something worse.
Who this actually matters for
If you are enlarging a photo of your own cat, none of this changes your day. The people for whom it changes something specific are narrower than "everyone who cares about privacy":
- Illustrators with unreleased work. A piece that has not been published yet is the one file you cannot casually put on someone else's server, because the leak risk and the reveal date are the same thing.
- Anyone handling client material. If you signed something about how the client's assets are stored and transmitted, an upload step is a clause you now have to read the service's policy against.
- Photographs of other people. The subject agreed to be photographed by you. That is not the same as agreeing to be processed by a third party.
- Work on a locked-down machine. If you cannot install software, a desktop app is not available to you even when it is the better tool.
How to actually do it
- Open Pixmend and drop the image onto the page.
- Pick x2 or x4, and choose the photo or illustration model.
- Compare before and after with the slider, then download the result.
There is no account step because there is no account. There is no upload step because there is no server.
Verify it yourself
This is the part you should not take on faith. Open your browser's developer tools, switch to the Network tab, and process an image. You will see the model file being fetched the first time, and after that, nothing carrying your image. The claim is falsifiable on the spot, which is the only kind of privacy claim worth making.
When something else is the better answer
Three cases where a browser is the wrong place to do this, stated plainly because a comparison that never loses is not a comparison:
You need more than x4. In-browser processing offers x2 and x4. If you specifically need 8x, Upscayl Cloud handles it — it is a paid product at $24.99 per month, with a free tier of 10 credits, and the file goes to their servers, which is the trade you are accepting in exchange for the larger factor.
Your browser has no WebGPU. The computation needs GPU access from the page. If your browser does not provide it, a desktop application is the better answer, and Upscayl is free.
The image is very large. There is a ceiling at 12.6MP, above which the browser cannot reliably hold the intermediate buffers. Beyond that, local processing on a desktop machine is the practical route.
What it does, and what it deliberately does not
Pixmend uses Real-ESRGAN models to restore detail that compression and scaling destroyed. It does not invent detail that was never in the frame. Generative upscalers can produce a sharper-looking result by imagining plausible texture — and in the process, a face can come back as a slightly different person.
For photographs of real people and for artwork that has to stay on-model, that is not a stylistic preference. It is the difference between a repaired image and a new one. Free to use, including commercially.