When your browser cannot run Pixmend
Pixmend needs WebGPU and has no slow fallback, so the fix is the browser or the machine, not a setting in the page.
If the page tells you that this browser cannot run Pixmend, nothing is broken and there is no setting inside the page to change. Pixmend runs the upscaling model on your own GPU through WebGPU, and if the browser cannot reach the GPU there is no second path for it to take. Open Pixmend in a browser that supports WebGPU and it works immediately; this page is about how to get to one.
Why there is no slower fallback
Most web tools that need a GPU keep a CPU version around for the browsers that lack one. Pixmend does not, and that is a decision rather than an omission.
Running this model on a CPU takes minutes per image instead of seconds. A result that arrives after several minutes is not a result anyone waits for, and it would set the impression of the tool for everyone who happened to land on the slow path. Telling you plainly that this browser will not work is more useful than making you find out slowly.
The obvious workaround does not exist either. Choosing x2 instead of x4 does not reduce the work: x2 is produced by running the same x4 model and scaling the output down, so it costs the same. There is no lighter setting hiding behind the notice.
What is actually being checked
The page runs a short sequence before it lets you start, and every step has to pass:
| Step | Why it can fail |
|---|---|
| Browser APIs are present | The processing runs in a background worker and caches models locally, which needs a handful of modern APIs. Very old browsers are missing them. |
| A WebGPU interface exists | The browser build has WebGPU compiled in and enabled. |
| A GPU adapter can be obtained | The browser will refuse if hardware acceleration is off, or if it blocks your graphics driver as unstable. |
| A GPU device can be created | An adapter that exists can still fail to hand out a device, so this is checked rather than assumed. |
Only after all of these does the page consider the environment usable. It is checked this way on purpose, because an environment that reports WebGPU support and then fails once your image is loaded is the worst possible outcome.
Things worth trying
Update the browser. WebGPU support has been arriving gradually, and an install that has not restarted in months is the single most common cause.
Turn hardware acceleration back on. It lives in the browser's settings, and it is frequently switched off to fix an unrelated problem and then forgotten. With it off, no browser can reach the GPU, however new it is.
Update your graphics driver. Browsers keep a list of driver versions known to crash and refuse to hand out a GPU on them. Updating the driver moves you off it.
Check where you are running. A remote desktop session, a virtual machine, or a locked-down work profile often has no usable GPU to give out. The same browser on the physical machine may be fine.
Try another browser before concluding your machine cannot do it. Chrome, Edge, or a recent Safari are the ones to try. This costs a minute and settles whether the problem is the browser or the hardware underneath it.
When it starts but then fails
A different message appears if the check passed and the engine failed afterwards, and it carries the underlying error. That usually means the model files or the runtime did not finish downloading — a network interruption, or a proxy that filters them. Reloading the page retries the download.
The first run fetches about 11.6MB in total: the models and the runtime. After that they are cached on your device, so later runs start immediately without touching the network again.
If processing is refused because of the size of the image, that is a separate limit rather than a support problem. Inputs are capped at 12.6MP, since above that the memory required exceeds what a browser tab can reasonably hold.
What you get when it does work
On a machine with a working GPU, a 1024×1024 image at x4 takes about 7.1 seconds with the photo model and about 3.9 seconds with the illustration model. There is no account, no install, and no upload — you can confirm the last one yourself by opening DevTools, switching to the Network tab, and watching it stay silent while your image is processed.