Upscale an image
Enlarge a picture with an AI model running on your own machine. Slow, and worth knowing why.
- Runs on your device — the file is never sent anywhere
- No sign-up, no watermarks, no daily quotas
- Free, with no limit on how often you use it
How it works
Enlarging a picture is either fast or good, and it is worth being clear about which you are choosing. This offers both, because for many jobs the instant option is genuinely the right one.
It is slow, and here is why
The AI option downloads a 21 MB model once, then runs a neural network over your picture on your own processor. Expect anywhere from several seconds to well over a minute depending on the image and the device. There is no server doing this faster in the background — that is the trade for the picture never leaving your machine. Pictures above about 1200 × 1200 are refused outright, because the browser runs out of memory.
What the AI actually does
It does not recover detail that was lost. It predicts what plausibly belongs between the pixels, based on millions of images it was trained on. On photographs and textures the result genuinely looks sharper. On text, faces and logos it can invent details that are subtly wrong, which matters if the picture is evidence of something.
The instant option is often better
High-quality resampling enlarges in steps rather than one jump, which avoids the blockiness of a naive resize. It adds no detail and makes no guesses. For printing something slightly larger, or for meeting a minimum dimension on a form, it is the right choice and takes no time at all.
The model
Swin2SR, published under the Apache 2.0 licence, run with ONNX Runtime in your browser.
Upscale an image — questions
Why is it so slow?
A neural network is running on your own processor rather than a server. That is what keeps the picture private.
My picture was refused for being too large — why?
The AI upscaler runs out of browser memory above roughly 1200 × 1200. Use the instant resize for larger pictures.
Can it recover detail from a blurry photo?
No. It predicts plausible detail rather than recovering real detail, and can get text and faces subtly wrong.
Is my image uploaded?
No. The model is downloaded to you; the picture never leaves your device.