Rebuild detail when you enlarge, instead of stretching pixels.
Photos, screenshots, old pictures that are too small to use. Up to 1400 pixels on the long edge. Nothing is uploaded.
Enlarging…
The model runs in this browser tab. Your image is never uploaded.
Makes an image larger without the softness that normally comes with it. Enlarging a picture the ordinary way means inventing each new pixel by averaging its neighbours, which is exactly why an enlarged photo looks blurred — the extra pixels carry no extra information.
This runs a small neural network trained on millions of pairs of small and large images, so it knows what a sharp edge looks like after it has been shrunk. Put the shrunken version in and it puts the edge back: a lip contour, the weave of a fabric, the slats of a chair. Against ordinary enlargement of the same photo it is a clear and visible improvement, not a subtle one.
It runs in this browser tab. Nothing is uploaded, there is no limit on how often you use it, and it costs nothing to run — which is only possible because the work happens on your device rather than on a server somebody has to pay for.
Pictures from an early digital camera or saved off a website years ago are often a few hundred pixels across. Three times larger with the edges intact is frequently the difference between usable and not.
Marketplaces set a minimum resolution. Enlarging in an image editor passes the check and looks soft; this passes the check and still looks like a photograph.
Something captured at screen size and now needed on a slide or a poster. Text and hard edges are exactly what ordinary enlargement damages most.
Cropping tightly throws away resolution. Enlarging afterwards gets some of the apparent sharpness back.
Up to 1400 pixels on the long edge — past that it is already large enough that enlarging gains little. Pasting works too.
The picture is processed in squares, with a progress bar. A small image takes a moment; a larger one takes longer, because all of the work is happening on your own processor.
Two, three or four times. Switching between them is instant — the slow part has already happened and does not depend on which you pick.
A PNG, so nothing is lost to compression on the way out. If the file is larger than you need, the Image Compressor will bring it down.
No, and that is a real distinction. The best-known AI upscalers generate plausible new detail — they will give a blurred face convincing eyelashes it never had, which is impressive and is also fiction. This one recovers structure that interpolation smooths away rather than inventing anything. The result is a more modest improvement and a truthful one, which matters if the photo is evidence of something.
Measurably: on a photograph shrunk and then restored, it recovers about a decibel more signal than ordinary bicubic enlargement. That number undersells it — the visible difference in edges and texture is larger than a decibel usually implies, because the improvement lands where the eye is looking.
No. The model is 240KB and runs here, on your device. Nothing leaves the tab, there is no queue, and there is no daily limit — none of which would be true if a server were doing the work.
Because the work is happening on your processor, and it grows with the number of pixels. A 1400 pixel image is already several seconds; something much larger would take minutes for an image that hardly needed enlarging. If yours is bigger than the limit it is probably large enough already.
Beyond that there is genuinely nothing left to recover — you would be stretching the result, and any tool offering ten times is either interpolating or making things up. Three times is where this model actually works; two and four are resampled from it.
No. This adds resolution, not focus. A picture that was out of focus or moved when it was taken is missing the detail entirely, and enlarging gives you a larger version of the same blur. It helps with photos that are small, not photos that are soft.