Upscale a small image
Four times bigger, reconstructed rather than stretched.
Your file
Drop a file here, or click to choose
It stays on this device. Nothing is sent anywhere.
Settings
Not the same as resizing
Scaling an image up in an editor spreads the pixels you already have over more space, which is why it goes soft. This runs a super-resolution model that was trained on millions of pairs of small and large images, so it produces plausible detail at the edges of things: text stays sharper, edges stay hard, and the result looks like a photograph rather than a blur.
It is still inventing. The detail it adds is what a model thinks belongs there, not what was in front of the camera. For a small logo, an avatar, an old thumbnail or a product shot, that is exactly what you want. As evidence of what something actually looked like, it is worthless.
Why it is capped at small pictures
The model is a transformer that looks at every pixel, and this site runs it on a single CPU thread inside your browser. The cost climbs faster than the pixel count: measurements on this build put a 64 by 64 input at a few seconds and a 256 by 256 input at around a minute. A twelve-megapixel phone photo would take hours.
So the input is capped at roughly 384 by 384, which gives you a 1536 by 1536 result in a minute or two. Larger images are refused outright rather than left running until you give up. If you want a big photo sharpened rather than enlarged, that is a different job and this is not the tool.
What it is good and bad at
Good: photographs, product shots, faces, screenshots of text, old low-resolution images. Poor: images that are already heavily compressed, where it will faithfully enlarge the compression artefacts too. Run it on the cleanest version of the file you have rather than a JPEG that has been through three chat apps.
Common questions
It refused my image.
It is bigger than the cap. This runs on one thread in a browser tab, and above roughly 384 by 384 the wait becomes unreasonable. Crop to the part you actually need and upscale that.
Can I upscale by 2x instead of 4x?
The model is fixed at four times. If you want less, upscale and then resize the result down to whatever you need, which still looks better than enlarging directly.
Is the detail it adds real?
No. It is a plausible reconstruction learned from other images. It looks right, and it is not evidence of anything.
Why is it faster than I expected, or slower?
It scales with the pixel count and with your CPU. The weights are quantised to 8-bit, which on this runtime is both smaller and quicker, and was measured as visually identical to the full-precision model.