Upscale an image 4×

No, and this is the single most important thing to understand about every AI upscaler. Detail that was never captured is gone. What the model does is generate new detail that is plausible given what surrounds it — leaves that look like leaves, bricks that look like bricks. On a photograph that reads as sharper and better. On anything where accuracy matters it is a liability: a blurred licence plate, a fuzzy serial number or an unreadable signature will come back looking crisp and confident, and the characters may simply be wrong. Never treat an upscaled image as evidence of anything.

A plain resize interpolates: it averages the pixels it has to fill the new ones, which is mathematically honest and always slightly soft. Measured against a known original, plain resizing actually scores BETTER on pixel-accuracy than this does — around 30 dB against 28 dB in our own test. It still looks worse, clearly so, because our eyes reward believable texture more than they reward numerical closeness. If you want faithful, use Resize Image; if you want it to look good, use this.

Because of what comes out the other end. Four times each side is sixteen times the pixels, so a 2 MP input already produces a 32 MP image, around 6400 by 5000. A 12 MP phone photo would come out at 192 MP, which most software will not even open and nobody actually wants. Upscaling is for pictures that are too SMALL — thumbnails, old web images, cropped details. If yours is already large, resizing down is far more likely to be what you need.

Roughly half a minute for an image near the limit, less for smaller ones. The work is done by our server rather than your browser, unlike the PDF tools here, because the alternative would mean sending you nearly 18 MB of runtime before the first image — and then running it single-threaded on your phone, which is slower still.

Less well than on photographs, and sometimes badly. The model was trained mostly on natural images, so it treats sharp letterforms as texture to embellish. Small text tends to come back looking confident and subtly deformed. For a screenshot you are usually better off retaking it at a higher resolution, or using Image to Text to pull the words out and set them again.

No. It is uploaded, processed, and both your file and the result are deleted automatically — uploads within two hours, finished downloads within twenty-four. Nothing is used to train anything; the model here is fixed and does not learn from what passes through it.

Files up to 15 MB are accepted. Every upload is stored under a random name, never served from a public folder, and deleted 24 hours after the job starts — the finished file goes at the same moment. We never share them and there is no account they could be tied to. The one thing we do is compare the original against the result in our own admin area, to be sure the compression has not softened the picture.