Upscale an image

Drop it on this page, choose how much bigger, and press Enlarge. Two engines are offered and the choice matters: the AI model redraws detail it has learned from photographs, which is what you want on a photo, a face or anything textured; the Lanczos engine invents nothing at all, which is what you want on a logo, a screenshot or a diagram. Both run in this browser tab and neither one uploads the picture.

It is real, and it is not sharpening. The model here is ESRGAN-slim, a residual dense network trained on thousands of pairs of large and small photographs, so it has learned what a pore, a brick, a blade of grass and a strand of hair look like at the larger size and draws that. Sharpening only exaggerates the edges that are already in the file. The difference is easiest to see at 100 percent on hair or fabric: sharpening gives you a crisper version of the same mush, and the model gives you texture.

No, and nothing can. A model draws what pictures like yours usually look like, so on a face it produces a plausible face rather than the real one, and on a number plate it produces plausible characters. That is fine for a poster and wrong for anything that will be relied on. If the answer has to be true rather than convincing, use the Lanczos engine, which can only rearrange pixels that were actually photographed.

The model adds detail from what it has learned. Lanczos adds nothing: it reconstructs the picture at a larger size using a windowed sinc filter, in linear light with the transparency premultiplied, then compares its own answer with your original by shrinking it back down and pushes the difference back in. On a photograph the model usually wins. On flat artwork, text, line drawings and screenshots the model can invent texture along an edge that is meant to be clean, and Lanczos is the better answer.

Two, three, four, six or eight times. Six and eight are done in two passes with the models that were trained for each step, because a model trained at four times has never seen an eight times job. There is a ceiling: an enlargement over 40 megapixels or 16,384 pixels on a side is more than a browser tab can hold, so the picture is enlarged by a smaller whole factor instead and the page says so rather than crashing.

Yes, and it takes some care over it. The model has three channels and cannot see alpha, so the transparent area is filled with the colour beside it before the model runs, which is what stops the dark rim a network draws when it sees black behind a cutout. The alpha channel itself is enlarged separately with the Lanczos path and put back afterwards, so a cutout comes out of here bigger with its soft edge intact.

Because the model is genuinely running on your device. It goes over the picture in overlapping tiles and each tile is a pass through the network on your graphics chip, so a twelve megapixel photo at four times is hundreds of tiles. The page shows how many are done and can be stopped at any point. If your device has no graphics engine the model can use, the Lanczos engine does the work instead and says so.

No. The model's weights are downloaded to your device, which is the opposite direction: about 900 KB the first time you use a scale, served by this site and then cached. The picture itself is read, enlarged and saved inside this browser tab, and there is no address on this site that could receive one.

An old photograph that is too small to print, a product shot a supplier sent at 600 pixels, a logo that has to go on a banner, a screenshot for a presentation, a thumbnail you no longer have the original of. For printing, work back from the size on paper: 300 dots per inch means an A4 page wants about 2,480 by 3,508 pixels.