AI upscale an image

No. The model runs inside this browser tab. The only download is the model file itself (4.6 MB, once, from our model file server, then cached by your browser) and the runtime that runs it (from this site). Neither request carries your picture, its name or anything about it.

It predicts detail. Real-ESRGAN was trained on pairs of small and large photographs, so given a small one it produces the large version it has learned is likely. The result is sharper than interpolation, and some of what it shows was never in front of the lens — which is why every result here is labelled “reconstructed” and a classical enlargement of the same patch is offered beside it.

The model enlarges by exactly four. ×4 gives you that output directly. ×2 runs the same pass and then averages each 2 × 2 block of its result, which halves the size and also calms the model's texture — often the better choice for photographs of people.

Up to 1 megapixel for ×4 and 2 megapixels for ×2, so the result stays inside what a browser tab can hold. It runs on the processor through WebAssembly, one tile of 128 px at a time, and shows its progress; a 0.5-megapixel picture typically takes from several seconds to a minute or two depending on the device. You can cancel at any point.

When the file has to be a faithful record — evidence, a scan for a form, a document where an invented detail would matter — or when the picture is larger than this mode takes. Classical kernels add no detail and invent nothing, and they have no size limit beyond the browser's.