Upscale an Image 2× or 4×

The fast tier enlarges in repeated 2x steps with adaptive sharpening applied at each stage - a pipeline that produces clean, natural enlargements with none of the stair-stepping of naive scaling. What it deliberately does not do is invent detail. Resampling can only redistribute the pixels you gave it, so this page will never claim the fast mode "enhances" anything: it is the right choice when you need bigger dimensions with no surprises, on any hardware, in about a second.

The ML tier runs Real-CUGAN, an MIT-licensed super-resolution network that is remarkably compact at 2.6 MB - smaller than many of the photos it enlarges. Instead of interpolating between existing pixels, it predicts plausible fine texture learned from training data, which is why fabric weave, foliage and hair come out convincingly detailed rather than merely smooth. The word to hold onto is plausible: the network makes an educated guess at what detail belonged there, not a recovery of what the camera saw.

Neural upscaling is memory-hungry in ways phones punish, so the ML mode is gated behind a quick benchmark of available memory on mobile devices. If your device fails the check, the option is disabled with an explanation instead of letting the tab crash halfway through a 4x pass - an honest trade most upscaler pages avoid by silently doing the work on a server instead.

When is upscaling actually worth it? Printing at a larger size, recovering a heavily cropped composition, and meeting a platform's minimum-dimension rule are the genuine cases. What no upscaler can legitimately do is CSI-style recovery: text and faces that are unreadable at the original size do not become trustworthy at 4x, because any legibility the ML mode adds is synthesized, not restored.