Upscale an Old Photo — Free, Fast & Private
An old photograph is usually two problems wearing one coat, and separating them is most of the work. There is a resolution problem — the file is small, because a 6x4 print scanned casually or snapped on a phone gives you a couple of megapixels at best. And there is a condition problem — the print has faded, the emulsion has gone soft, the faces have lost definition, there is a crease across one corner and a coffee ring in the sky. Enlargement addresses only the first. Point a super-resolution model at an unrepaired scan and you get a larger copy of every flaw, rendered with more conviction than before. The sequence on this page exists because each of these tools is good at exactly one thing, and the order they run in decides how much of the picture survives.
Step one is capture, and it caps everything downstream. No model can recover what the capture never collected, so this is the step worth spending an afternoon on. A flatbed scanner beats a phone for three reasons that have nothing to do with megapixels: it holds the print flat so the whole frame is in focus, it lights evenly so there is no gradient across the picture, and it samples at a genuine optical resolution rather than an interpolated one. Scan at the scanner true optical setting; at 600 ppi a 4x6 print yields roughly 2400x3600 pixels, which is enough to reprint at the original size before any enlargement at all. If the only option is a phone, work near a window with the flash off, lay the print on a matt surface, shoot square-on so the rectangle stays a rectangle, and fill the frame. Curl the print flat under glass only if you can avoid the reflection, since glare removes information as thoroughly as a scratch does.
Step two is repair, and it happens before any enlargement because those tools hand back the same size. The Cloud HD photo restoration tool runs RestoreFormer++ on our GPU and brings faces back in old, blurry, faded pictures — it is Pro, it accepts up to 25 MB and 24 megapixels, and it returns an image the same size as the one it was given. It does not repair tears, stains, scratches or missing areas; nothing here does, except the Cloud HD object remover, where you paint a brush mask over a small region with predictable surroundings and the scene is filled back in. If the source came off a scanner, restore scanned photos covers the specifics of that path. Because those tools return the same dimensions, running them first costs nothing and gives the upscaler a clean picture to work from — while running them second would ask them to process four or sixteen times as many pixels for no additional detail, and can push a file into their 24 megapixel ceiling.
Colour, if you want it, belongs in the middle. DDColor produces plausible colour for a monochrome photograph, and plausible is the operative word: nothing in a black-and-white negative records which of several possible hues a coat or a wall actually was, so the model gives you a convincing guess rather than a recovery. Run it after the face repair and before the enlargement, for the same reason as above — it returns a PNG at the input size, and colouring a small clean image then enlarging it carries the colour up with everything else. Colorize photo for printing covers that step where the destination is paper. Both restoration and colour are Cloud HD tools: Pro only, on our server, off until you ask for them.
Step three is the enlargement, which is the free part. The upscaler on this page runs realesr-general-x4v3 in your own browser — 4.9 MB, checksum-verified, executed through WebGPU or WebAssembly in 512px tiles so memory stays flat. Choose 4x for a small scan and 2x when the capture was already generous; both run the same network, with 2x downsampling on the way out for a subtler result. The free tier caps input at 1080px on the long edge, produces up to 4320px at 4x, and stamps a small "minipx.com" watermark in a corner; it includes 2 AI runs shared across the three on-device AI tools before a cooldown begins. Pro removes the watermark, makes on-device runs unlimited and raises the caps to 2048px at 4x and 4096px at 2x, ceilinged at 8192px of output. Crop the scanner border and any damaged edges first in the cropper — reconstructing a strip of album card is an expensive way to make nothing. The tiling and the model itself are described on the on-device upscaler page.
Cloud HD as the last resort on a difficult scan. When the capture is small and the print was soft to begin with, Pro can switch on Cloud HD beside the run button and send that one photo to our GPU, where a JPEG pre-clean pass runs before HAT-L rebuilds it at 4x. Intake is 25 MB and 4 megapixels, output reaches up to 16× the pixels of the input — a 4 MP photo comes back at about 64 MP, and a full 4 megapixel source takes four to five minutes — long, and the tool says so before you commit. One run spends one of the 10 HD runs in your month, inside 150 runs on Pro Monthly or 200 on Pro Annual; a failed run costs nothing. For a print destined for a frame, that is usually the right place to spend it, and upscale a photo for printing has the paper arithmetic to go with it. The AI image upscaler hub documents both paths in full.
Keep the scan, and be clear with yourself about what is real. Archive the untouched capture separately from every processed version. The original is the record; the restored, coloured, enlarged file is a presentation of it, and a good one — but its fine texture is reconstruction, its colour on a monochrome original is inference, and neither should be relied on for genealogy, identification or any other documentary purpose. On the machinery: the enlargement runs on your own device and nothing image-shaped is transmitted, while the restoration, colour and Cloud HD steps are the exception — a Pro-only path where that one photo is uploaded encrypted, processed in the United States on Cloudflare R2 and a Modal GPU, deleted when the run finishes, and in all cases within 24 hours, and never used to train anything.