Upscaling images for printing with AI

What upscaling actually does

Enlarging an image the ordinary way spreads the pixels you already have over a larger area and interpolates between them. Nothing is added; the softness is simply magnified along with everything else.

An AI upscaler works differently. The model has been trained on a very large number of image pairs — the same picture in low and high resolution — and has learned what kinds of detail typically belong at the higher resolution. Given a small image, it generates detail that is plausible for what it sees: it sharpens edges, rebuilds texture, and cleans up compression artefacts.

The word doing the work is "plausible". The model is not recovering information that was lost, because that information is gone. It is inventing something convincing in its place. For photographs, textures and illustrations this is usually indistinguishable from the real thing at print size. For fine text, logos and faces at small scale it can go visibly wrong.

When it helps, and when it will not

A quick way to tell the difference before you spend the processing time.

• Worth doing: the source is genuinely small, and you want to print it larger than its resolution supports. A web-sized image, an old scan, a photo from a phone several generations old, artwork exported too small.
• Worth doing: the image is soft or slightly compressed, and you want the edges to hold up at print size.
• Not worth doing: the image is already sharp and already large enough for your print size. Upscaling will change it without improving anything you would notice.
• Risky: heavy JPEG artefacts, screenshots of text, small logos, and low-resolution faces. The model will confidently invent detail, and here you are likely to notice that it did.

Working out whether you need it

Divide your image width in pixels by the width you want to print in inches. That is the DPI you will get. An 800-pixel-wide image printed 20 cm (7.9 in) wide gives about 100 DPI — noticeably soft in the hand, fine on a wall. Upscaled 2×, the same image gives about 200 DPI at that size, which is comfortably good.

This is also the honest way to decide the factor. 2× is fast and enough for most rescues; 4× is worth it when you are going from something genuinely tiny to something genuinely large, and it takes considerably longer.

Why running it on your own device matters

PrintReal runs the upscaling model in your browser, using Real-CUGAN and Real-ESRGAN through your device's graphics hardware. Nothing is uploaded, which has consequences beyond privacy: no file size cap imposed by a server, no queue behind other people's jobs, no watermark, and no limit on how many images you run.

The trade is that the work is done by your machine. On a recent laptop or phone a 2× upscale takes seconds to a minute; on older hardware, or at 4× on a large image, it takes longer, and the tool will feel like it is thinking — because it is. The model has to be downloaded once, after which it is cached.

Then print it at the size you meant

Upscaling gives you pixels; it does not decide how big the result appears on paper. When it finishes you can send the image straight into the Real size tool to print it at exact dimensions, or into the Poster tool to tile it across several sheets — which is the usual reason for upscaling in the first place.