Small picture, bigger print

When a picture is enlarged, no program can look at what was never recorded. A 640-pixel photo contains 307,200 coloured squares and nothing else; a 2560-pixel version needs 4.9 million. The extra 4.6 million squares have to be invented from the ones that exist. That single fact decides everything about upscaling: the question is never whether detail is invented, only how convincingly it is done.

There are two families of methods. The first fills the new pixels with an average of their neighbours — bilinear or bicubic interpolation. It is fast and it is why a stretched picture looks soft: every edge is smoothed across two or three pixels, so contrast disappears exactly where the eye looks first. The second family tries to rebuild the edge instead of smoothing it: the direction of a line is estimated, and new pixels follow that direction rather than averaging across it. The tool on this page uses the second approach, in the modest form that a browser can afford.

Why the enlargement happens in steps

Doubling a picture is well behaved: every source pixel becomes a two-by-two block, and the filter knows exactly where the original samples were. Quadrupling in a single call forces the filter to guess across four times the distance, which blurs strongly and sometimes leaves visible squares around hard edges. Two doublings cost a little more memory, but each one has a small, predictable error, and the sharpening pass after each step can work on edges that are still where they belong.

Between the steps the tool removes noise. Grain is the enemy of enlargement: a sharpening filter cannot tell a dust speck from an eyelash, so it amplifies both. A gentle box filter first, then a stronger edge rebuild, produces a picture that is clean without looking plastic.

What the detail slider really controls

The slider sets the strength of an unsharp mask: the picture is compared with a slightly blurred copy of itself, and the difference is added back to the original. At low values the result is soft but natural. Around 50% the edges of buildings, letters and product shapes look right. Above 70% the halos around high-contrast edges become visible when you zoom in, and skin starts to look tired. For photographs, stay near the middle; for screenshots and line art, go higher.

The honest limits

Three things do not survive. Faces lose micro-texture: pores, fine hair and stubble cannot come back, so a face enlarged 4× reads well at arm's length and poorly in a close crop. Small print loses letterforms — a 12-pixel-tall word becomes a 48-pixel-tall word that is still only 12 pixels of information. And out-of-focus areas stay out of focus; blur has no edges to rebuild.

The right way to use an upscaler is therefore narrow and useful: take a picture that is sharp but too small, enlarge it by the factor the destination needs, choose the matching mode and export. If the destination needs 300 dpi at A3 from a phone photo, no upscaler is the answer — a new photograph is.

A short workflow that avoids trouble

• Crop and colour-correct first, at the original size, then enlarge once at the end.
• Keep the original file. The enlarged version should always be a new file with the factor in its name.
• Export PNG for logos, diagrams and text; JPG or WebP for photographs.
• Check the result at 100% on the screen where it will be seen, not at thumbnail size.

One more practical note: an image that was already enlarged by another program carries interpolation artefacts. Enlarging it again doubles those artefacts rather than the subject, so always work from the biggest original you can find — the camera roll, the client's folder, the design file.

Print size, honestly calculated

Printers do not have pixels, they have dots per inch. A 300 dpi print of a 2× enlargement from a 640-pixel source is 1280 pixels across, which is 4.3 inches of paper, roughly a postcard. The same file at A4 would be 155 dpi: acceptable for a photo held in the hand, soft for a poster read from a metre away. Doing that arithmetic before exporting saves a reprint.

The reverse calculation is the useful one. Decide the physical size, multiply by the dpi the printer asks for, and you have the pixel width you need. If the number is more than four times the source, the honest answer is a better source, not a bigger slider setting. Enlarging is a repair, and repairs have a budget.