How to enlarge an image without it turning into mush

You have a photograph that is perfect except that it is small. You need it larger, so you stretch it, and the result is soft and blocky. The natural next question is which software will enlarge it properly.

The honest answer starts with an uncomfortable fact about what enlargement can be.

The detail is not there to recover

An image is a grid of pixels. A picture 400 pixels wide contains exactly 400 pixels of horizontal information, and that is the entire record of what the camera saw. Displaying it at 1600 pixels wide means three quarters of what you are looking at was never captured.

Traditional enlargement handles this by interpolating: for each new pixel it averages the neighbours. The result is smooth and therefore soft, because averaging is precisely how you remove detail. That is the blurriness you get from simply stretching an image, and no setting fixes it, because the information genuinely is not present.

What modern upscaling does differently

Machine-learning upscalers take a different approach. Having seen enormous numbers of images, they have learned what things tend to look like at higher resolution, and they generate plausible detail rather than averaging existing pixels.

The results can be striking. Edges stay crisp, textures reappear, and text becomes legible again. On photographs of everyday subjects the output frequently looks like a genuinely higher-resolution capture.

But read that carefully: it generates plausible detail. It is not recovering what the camera saw, because that information is gone. It is producing something consistent with what it sees, which is a different claim, and the distinction matters more in some situations than others.

• Photographs of common subjects: usually excellent results
• Faces at very low resolution: features may be invented convincingly and wrongly
• Text and numbers: characters can be reconstructed as the wrong characters
• Logos and precise shapes: proportions may drift subtly
• Anything used as evidence: invented detail is a serious problem

The text case is worth dwelling on. An upscaler asked to sharpen a blurred serial number will produce crisp, confident digits. Those digits may not be the ones that were photographed, and nothing in the output signals which are real.

Getting the best result

1. Start from the largest original you have. An upscale of an original always beats an upscale of a compressed copy someone sent you.
2. Avoid upscaling a heavily compressed JPG if you can, since the upscaler will faithfully enlarge the compression artefacts along with the picture.
3. Enlarge in one step to the size you need rather than repeatedly, because each pass builds on the last one's inventions.
4. Do not exceed about four times the original dimensions. Beyond that the output is mostly generated rather than derived.
5. Look at the result at full size before using it, particularly at faces, text and edges.

The question to ask first

Before upscaling anything, check whether you actually need to. A common mistake is enlarging an image far beyond the size it will be displayed. If it appears in a 600 pixel column on a web page, a 600 pixel image is already sufficient and enlarging it to 2400 helps nobody while costing everyone loading time.

Print is where enlargement genuinely matters, because print resolution is much higher than screen. As a rough guide, printing at good quality wants around 300 pixels per inch, so a photograph printed six inches wide wants about 1800 pixels across.

Upscaling makes an image look right. It does not make it true. Know which of those you need before you trust the result.