Enlarge an image without it going blurry

What this does, and the thing it cannot do

This makes your image bigger and keeps the edges clean. It cannot add detail that the camera never recorded — and neither can anything else, including every tool that calls itself an AI upscaler.

That distinction is worth being plain about, because the expectation it sets is the source of nearly every complaint in this category. Enlarging invents pixels between the ones you have. A good algorithm invents plausible ones and the result looks sharp; a bad one averages its neighbours and the result looks like a smear. Neither recovers a licence plate, a face at the back of a crowd, or text that was four pixels tall.

Lanczos, and why it is not a compromise

Enlarging here uses Lanczos resampling over an eight-by-eight neighbourhood. It is the sharpest of the classical filters and the difference against the bicubic most software defaults to is visible on any hard edge.

It is also worth knowing how close that gets you. OpenCV's own super-resolution benchmark ranks plain Lanczos below bicubic on one common quality metric, and neural models above both — but the gap for a 2× enlargement of a clean, sharp photograph is much smaller than the marketing suggests. Where a model genuinely wins is on rough input: a compressed JPEG, a screenshot, a phone photo taken indoors.

The order of operations is the part that matters

Denoising happens before the enlargement and sharpening after, and reversing either is a mistake.

Denoise after enlarging and you are smoothing interpolated pixels that were never real detail — you smear the guesses and lose the genuine texture along with them. Sharpen before enlarging and you amplify the noise first, then blur it across four times the area.

Sharpening that knows what to leave alone

Enlarging always softens edges a little, so a small amount of sharpening puts the crispness back. Applied naively it also amplifies sensor noise in exactly the places that should stay smooth — sky, skin, paper — which is what makes an over-processed photo look gritty.

So the sharpening has a threshold: it only acts where there is a real edge. And rather than making you guess the number, it measures the noise in your own image and sets the threshold to suit. That is not cosmetic. Measured on a photo with a noise level of 9 grey levels, a fixed threshold of 8 let flat-area noise more than double, from 6.0 to 14.4; measuring the image and choosing 22 held it at 7.2. On a clean image it picks 3, so fine detail is still sharpened properly. The result tells you which figure it chose.

When to use the noise clean-up

Turn it on for a phone photo taken in low light, a heavily compressed JPEG, or a scan of an old print. Leave it at zero for anything clean — on a sharp original it removes real texture and gains you nothing.

Choosing a size

Two, three or four times, or an exact width when a form or a print shop has given you a number. There is a 40-megapixel ceiling on the output, which is generous — a 4× enlargement of a 12-megapixel photo is 192 megapixels and would be a 500 MB PNG that no software will thank you for.

Going the other way is a different operation with a different filter, so resize image handles shrinking and this refuses it rather than doing it badly.

What would actually recover detail

Nothing, if it is not in the file. If you still have the original — the camera file, the un-cropped version, the PDF the screenshot came from — go back to it. A single re-export from the source beats any amount of processing on a copy.