If nothing is gained, why does a larger version sometimes look better?

Because a display is doing its own scaling, and doing it worse. A 400 pixel image shown in an 800 pixel frame is being interpolated by the browser with a fixed, cheap filter. Producing the 800 pixel version deliberately, with a kernel you chose, can look better than letting the page do it — the information content is identical, the presentation is not.

Does a higher-numbered kernel recover more?

No. Lanczos, bicubic and bilinear differ in how they distribute the error, not in how much information they have. A windowed sinc makes edges look crisper by overshooting at them, which reads as sharper and is not more faithful. There is no ordering of these three in which one recovers detail another cannot.

What about the tools that clearly do add detail?

They are not interpolating. A model trained on millions of photographs predicts what plausibly sat between the samples, and produces a result that is convincing rather than recorded. That is a legitimate and often desirable thing to want, and it is a different operation with a different set of honesty problems, which is why it lives on a different site.

Why upscaling loses detail

A kernel computes values between the samples it has. Where a sample was never taken, there is nothing to compute from — and no arrangement of weights changes that.

What a kernel is doing when it goes up

Enlarging a grid means asking for the colour at positions that were never measured. A source pixel is a measurement of the light that fell on one photosite; between two of them there is no measurement at all, only whatever the scene happened to contain. A kernel answers by taking a weighted average of the nearby measurements, and the weights are what distinguishes one kernel from another.

That answer is always a smooth blend of things already known. It follows directly that the output contains no frequency the input did not — the fine texture on a brick, the thread in a fabric, the boundary between two letters that were three pixels apart. Where those things were lost in the original sampling, the new grid describes their absence at higher precision.

The result is not damaged. Nothing is subtracted. It is that the ratio of recorded detail to stored pixels falls, and that ratio is what the eye responds to at a given print size.

The arithmetic, on a real case

Take a 600 by 400 pixel image and ask for a 20 cm wide print at 300 dpi. The target grid is 2362 pixels across. The source had 600.

600 x 400 taken to a 20 cm print at 300 dpi
	Source	After resampling
Pixels across	600	2362
Recorded samples across	600	600
Computed samples across	0	1762
Stored megapixels	0.24	3.72
Effective resolution at 20 cm	76 ppi	76 ppi

The bottom row is the one that decides how the print looks. Of the 3.72 million pixels in the exported file, 240,000 carry a recorded sample and the rest were computed from them — and the density of recorded detail on the paper has not moved at all. A viewer at reading distance sees a 76 ppi print, because that is what it is.

Softness, and the thing that is often mistaken for it

An enlarged image usually reads as soft, and the instinct is to sharpen it. Sharpening raises local contrast at edges; it does not restore the samples that would have made the edge crisp, and pushed far enough it produces a bright rim beside every boundary that is more distracting than the softness was.

There is one exception worth naming, and it is the reverse case: a downscale genuinely does lose acuity that sharpening can legitimately restore, because the information was present and the filter attenuated it. That is a different operation from rescuing an enlargement, and confusing the two is how images end up over-processed.

When you actually need a bigger image with more in it

Sometimes the requirement is real: the only copy of a photograph is small, and it has to fill a page. Interpolation will not meet that requirement, and this site will keep saying so while it produces the file. What meets it is reconstruction — a model that has learned what skin, brickwork, foliage and text look like at higher magnification and predicts plausible detail rather than averaging what is there. That work belongs to enlargepicture.com, which is built for it.

The distinction is worth holding on to when you are judging a result. A kernel that invented nothing has produced an honest, soft enlargement. A model that invented something has produced a convincing image containing details that were never in front of the lens. Both are useful. Only one of them is a measurement, and this is the site that measures.

Questions people bring to this

If nothing is gained, why does a larger version sometimes look better?: Because a display is doing its own scaling, and doing it worse. A 400 pixel image shown in an 800 pixel frame is being interpolated by the browser with a fixed, cheap filter. Producing the 800 pixel version deliberately, with a kernel you chose, can look better than letting the page do it — the information content is identical, the presentation is not.
Does a higher-numbered kernel recover more?: No. Lanczos, bicubic and bilinear differ in how they distribute the error, not in how much information they have. A windowed sinc makes edges look crisper by overshooting at them, which reads as sharper and is not more faithful. There is no ordering of these three in which one recovers detail another cannot.
What about the tools that clearly do add detail?: They are not interpolating. A model trained on millions of photographs predicts what plausibly sat between the samples, and produces a result that is convincing rather than recorded. That is a legitimate and often desirable thing to want, and it is a different operation with a different set of honesty problems, which is why it lives on a different site.