Image enlarger

Lanczos resampling, shown beside your browser’s own scaler so you can judge it.

Or drop it here, or paste. PNG, JPEG, WebP, GIF or AVIF, one at a time. Nothing is uploaded.

Options

Enlarge by

Presets stop at 4×: past that a measured pixel is asked to become 16 or more, and what fills them is invention.

More optionssaved as the smallest of WebP, JPEG or PNG

Applied when you press Enter or leave the box. Between 1.01× and 4× of the original.

Save as

Encodes WebP, JPEG and PNG, each where this browser can really write it, and keeps the smallest file. JPEG is left out when the picture has see-through pixels.

Add an image, or try the sample, to start.

What image resampling is

Computing new pixels from the ones around them when an image changes size. A resampler redistributes the information that is already there; it cannot add any, which is the distinction between resampling and the neural super-resolution the term "enlarger" now implies to most people.

What this tool does that others do not

  • Lanczos3, shown beside the browser’s own scaler

    The right-hand pane is the browser scaling the same image at its highest quality setting, which is Lanczos-like in Chrome and Firefox and bicubic in Safari, and no browser says which you got. Lanczos3 on the left samples a wider neighbourhood with a windowed sinc function and gives the same result in every browser. Both are drawn at the same size so the difference is visible rather than asserted — if you cannot see one on your image, either is fine.

  • It says what it cannot invent

    A four-pixel face does not become a recognisable one. AI super-resolution models hallucinate plausible detail, which is useful for a wallpaper and dangerous for anything you intend to read, identify or rely on. Shipping one here was measured and rejected: the WebGPU build of the runtime came to 27,648 KB, over the 25 MiB per-asset ceiling this site’s host serves, and the 800 KB of model weights compress by about 7% because trained float weights are close to random.

  • Factors that reflect what the information allows

    Presets stop at 4×, where competitors offer 8× and 16×. At those factors a single measured pixel is being asked to become 64 or 256, and what fills them is invention. Two times is usually indistinguishable at normal viewing distance; three is acceptable for print at arm’s length; four starts to look soft on fine texture.

How to use it

  1. Drop the image in and pick a factor

    Start at 2×. If the result is soft, the limit is the information in the file rather than the setting.

  2. Compare the two panes on your own image

    The difference is small at 1.5× and obvious at 3×, and it depends entirely on the picture. Judge it rather than take the claim.

  3. Stop when the enlargement stops adding anything

    Past about 4× you are enlarging the blur rather than the subject. A 200-pixel image contains 200 pixels of detail at every size you display it.

Worked example

Real input, real figures.

Choosing between the two resamplers

You give it

A photograph with fine texture, enlarged 3×.

You get

In Safari, whose high-quality path is bicubic, Lanczos3 holds visibly more edge contrast on fine texture; in Chrome and Firefox the two are often close. On flat graphics they are nearly identical everywhere. The comparison is on the page so you judge your own image, not this sentence.

Against an “AI upscaler”

AspectThis toolA neural upscaler
MethodLanczos3 resamplingA trained model
Invented detailNonePlausible, and sometimes wrong
Faces and textSoft but unalteredFrequently mangled
Maximum factor4×, because the information runs out8× or 16×
Download cost14 KB, this page onlyMegabytes of runtime and weights

Guide

What it will not do

  • Output is capped at 16,777,216 pixels (16 megapixels) in every browser, because some browsers return a blank canvas past that rather than an error. A 4000 × 3000 photo therefore comes out at 4729 × 3547 whichever factor you pick, about 1.18×, and the tool reports the size it used.

What enlarging can and cannot do

Every enlarger faces the same wall: the image contains a fixed amount of information, and making it bigger does not add any. What a good resampler does is distribute what is there as cleanly as possible, so edges stay edges instead of turning into stairs or smears.

Lanczos3 samples a wider neighbourhood and weights it with a windowed sinc function, and it gives the same result in every browser. The browser’s own high-quality scaling is close to it in Chrome and Firefox and bicubic in Safari, so how much the difference matters depends on where you are. Both are shown side by side above so you can judge it on your own image rather than take our word for it.

What it will not do is invent. If you need a four-pixel face to become a recognisable one, no resampler can help and any tool that appears to has made the face up.

Questions people ask

Straight answers. Each one links directly: copy the address after opening it.

Does this use AI to add detail?

No, and it is worth being exact about what that means. This resamples — it computes new pixels from the ones around them using Lanczos3, which is the best general-purpose filter for the job. It cannot invent a face that was four pixels wide, because that information was never recorded. AI super-resolution models hallucinate plausible detail, which is genuinely useful for a wallpaper and genuinely dangerous for anything you intend to read, identify or rely on. What we could ship in a browser is a small model whose weights are 800 KB and do not compress, behind a runtime that is over the per-file limit our host allows. So we ship the honest version instead.

Then why not just use the browser’s own scaling?

Because what the browser does depends on the browser. Asked for its highest quality, Chrome and Firefox use something close to Lanczos and Safari uses bicubic, and none of them says which you got. Lanczos3 here gives the same result everywhere. The two are drawn side by side at the same size, so you can judge on your own image — if you cannot see a difference, either is fine.

How much can I enlarge before it looks bad?

Two times is usually indistinguishable from the original at normal viewing distance. Three is acceptable for print at arm’s length. Four starts to look soft on anything with fine texture, and beyond that you are enlarging the blur rather than the subject. The limit is not the algorithm, it is information: a 200-pixel image contains 200 pixels of detail at every size you display it.

Is there a size limit?

Output is capped at 16 megapixels (16,777,216 pixels) in every browser, because past that some browsers return a blank image rather than an error. So a 4000 × 3000 photo comes out at 4729 × 3547 whichever factor you choose, and the tool tells you the size it used instead of handing back a white rectangle. Nothing is uploaded, so there is no file-size limit of the kind a server-based tool imposes.