Does resizing lose quality?

What downscaling actually does

A 4000 × 3000 photograph holds twelve million samples. An 800 × 600 version holds four hundred and eighty thousand. Ninety-six percent of the measurements are gone, and no amount of cleverness in how they are combined changes that number. The question is never whether information is lost — it is whether the surviving samples are a fair summary of the ones that went.

The bad way to do it is to pick every fifth pixel and throw the rest away. That is fast, and it produces the specific ugliness people mean when they say a resize looks wrong: fine repeating detail turns into a pattern that was never there. Brick walls shimmer. Striped shirts grow rainbows. Hair grows a fringe. This is aliasing, and it is not blur — it is detail being replaced with different, false detail.

Why this page steps down instead of jumping

A resampling filter looks at a window of source pixels around each output pixel. Going from 4000 wide to 800 wide, each output pixel is a summary of five source pixels, so the window has to be at least five wide or most of the source is never read at all. Filters can be widened to compensate, and this one is, but a very wide window is both slow and progressively less accurate.

So the work is split. The picture is halved, and halved again, by averaging each 2 × 2 block into one pixel — a box average, which reads every single source pixel exactly once and cannot alias. That repeats until the picture is within twice the target size. Then one Lanczos3 pass lands it exactly on the numbers you typed. From 4000 to 800 that is two halvings, to 1000, and then one short pass.

The result is a picture where every original pixel contributed something, resampled at the end with a filter that is sharp because it only has a short distance to travel. Going up, the filter is Mitchell instead: it is softer than Lanczos and does not ring, and ringing is what you get when you ask a sharp filter to invent pixels between the real ones.

Why there is no filter menu

Because on this page it would be a dropdown that makes results worse. The stepped box-and-Lanczos3 path is the right answer for photographs going down to screen sizes, which is what almost everyone types two numbers for, and offering four other kernels beside it mostly buys people a chance to pick a worse one.

There is a real case for choosing — print production, where the paper, the ink and the screening make a genuine difference and the person choosing knows why. That case is not this page. It belongs with the physical sizing tools, along with DPI and centimetres and everything else that is a decision about paper rather than about pixels.

What this page does instead is tell you what ran. Under the control bar, on every resize, there is a line saying how many halvings happened and which filter finished the job. A tool that makes the choice for you owes you the choice it made.

The trip back does not return anything

Making a picture smaller throws pixels away. We choose the order of operations that throws away the least, and there is no setting on this page that gets them back — a 4000 px photo resized to 800 and then back to 4000 is not the photo you started with.

This is worth being blunt about because it is the single most common thing people hope a resizer will do. The 800-wide file contains four hundred and eighty thousand samples. Asking for 4000 wide again produces a file with twelve million samples in it, eleven and a half million of which are interpolations between the ones that survived. It will be larger in kilobytes and no more detailed.

Keep the original. If you need three sizes, make all three from it rather than making each one from the last. That single habit avoids most of what this guide is about.

Where quality is lost that is not resampling

• The encoder. Saving as JPG at quality 60 will do more visible damage to a photograph than any resampling decision on this page. The number beside the download button is the real encoded size, so you can move the quality slider until the trade is one you are happy with.
• Repeated JPG saves. Every save re-encodes and every re-encode loses a little more, even if nothing changed. Resize once from the original rather than editing a JPG four times.
• Colour space. The working space here is 8-bit sRGB. A photograph in a wider space is flattened into sRGB on the way through, and that is a real loss even though most screens would never have shown the difference.
• Transparency meeting JPG. There is no alpha channel in a JPG, so anything see-through lands on white. The page warns before the save rather than after it.

Questions we get asked

Can I resize down and then back up to get the original?

No. The information was discarded, not compressed, and there is nothing left in the file to reconstruct it from. Going back up will produce a file with the original dimensions and roughly the detail of the small one.

Why does my picture look worse after resizing when other tools look fine?

Usually aliasing rather than blur, and usually because the tool made one large resampling jump instead of stepping down. The give-away is a shimmer or a moiré pattern on fine repeating detail — brickwork, hair, fabric, small text — rather than a general softness.

Does saving as PNG avoid the loss?

It avoids a second, separate loss. PNG stores the resampled pixels exactly, so nothing further is thrown away by the encoder. It does not undo the resampling, because the resampling happened first.

Is there a size below which it stops mattering?

Not really, but there is a size below which the source stops mattering: once the target is small enough, a 45 megapixel original and a 12 megapixel one produce nearly identical files. The extra pixels were only ever going to be averaged together.