How to resize an image without losing quality

Shrinking and enlarging are not two directions of the same operation. One is nearly free. The other is a guess, however expensive the software making it.

Updated 28 August 2026 · 11 min read

“Resize without losing quality” is one request, but it hides two completely different problems. One of them has a good answer. The other has an honest answer, which is not the same thing.

Making an image smaller means deciding what to throw away, and there are careful ways to do that. Making an image bigger means deciding what to make up, and there are no careful ways to do that — only more or less convincing ones. Most disappointment with resizing comes from expecting the second problem to behave like the first.

Shrinking and enlarging are not the same operation

It is tempting to think of a resize slider as a single dial with a middle at 100%, where dragging left and dragging right are symmetrical. They are not.

When you reduce a 4000 px photo to 1000 px, everything in the result was recorded by the camera. Sixteen source pixels are being summarised into one output pixel. The summary is lossy in the sense that you cannot get back to the original, but nothing in the output is fictional. Done properly, a downscaled image can look flawless — often better than the original at that display size, because averaging suppresses sensor noise.

When you enlarge a 1000 px photo to 4000 px, you are asking for sixteen times as many pixels as were ever recorded. Fifteen out of every sixteen have to come from somewhere other than the scene in front of the lens. That is not a limitation of your software or a sign that you should pay for better software. It is arithmetic.

So the useful way to hold this: downscaling is a summarising problem, upscaling is a guessing problem. The rest of this guide is about doing the first one well and being realistic about the second.

What resizing actually does to the pixels

An image is a grid of coloured squares. Resizing it means building a new grid, of a different size, and working out what colour each square in the new grid should be. That process is called resampling, and every output pixel is a weighted blend of the input pixels that fall near its position.

The differences between resampling methods come down to how wide a net each output pixel casts and how it weights what it finds.

• Nearest neighbour takes the single closest input pixel and copies it. No blending at all. It produces hard blocky edges, which is exactly right for pixel art and exactly wrong for photographs.
• Bilinear blends the four surrounding pixels. Cheap, smooth, and inclined to look soft.
• Bicubic considers a wider neighbourhood with a curve that slightly overshoots at edges, which reads as more contrast. It is the default in most photo software for good reason.
• Lanczos casts a wider net still and tends to preserve fine detail crisply, at the cost of occasionally ringing — faint halos beside high-contrast edges.

You rarely get to choose the method in a browser tool, and for ordinary photographs the choice matters far less than the thing described in the next section. A well-executed downscale with a plain algorithm beats a careless downscale with a fancy one.

Why one big jump loses detail, and stepping down does not

Here is the failure that produces most “my resized photo looks terrible” complaints.

Suppose you reduce a 4000 px image to 400 px in a single pass, and the resampler is blending a small neighbourhood — say four or sixteen nearby pixels — around each output position. Each output pixel now stands for a ten-by-ten block of the original — a hundred source pixels in total — but it only looked at a handful of them. Ninety-odd pixels out of every hundred are never consulted. They are simply skipped.

The consequences are visible and specific. Fine detail vanishes unpredictably, because whether a thin feature survives depends on whether it happened to sit under a sampled point. Diagonal lines go jagged. Regular patterns — brickwork, roof tiles, suit fabric, mesh, printed halftone dots — develop moiré: broad wavy bands of false colour that were never in the scene. That is aliasing, and it is the signature of sampling a detailed signal too sparsely.

The standard fix is to reduce in stages. Halve the image, halve it again, and keep halving until you are close to the target, then make the final short step. Each halving averages a 2 × 2 block into one pixel, so nothing gets skipped, and by the time you arrive at 400 px every source pixel has contributed to the result through the chain. Edges stay clean, patterns stay patterns, and the whole image reads as sharper.

This is not a proprietary trick or a secret setting. It is a well-known technique that image libraries and game engines have used for decades, and its only real cost is a few extra fractions of a second.

One more habit worth adopting: a slight sharpen after downscaling is normal, not cheating. Averaging pixels together inherently softens edges, so a modest amount of sharpening restores the apparent crispness the averaging cost you. The word doing the work there is modest — heavy sharpening produces bright fringes along every edge, and that is far more obvious than softness.

Enlarging, and what no software can do

Now the harder half. If the camera recorded 1000 px across, the information about what sits between two neighbouring pixels does not exist anywhere. It was never measured. No amount of processing recovers it, because recovery is not what is happening — invention is.

Conventional interpolation is honest about this. Bicubic or Lanczos enlargement fills the new pixels by smoothly continuing the values around them, so a doubled image looks like the original seen slightly out of focus. Nothing false is added; the result is simply soft, and softness scales with how far you pushed it. A 25% enlargement is usually unnoticeable. Doubling is noticeable but often acceptable. Quadrupling looks like what it is.

Machine-learning upscalers work differently. They have been trained on enormous numbers of images and have learned what plausible detail looks like at various scales, so instead of smoothly interpolating they synthesise texture: pores on skin, individual hairs, thread in fabric, letterforms in signage. Often the result is genuinely more pleasing than a blurred one, and for a landscape photo destined for a wall, that may be exactly what you want.

But the detail is generated, not recovered. The model is answering the question “what would probably be here?” — and probably is not the same as actually. Most of the time this is harmless. Sometimes it is not.

The cases where an invented detail is worse than a soft one deserve to be named. A face is the obvious one: an upscaler resolving an ambiguous cluster of pixels into a nose, a jawline, or an expression is producing a person who is subtly not the person photographed, and the result carries a confidence the original never had. Text is the same problem with sharper edges — a blurred digit is visibly unreadable, whereas an invented digit is crisply, legibly wrong, and nothing in the image warns you which one you are looking at. Scanned contracts, invoices, identity documents, meter readings and medical images all inherit this hazard. For anything that will be used as evidence of what something said or looked like, softness is a feature: it tells the viewer honestly how much was known. If you have any doubt about an image's provenance or purpose, leave it soft and say it is soft.

The practical version of all this is unglamorous. Before enlarging, check whether a larger original exists — the full-resolution file on the camera, the version before it went through a chat app, the original send rather than the forwarded thumbnail. Recovering the real file beats every upscaler in existence, and it takes two minutes.

How many pixels you actually need

Most people resize larger than necessary because they have never worked out the target. The numbers below are ordinary conventions, not rules, and any of them can reasonably be doubled for a display that needs to look sharp on a high-density screen.

Exact platform dimensions change often enough to be worth checking against a current list rather than guessing; the social media image sizes guide keeps those in one place. And if what you are really up against is a byte limit rather than a pixel limit, resizing is only one of the two dials involved — compressing to an exact file size covers how they interact.

DPI is a printing instruction, not a quality setting

This one causes more wasted effort than any other misunderstanding in the subject, so it is worth being blunt.

A digital image has exactly one measure of how much detail it holds: its pixel dimensions. DPI — or more correctly PPI, pixels per inch — is a number written into the file's header that says how large the image would like to be printed. It is metadata. It is a note to a printer.

Setting a 1000 × 1000 px image to 300 DPI does not add detail; it says “print me at about 3.3 inches square.” Setting the same file to 72 DPI does not remove detail; it says “print me at about 14 inches square.” The pixels are byte-for-byte identical in both cases. On screen the setting is ignored entirely. And a file with no DPI field at all is not defective — most software just assumes a default.

So when a form or a print shop asks for 300 DPI, what they need is enough pixels for the size they intend to print. The arithmetic is one division: pixels ÷ desired inches = effective PPI. A 3000 px wide image printed 10 inches wide gives 300 PPI. The same image printed 20 inches wide gives 150 PPI, which is a perfectly normal figure for something viewed from further back. Three hundred is the traditional target for print held at reading distance, not a threshold below which images stop working.

Keeping the proportions honest

Width and height are not independent. If you type both numbers and they do not match the original ratio, the image is stretched — faces widen, circles turn into ovals, and the effect is instantly recognisable even to people who cannot say what is wrong.

Fixing one dimension and letting the other follow is the safe default. Constraining the longest side is safer still for a mixed batch, because portrait and landscape photos then land inside the same bounding box without either being distorted. When a form genuinely demands exact pixel dimensions that do not match your image, the correct answer is to crop to the required ratio first and then resize — cropping removes content you can see, whereas stretching damages everything at once. The aspect ratio guide goes through the arithmetic and the common ratios.

Common questions

Does resizing reduce quality? Downscaling removes pixels, so information is genuinely gone, but if the result is only ever viewed at the smaller size you will not see a loss. Upscaling does not remove anything and still looks worse, because the added pixels carry no new information.

Can I resize back to the original size afterwards? No. Once the pixels are discarded they are not stored anywhere, and enlarging back gives you a soft approximation. Keep the original file and derive every version from it.

Why does my image look blurry after resizing down? Averaging softens edges, which is normal — a slight sharpen usually fixes it. If it looks jagged or shows wavy patterns rather than soft, that is aliasing from too large a single step, and a tool that steps down in stages will produce a cleaner result from the same original.

Do AI upscalers actually work? They produce convincing texture, and for photographs viewed for their own sake the result is often better than a blurred enlargement. They do not recover what was never recorded. For faces, text and any document that will be relied upon, treat the added detail as an illustration rather than as information.

Should I change the DPI before printing? Only if the printer explicitly asks for a particular value in the file. What determines print quality is pixels divided by inches, and no header field alters that.

Is it better to resize or to compress? They solve different problems and often work together. If the picture is far larger than it will ever be displayed, resize first — it is the more powerful lever and costs nothing visible. Reach for heavier compression only once the dimensions are sensible.