Image resolution explained: pixels, dpi and how big you can go

Resolution is the most misused word in photography. It gets used for pixel count, for dpi, for perceived sharpness and for file size, and those are four different things that only sometimes move together. Getting them straight takes about five minutes and saves a great deal of arguing with a printer, a marketplace listing or a designer.

This is the hub for everything on this site about size and sharpness: what the numbers mean, how to work out what a job needs, what an upscaler can add and what it cannot, and what to do when a picture is simply too small.

Resolution is a count of pixels

A digital picture is a grid. Resolution is how many cells that grid has: 3000 across by 2000 down, which is 6,000,000 cells, which is six megapixels. That is the whole definition. It is decided when the shutter fires or when the file is exported, and nothing afterwards changes how much of the world was actually sampled.

Every other number people call resolution is derived from that one, or is about display rather than content:

• Megapixels are width times height, divided by a million. A 12 megapixel phone photo is about 4000 by 3000.
• File size depends on the pixel count and on the format and its compression. A 6 megapixel JPG might be 2 MB and the same picture as a PNG 15 MB, with identical resolution.
• Dpi or ppi is a ratio, not an amount, and is covered below.
• Sharpness is about focus, camera shake, lens quality and compression. A blurry 24 megapixel photo is high resolution and still blurry.

That last one is worth sitting with. High resolution does not mean sharp, and a picture can fail for either reason. Why does my image look blurry separates the causes, because the fix is completely different for each.

Dpi is an instruction, not a property

Dots per inch is a relationship between pixels and physical size. It only becomes real when the picture meets something physical: paper, or a screen at a known size.

The arithmetic is simple. Pixels divided by dpi gives inches. A 3000 pixel wide photo is:

Notice that the file did not change in any of those rows. The same pixels are simply being spread over more or less paper. This is why “please send it at 300 dpi” is an incomplete request: 300 dpi at what size? A designer asking for that almost always means “send me enough pixels that it can be printed at 300 dpi at the size it will appear”, and the useful reply is to ask how big it is going on the page.

Files do store a dpi number, in the JPEG header or in a PNG’s pHYs chunk, and it is only a note. It tells software what size to assume if nobody says otherwise. Changing it in an editor changes nothing about the picture. Dpi explained goes through what to scan and export at, and where the number does matter.

On a screen, dpi is ignored entirely. A browser, a slide deck and a phone all place a picture by its pixel count. That is why the advice for the web is always in pixels and never in dpi.

What a job actually needs

Work backwards from where the picture is going.

Upscaling an image for print goes through this in more detail, including how the requirement falls with viewing distance.

For print, at 300 dpi:

For print at a distance, 300 dpi is generous. A poster read from two metres is fine at 150 dpi, and a banner read from ten metres is fine at 50. Print shops publish their own minimums and they are usually lower than people expect.

For a website in particular, the pixels a picture is saved at and the pixels it is displayed at drift apart quickly, and the gap is pure weight nobody sees. getReport’s image sizes that do not hurt is that argument from the page’s side, and its image size checker measures the gap on a page you have published.

For screens, take the size in CSS pixels the picture will occupy and double it, because dense displays draw two device pixels for each CSS pixel. A hero image spanning a 1200 pixel wide layout wants roughly 2400 pixels. More than that is bytes nobody sees, which is why reducing PNG file size matters as much as having enough pixels.

For marketplaces, the platform sets the number. Amazon wants at least 1000 pixels on the longest side for zoom to work and recommends 1600. Amazon’s product image requirements has the detail.

When the picture is too small

There are only four honest answers, in order of preference.

Get the original. The version you have has often been resized on the way to you: by a messaging app, by a website’s thumbnail generator, by an email client. The photographer, the supplier or your own phone’s library usually still has the full size file. This is worth ten minutes of asking before anything else.

Use it smaller. A picture at its own resolution always looks better than the same picture stretched. Reducing the size on the page or cropping the layout around it is not a defeat.

Enlarge it properly. A good upscaler genuinely helps, especially on photographs. How to upscale an image covers the method and the settings, and the image upscaler on this site does it in the browser with either a trained model or a classical filter.

Recreate it. For a logo or a diagram this is the right answer and often a quick one: a vector version has no resolution at all and can be output at any size. Enlarging a logo without blur covers getting from a small bitmap to something that scales.

What enlarging can and cannot do

Enlarging adds pixels. The question is what goes in them.

A classical resampler, such as Lanczos, computes each new pixel from the ones around it. It cannot add detail because it has nothing to add: every value it produces is a weighted mixture of values that were photographed. Done well, with the arithmetic in linear light and a proper kernel, the result is a clean, faithful, slightly soft version of the original at a larger size.

A trained model does something different. It has seen millions of pairs of large and small patches, so it has learned what skin, brick, fabric and hair look like at the larger size, and it draws that. The output genuinely looks more detailed, and on a photograph it is not close: a model beats a filter easily.

The catch is in the word “plausible”. A model draws what a picture like yours usually looks like, not what yours actually looked like. On a face it produces a convincing face that is not quite the person. On a number plate it produces convincing characters that may be the wrong ones. That is fine for a poster and wrong for a record, and it is why this site lets you choose the engine rather than deciding for you. What AI upscaling can and cannot do has the full argument with examples.

Resolution and transparency

None of this changes when a picture has an alpha channel, but two practical things are worth knowing.

A cutout’s soft edge is a long ramp of half-transparent pixels, and it enlarges badly if the tool is careless. Resampling straight RGBA drags the colour of fully transparent pixels into the visible edge, which produces the grey halo people see around an enlarged cutout. The fix is to premultiply the alpha before any arithmetic, which the upscaler here does. White halo around a cutout covers the related problem at the point the cutout is made.

The other is that a model has three channels and cannot see alpha at all. If nobody handles that, the transparent area, which usually holds black, is fed to the network as a black wall and it draws a dark fringe along the subject. Filling the transparent area with the colour beside it first, then enlarging the alpha separately, removes the problem completely.

The short version

Count pixels, not dpi. Work backwards from where the picture is going. Get the original if you can. If you cannot, enlarge with a model when the picture is photographic and with a filter when it has to stay faithful, and be honest with yourself about which of those the job is.

Questions

What is image resolution?

The number of pixels in the picture, written as width by height: 3000 by 2000. Multiply them and you get megapixels, so that example is six. It is decided when the picture is captured or exported and nothing you do afterwards changes what was actually recorded.

Is 300 dpi the same as high resolution?

No. Dpi on its own says nothing, because it is a ratio rather than an amount. A picture is 300 dpi at some size and 72 dpi at another, and the number stored in the file is only a note to whatever prints it. What matters is the pixel count and the size on paper.

How many pixels do I need for an A4 print?

About 2480 by 3508 at 300 dots per inch, which is the usual standard for photographic printing. At 150 dpi, which is fine for a poster seen from a metre away, half of that is enough.

Can I increase the resolution of an image?

You can increase the pixel count, which is what an upscaler does, and on a photograph a good model produces a genuinely better looking result than stretching. You cannot recover detail that was not recorded. The distinction matters for anything that has to be accurate rather than merely convincing.

Why does my picture look sharp on my phone and soft when printed?

Because a phone screen is small and dense and hides a shortage of pixels, while paper spreads the same pixels over a much larger area. A 1200 pixel wide picture is beautiful on a phone and about four inches wide at print quality.

Does saving as PNG increase resolution?

No. Changing format never adds pixels. PNG is lossless, so it preserves what is there, but a 600 pixel JPG saved as a PNG is still 600 pixels.