Expand pictures without cropping

It makes the canvas bigger. That is the whole idea, and it is worth being precise about what it means, because three other things sound like it and are not. Your photograph is copied into a larger grid at its own size, landing on whole-number coordinates with smoothing turned off for that one draw, and then the space around it is painted. The rows and columns you photographed are the same rows and columns when you save.

Nothing here resamples the original photograph.

So the thing that grows is the frame, not the picture. A 1080 by 1920 photo from a phone asked for 16:9 comes back as 3,413 by 1,920: the height is untouched, 2,333 columns of new material appear at the sides, and roughly two thirds of what you are looking at was built rather than taken. That proportion is printed under the canvas the whole time you are working, because it is the number that tells you whether the shape you asked for was a reasonable thing to ask of this photograph.

The other direction works the same way round. A wide photograph asked for 9:16 grows downward and upward rather than losing its sides, and a photograph of any shape asked for a square grows whichever axis is short. One rule, applied consistently: the axis that is too small grows, and the axis that is already big enough is left alone.

Coming in: JPG, PNG, WebP, AVIF, GIF, BMP and ICO — TIFF where your browser can read it — plus HEIC from an iPhone, which pulls in a decoder the first time one arrives rather than making every visitor carry it. Going out: JPG, PNG and WebP, which are the three every browser can genuinely write. A file heavier than 120 MB is refused before it is opened, and a photograph past roughly 32 megapixels — 16 on an iPhone — stops and asks whether a lighter working copy is acceptable, because the canvas being built is larger than the photograph going into it.

A few file types arrive with something worth mentioning, and the tool mentions it under the canvas rather than burying it here. A JPG written for a printing press carries four ink channels instead of three screen ones, so the browser converts it and the colour shifts very slightly. A picture that already has see-through pixels at its edge makes mirror and stretch copy that emptiness outward, so the rail points you at a colour or a gradient instead. And a file with more than one frame is opened at the first.

• Nothing here knows what was actually outside the frame

  Seven of the eight fills are arithmetic on the pixels you already have: reflected, smeared, extended, blurred, averaged or swept. The eighth is a model, and it does not know either — it produces what usually carries on past an edge like yours, which is a good guess and never a record. On sky, grass, water, wallpaper and fabric the distinction does not matter, because more of the same thing is the truthful answer. On a straight architectural line, a tiled floor, a line of text or a face at the edge, it matters enormously, and nothing in the rail will rescue it.

• The generated fill is real now, and this page was wrong about it for months

  This site used to say a model small enough to run in a browser tab was several hundred megabytes, needed graphics support most phones lack, and took the better part of a minute. Those figures were estimates and they were wrong. The model it now offers is 26,639,447 bytes, MIT-licensed (code and weights), runs on plain WebAssembly with no graphics requirement, and takes about two and a half seconds for each 512-pixel square it fills. What we said about the subjects it fails on was right, and it is still offered only up to twice the area of your photograph, because past that it stops continuing an edge and starts inventing a scene.

• An animated file comes out as a single still

  A GIF or an animated WebP is opened at its first frame and saved as one picture. Expanding every frame of an animation, keeping the timing and writing the container back out is a genuinely different program, and a tool that quietly threw away ninety frames without saying so would be worse than one that says this out loud.

Every route on this site drives the same engine. What differs is how much has already been decided when you arrive: the front page starts neutral and waits for you to type a ratio, and the six others open with a target, a fill and a seam setting chosen for the job in their name. That saves three decisions on the way in and costs nothing, because every control is still sitting in the rail and every one of them still moves.

So pick by the sentence you would use to describe the job. “I have a still from a phone video and the slot is widescreen” is the 9:16 page. “This is tall and it has to be wide”, with no particular numbers attached, is portrait to landscape. “It has to be square and I am not cutting anybody’s feet off” is the square page. “I want the soft version everybody else posts” is blurred edges. “The printer asked for a border” is margins. “I want the new edge to carry on from the photograph rather than repeat it” is the model, which is a download and a wait and is offered only for modest extensions. If none of those is your sentence, stay here and type the ratio in.

What the route never changes is the arithmetic. How much of the finished frame gets built rather than photographed depends on your picture’s proportions and the target’s, and on nothing else — the door you came in by does not move that figure by a single pixel. Past about two thirds the fill has stopped decorating the picture and started being it, which is the point at which the choice in the fills guide matters more than the shape you asked for.