Put a cut-out person on another photo

Placing one picture on another is what the rest of this site does. The thing that makes it look wrong is almost never the placement — it is that the person you dropped in arrived inside a rectangle of their own living room. So the work happens somewhere else first: you cut them out, save a transparent PNG, come back, and place it. This page folds that step into the placement, so the two are one pass.

What it will not do is give you the cut-out on its own. There is no transparent PNG at the end of this and no button that offers one; the file this page writes is the composite. That is a line rather than an oversight: a tool whose product is the cutout is a background remover, a different job with its own controls, and this page would do it badly next to what it does well — put the person on the scene.

1. 1Open the scene. The photograph that everything else lands on — a beach, a street, a plain colour you made earlier. This is the picture you will be saving.
2. 2Drop in the photo of the person. A full-length shot or a portrait, exactly as it came off the camera or the phone. It does not need a clean background and it does not need to be a PNG.
3. 3Press the cutout control, once. Nothing is downloaded before you press it. The first press fetches 5.9 MB of model from our model file server and 13.3 MB of runtime from this domain, and the button says so before it does. Your two photographs stay in this tab throughout: the model travels to you, not the other way round.
4. 4Place them on the scene. The subject is now an element like any other — drag it, pull a corner, turn it with the grip, drop the opacity, choose one of the sixteen mixing modes. An outline or a soft shadow underneath is what makes somebody look like they are standing in a picture rather than pasted onto it.
5. 5Save the finished picture. The composite is drawn again at the scene's full size and written as PNG, JPG or WebP. What you get is the photograph with the person in it — this page never hands back the person on their own.

Before you press anything

Nothing. No model, no runtime, not a byte. Opening both photographs and placing one on the other fetches nothing from anywhere.

The model

Portrait cutout, 6,163,938 bytes exactly — 5.9 MB — Apache-2.0-licensed, from our model file server.

The runtime that executes it

13.3 MB of WebAssembly, served from this domain rather than anywhere else, so this page talks to exactly one other host.

First press, all in

19.2 MB. Every press after that: nothing, on this device or any other you have already used it on.

Time for one cutout

About 90 ms on a single WebAssembly thread — the figure measured in a browser, not quoted from a model card, and the one a real visitor sees.

What it reports

Two numbers per pixel of a 192-pixel square: how much it thinks that pixel is a person, and how much it thinks it is not. No identity, no estimate, nothing kept.

Where your photographs go

Nowhere. They are decoded into memory in this tab and thrown away when it closes.

• It finds people. A portrait or a full-length figure is what it handles well; a product, a pet or a car is not what it was trained for, and it will keep almost nothing rather than guess at an edge.
• The edge is decided at 192 pixels square and stretched over your element. A shoulder comes back clean; individual strands of hair come back as a suggestion of hair.
• The cutout works on one layer at a time. Two people on one scene means adding both pictures and lifting each one from its own card; each keeps its own cut.
• The model is stored in this browser once you use it. That is the only thing this site writes to your device, it is a file that came from the network rather than anything of yours, and the control beside the cutout removes it.
• Decline the download and nothing here stops working. A transparent PNG cut elsewhere places exactly as it always did.

A clean cut is about half of it. The other half is that a figure lifted from one photograph carries that photograph’s light, and the scene underneath has its own. Three things close most of that gap, and all three are controls you already have.

Drop the opacity a few per cent so the scene shows through the edges very slightly. Turn the shadow on and push it in the direction the scene’s own shadows fall — the offset turns with the element, so it stays correct if you rotate. And if the subject was shot warm and the scene is cool, a mixing mode is a blunt but quick correction: Luminosity keeps their shape and takes the scene’s colour, which is usually too much, while twenty per cent of Soft light is usually about right.

Does my photo get uploaded to do the cutout?

No, and the direction is the point. A file does cross the network the first time you press the control — 6,163,938 bytes of model weights from our model file server — and it travels towards you. Both of your photographs are decoded into memory in this tab, and the separation runs on this processor, in about 90 milliseconds. There is no server here to send a picture to. The privacy page names the one request and what it carries.

It cut out the person but ignored the dog beside them. Why?

Because the model was trained on people and on nothing else. It is a human segmenter: give it a portrait or a full-length figure and it separates them from what is behind them well; give it a chair, a bottle, a car or a pet and it will find very little to keep and say so rather than inventing an edge. That is the honest boundary of what is running here, and widening it is not a setting — it is a different model.

The hair looks chewed. Can I fix it?

Partly, with the two controls under the cutout. The model decides the edge on a 192-pixel square and that decision is stretched over your element, so it follows a shoulder closely and a strand of hair loosely — there is genuinely less detail there than the photograph has. Soft keeps the uncertain band, which reads better against a busy scene; crisp cuts through it, which reads better against a flat one. Moving the keep slider a few per cent either way often does more than either. If the picture depends on the hair, a photograph shot against a plain wall is worth more than any slider.

Can I download just the cut-out person as a transparent PNG?

No. That is a deliberate line rather than a missing feature: the cutout on this page is a step inside placing somebody on a photograph, and the file this site writes is the finished composite. A tool whose output is the subject on its own is a background remover, which is a different job with different controls — refining hair, choosing a new background, exporting the subject alone — and this page does not pretend to be one.

What happens if I say no to the download?

Everything else. The download only starts when you press the control that names its size, and the page works without it exactly as it worked before the model existed: place a transparent PNG you cut somewhere else, or drop the whole photograph on and set it to Multiply or Screen so the background falls away. Nothing on this page is gated behind the model, and nothing is fetched until you ask.

Is it kept on my device afterwards?

The model file is, on purpose, so the second time you use it nothing is requested at all and the cutout works with the network switched off. It sits in this browser's cache store under a name with a version in it. There is a control beside the cutout that deletes it again, and your photographs are not in there — they are never written to storage of any kind.

• What the blend modes doMultiply, screen, overlay and soft light, with the arithmetic said in words.
• Place something on a photoThe stage and a deck of layers: logos, stickers, textures or a second photo, each moved until it looks right.
• PrivacyWhat happens to a file you open, the cookieless visit count, and the one model download.