How to spot a face swap in a photo

Look at the boundary rather than the face. Swaps are convincing in the middle and weak at the hairline, the jaw and the neck, where the fitted region has to meet a real photograph.

A face swap is not a fake picture. It is a real picture with a fake region in it, and that difference decides how you find one. The room is genuine, the clothing is genuine, the light is genuine, and one area roughly the size of a face is not.

This is the case that whole-image scoring handles worst, and the case that turns up most often in fake profiles, impersonation and harassment. It is worth knowing how to approach on its own terms.

Look at the edges, not the middle

The instinct is to stare at the face. The face is the part the model spent all its effort on, and on a good swap it is convincing. The weakness is at the boundary, where a generated region has to be fitted against pixels that came from a camera.

The hairline, where the fitted face meets real hair. Look for a soft band, a slight change in sharpness, or hair that stops rather than overlapping the forehead as it should.
The jaw and chin against the neck, where skin tone and lighting have to match across the seam and often nearly do.
Ears, which frequently belong to the original person while the face does not. Mismatched ears against a face are a strong tell.
The neck and throat, usually untouched, so any difference in grain or colour cast between face and neck is worth noticing.
Glasses and jewellery crossing the boundary, where a frame can change thickness or fail to line up on the far side.
Sharpness. A swapped face is often slightly softer or slightly crisper than the rest of the frame, because it was rendered separately.

Lighting direction is the other reliable check. Work out where the light is coming from by looking at the shadows in the room, then check whether the shadows on the face agree. A face lit from slightly the wrong side is a common failure, because the source face was photographed somewhere else. Reflections in glasses and in eyes are worth the same check, since a face borrowed from another photograph brings that room with it.

Why a single score under-reports it

A face occupies a small fraction of a frame. Score the picture as one image and the authentic majority outweighs the edited minority, dragging the average down. The result is a number in the middle band, which reads as uncertainty rather than as a finding.

The picture is not ambiguous. The measurement was averaged across a frame that is mostly real, and averaging is the wrong operation for a localised change.

14 11 9 17 91 12 8 13 10

The pattern that says edited rather than generated.

A swapped face on the region grid. One tile at 91 against eight in the low teens, from a headline score of only 61.

Reading the frame in overlapping regions fixes the arithmetic. Each region is scored on its own, so a face-sized area of generated pixels is measured against itself rather than diluted into a picture-wide average.

Three things people call a deepfake

They are made differently and they are found differently, so it helps to know which one you are looking at.

Generated outright
No camera anywhere. The whole picture came from a prompt, including a person who does not exist. On the grid this is the even pattern, every region reading high together.
A face swapped in
A real photograph with somebody else's face fitted onto the body, angle and lighting adapted to match. One region hot, neighbours quiet.
A face altered
The same person, changed. Expression edited, mouth reshaped to fit words never spoken, likeness nudged. In a still this looks like a swap. Across a clip the evidence is in how the face moves, which is a harder problem.

Checking a video

Pause on a clear, well-lit, front-facing frame and check that still. A swapped face carries the same local signal in one frame as it does across a thousand, so a single good frame usually answers the question.

Two cautions. Take the screenshot at full resolution rather than photographing the screen, because a photograph of a screen adds a second layer of processing that weakens everything. And this approach will not catch a clip that is genuine throughout and edited for one second, since you would have to pause on exactly that second.

Before you act on it

Face swaps turn up in situations where being wrong is expensive: a dating profile, a job applicant, a supplier's team page, an account claiming to be someone you know. The check is worth running and it is not sufficient on its own.

Run a reverse image search. A swapped face is often fitted onto a stock photograph, and finding the unedited original settles it immediately.
Ask for a second picture in a different pose. Producing a convincing swap from a new angle is more work than most people bother with.
Check the account rather than only the picture, including when it was created and what else it has posted.
Keep the original file rather than a screenshot, in case it needs to be looked at properly later.

Questions people ask

Can a face swap score low overall and still be fake?

Yes, and that is the normal outcome. A face is a small part of a frame, so the authentic majority pulls the average into the middle band. The region grid is what separates that from a picture that is genuinely uncertain.

Does this work on video?

Not directly, because the detector reads stills. Pause on a clear frame and check that. It handles a face swapped through the whole clip well, and it will not find a clip that is real except for one edited second.

The face looks perfect. Does that rule out a swap?

No. Good swaps are convincing in the middle of the face, which is where the model concentrates. Look at the hairline, the jaw against the neck, the ears, and whether the light on the face agrees with the light in the room.