How to spot a face swap
A face swap moves one person's face onto someone else's head. The face itself often looks fine. The trouble is at the edges, where it has to meet a head that was never its own.
Updated
What counts as a face swap
In a face swap, the face comes from one person and everything else (hair, head shape, body, background) comes from another photo. Tools range from one-click phone apps and live webcam swappers to AI image editors that will redraw a head on request.
It helps to know the neighbours, because they leave different traces:
- AI face edits change a face that was already there: a new expression, a younger look, different eyes.
- Fully generated faces belong to nobody. The whole image was made from scratch.
- Talking-head and lip-sync videos animate a face from a photo or move its mouth to new audio.
Where swaps leave traces
A swap tool has to blend a pasted-in face into a head that wasn't made for it. The blend is where most mistakes live: along the jaw, around the hairline, near the ears and down the neck.
The face is also often processed at a different resolution from the rest of the photo, so it can look slightly sharper or slightly softer than the hair and skin next to it. And it carries the lighting of its original photo, which may not match the new scene.
Signs to check, in order
Zoom in to full size on the best copy you have. Then work through these:
- The hairline and jaw edge: a soft halo, a smear, a faint seam, or a sudden change in sharpness.
- Sharpness: is the face crisper or blurrier than the hair, ears and neck at the same distance from the camera?
- Skin: does the face match the neck, ears and hands in tone, texture and age?
- Light: do shadows on the face fall the same way as shadows on the body and in the room?
- Things that cross the edge: glasses arms, earrings, hair strands and hands in front of the face often come out melted or cut off.
- Teeth and eyes: teeth that merge into one band, or catchlights in the eyes that don't match each other or the light source.
- Proportions: a face that seems slightly too small or too large for the head, or sits a little off-centre.
- Angle: swaps are most convincing face-on. Turned heads, profiles and tilted faces are harder to fake.
Signs that mislead
Plenty of genuine photos show some of the signs above. Before you read too much into one, rule out the usual suspects:
- Beauty filters and retouching smooth skin and reshape faces, and can leave soft edges at the jaw.
- Portrait mode on phones blurs the background with software, and often gets hair and ear edges wrong.
- Heavy compression adds blocky edges and colour smears everywhere, faces included.
- Low light makes phones denoise hard, which can turn skin waxy.
- Makeup, contouring and stage lighting change how skin and shadows look.
Check the photo's story
Swaps are often built on a photo that already exists, with only the face changed. That makes a reverse image search unusually useful: if it finds the same body, outfit and background with a different face, you have found the original.
Also check where the image first appeared, whether the person's own accounts or other photos of the same event show the same scene, and whether the photo says something about the person that would be very convenient for whoever posted it.
Use a face swap detector
Looking only gets you so far: good swaps hide most of the visible signs. Checkobot's face analysis was trained on face swaps and AI face edits from many tools. It looks at a wide crop around the face, including the jaw, neck, hairline and some background, because that is where swaps have to blend. A whole-image analysis runs alongside it, and the result says which one fired and why.
For a useful check, use the largest copy you can find. The image needs to be at least 256 pixels on its short side, and the face at least 64 pixels across for the face analysis to run. If there are several people, it checks the largest face, so crop the image to check someone else.
If you find one
Remember who is usually harmed. The face in a swap generally belongs to someone who never agreed to it, and the body often belongs to someone else again. Don't forward it to prove a point, and don't name or blame the people in it.
If it's being used to harass, scam or impersonate someone, report it to the platform it's on. If it's of you, keep the links and screenshots with dates before you report, in case you need them later. And treat any detector result, ours included, as a signal rather than proof.
Questions
Can a face swap be impossible to spot?
By eye, yes: the best swaps on a clean, face-on photo can hide every visible sign. That's why the source, a reverse image search and a detector matter more than any single visual clue.
What about deepfake videos?
Checko checks images. You can check a still frame from a video. Talking-head and lip-sync videos from new engines are Checko's weakest area, so a No AI detected on a frame means even less there.
Will a beauty filter look like a face swap?
Sometimes, to people and to detectors. Strong filters and heavy retouching smooth and reshape faces in ways that can look like AI editing.
Which face does the face analysis check?
The largest one, if it's at least 64 pixels across. To check a different person in the photo, crop around their face and check the crop.