How Do Face Scan Tests Work? Landmarks, Selfies, and Scores

The pipeline in one paragraph

A face scan test takes your selfie through three stages: it detects a face, it places a set of landmark points on that face, and it runs a formula over those points to produce a number. Stage one and stage two are ordinary, well-established computer vision. Stage three is where the entertainment begins, because the formula is an editorial choice, not a measurement.

Understanding which stage is which is the whole point of this article. Once you can separate "the software located your jaw accurately" from "the number means something", the results become much easier to read correctly.

Stage one: finding a face

The first job is deciding which pixels in the photo belong to a face at all. This is a detection problem, and modern detectors handle it well: they work on a crowded group photo, they cope with glasses and beards, they tolerate a phone held at an angle.

What they need from you is a photo where a face is actually visible. Common reasons a scan fails outright:

• The face is too small. A full-body shot gives the model very few pixels to work with.
• The lighting is behind you. A backlit silhouette removes the contrast the detector relies on.
• The face is heavily occluded. A hand, a mask, or a hair across the face can push it below the detection threshold.
• The image is not a photo of a person. Obviously. But it is worth stating, since people test the software.

If the scan returns an error rather than a score, detection is usually the reason. That error is the one honest piece of feedback in the entire pipeline: it means the software genuinely could not find a face, and it did not invent one.

Stage two: landmark points

Once a face is found, the model places a defined set of points on it. These are landmarks: precise pixel coordinates for the corners of the eyes, the tip and bridge of the nose, the jawline, the mouth, and the eyebrows.

The output of this stage is not a description and not a judgement. It is a table of numbers. A typical entry looks like (312.4, 188.9) — an x and a y position on the image.

Two things follow from that, and both matter:

• The points are geometric, not semantic. The model knows that a point sits at the outer corner of your left eye. It has no idea what an eye means, and no capacity to draw conclusions from one.
• The points scale with the photo. If the same face is photographed close up and far away, the coordinates describe the same geometry in different units. Anything the formula wants to use has to be normalised by a reference length first — usually the distance between the eyes.

When a site advertises a landmark count, that number is simply how many points the model places. A larger count means a denser mesh of coordinates, which can make the geometry slightly more robust to noise. It does not mean the points are more "truthful", and it says nothing at all about whether the final score is meaningful.

Stage three: the formula

This is where a landmark mesh becomes a body count, a loyalty score, or a financial destiny. The transformation is arithmetic: distances, ratios, angles, and weighted sums, written by a developer.

Here is the part worth being blunt about. The formula is a design decision, not a discovery. Nobody extracted it from data about real people's lives, because the data required to do that does not exist. Someone decided that a wider ratio should push the score up and a narrower one should push it down, chose coefficients, and named the result something memorable.

That is not a secret conspiracy — it is the normal way entertainment products are built. A fortune cookie is also written by someone. The problem only starts when the output is presented as a measurement rather than a script.

Because the formula is fixed, the tests are at least internally consistent. The same face photographed the same way will produce the same score. This consistency is easy to mistake for accuracy: if a score stops moving, it feels anchored to something real. It is anchored to the formula, which is stable for the same reason a stopped clock is right twice a day.

What a selfie can and cannot carry

A single front-facing photo is a narrow slice of information. Worth knowing what is actually in it:

The left column is genuinely informative for other purposes — face geometry is used in animation, in accessibility tooling, in photo editing, and in medical imaging research where a clinician interprets the result. The right column is the entertainment layer. Nothing in the left column becomes the right column by adding more points or more compute.

Why the same face can score differently

People often notice their score changing between attempts and conclude the software is broken. It is not; it is doing exactly what it was designed to do, and the inputs changed.

• Pose and angle. A slight head tilt changes the projected distances the formula reads.
• Lighting and shadow. Shadows shift where the contours appear to be, which moves the landmark positions.
• Expression. Smiling changes the mouth and jaw geometry, and some formulas read the smile as a measurement.
• Filters and beautification. Automatic smoothing, face slimming, and eye enlargement are applied before the image reaches the model.
• Compression. A heavily compressed screenshot loses edge detail, and with it some precision.

If you want scores that are at least consistent with each other, control for those variables: front-facing, even lighting, neutral expression, no filter, no screenshot. You will get a more stable number. You still will not get a more meaningful one.

The honest summary

Three stages, and only the first two are engineering. Face detection and landmark placement are competent, well-understood techniques that do a real job. The score that comes out the far end is an authored formula with an entertainment purpose.

Both halves of that sentence can be true at once, and both are. The technology is real. The claim is not. If you want to try the tests, the honest reason to do it is that they are fun to compare with friends — and they are. Just keep the score in the column it belongs to.