How it measures
A scan is one front-facing photo, measured in your browser. This page says what is measured, how far to trust each number, and what the scan cannot see.
Finding the face
A face-mesh model places 478 points on the photo. From them the scan works out how your head is turned and tilted, where the midline of your face runs, and the scale: an adult iris averages about 11.7 mm across, give or take half a millimetre, so its size in the photo gives pixels per millimetre and every distance can be a few percent off.
The model runs twice, on the photo and on its mirror image, and the two answers are averaged. That removes a small left-right bias in the model itself.
Comparing the two sides
The mesh is fitted to an average face, and it under-reports real differences between sides. So the differences are measured from the photo: the skin around a feature on one side is matched against the mirrored patch from the other side, and the shift that lines them up is the asymmetry.
To check this, features in two test photos were moved by a known 2 to 4 mm and the scan was asked to read the change back. The mesh alone missed most of it. The photo-matching method recovered most of it.
| Feature moved | Mesh alone | This scan |
|---|---|---|
| Brow height | 44% | 76% |
| Eye opening | 25% | 97% |
| Mouth corner | 67% | 102% |
| Nose off midline | 6% | 86% |
| Chin off midline | 20% | 98% |
Share of the injected difference read back, averaged over two computer-generated test faces. Brow readings run low, so a real brow difference is likely a little larger than shown.
Each structure reading is set against the size of difference that observers start to notice in published studies: about 2 mm for eyelid position, 3 mm for the mouth corner, 3.5 mm for the brow, 4 mm for the nose tip and 6 mm for the chin.
Reading folds, shadow and skin
The face is resampled upright at four pixels per millimetre and converted to a color space built around human vision. Folds and lines are measured as how much darker they are than the skin beside them, which cancels out how bright the photo is. Under-eye shadow is the lightness difference between the under-eye area and the cheek. Texture and evenness are the amount of light-and-dark variation at fine and at patch scales.
These readings have no clinical scale yet. The level marked as standing out is a working value: about where the top fifth of thirty test faces fell. They are best used to compare your own scans taken the same way, and not yet to compare you with anyone else.
Skin tone uses the individual typology angle, a standard measure in dermatology research, which sorts skin into six groups from very light to dark. It matters here because laser risk rises with skin pigment. A provider judges that in person, by how your skin tans and heals, not from a photo.
How steady the numbers are
Five test photos were each re-processed thirteen ways: rotated, resized, darkened, brightened, re-compressed, blurred, shifted and mirrored. Structure readings moved by less than 0.5 mm in 95% of cases. The scan shows a wider margin than that on every finding, about 1 mm, because real retakes vary more than re-processed copies, and that has not been measured yet.
Movement, from the photo set
The full photo set adds a big smile, a surprised face and an “eek” face. Each is lined up with the relaxed photo on the parts of the face that don’t move with expression (the bony nose, the eye corners, the sides of the face), and each brow and mouth corner is followed from one photo to the other. On simulated expressions this read back 91% to 100% of the true movement, and the left-against-right comparison was within 10 points.
Simulated expressions are the easy case: a feature moved without changing shape. A real smile bunches the cheeks and shows teeth, so real readings will be rougher, and a reading the scan can’t follow is left out. The level at which uneven movement is marked as standing out, one side moving less than 70% as far as the other, is a working value.
The looking-down and side photos aren’t measured. They are there to compare by eye, and to bring to a consultation: they show jowls, the jawline, temples and cheek contour, which a front photo can’t.
What the capture checks
- Head turned no more than 6 degrees. Past 10 degrees the two sides are not compared at all.
- The whole face in the photo, with the head not tilted far back or leaning over.
- A resting face. A smile lifts the cheeks and mouth corners unevenly and deepens every fold, so those readings are left out when you smile.
- Light within a few points of even across both cheeks, with skin neither blown out nor too dark.
- Enough detail: at least 2 pixels per millimetre for structure and 3 for fine lines and texture.
Color, previews and comparisons
Skin color. Cheek and forehead skin is read from the middle of its brightness range, so shine and shadow drop out. A phone’s white balance moves skin color by more than one foundation shade, so a photo with plain white paper beside your face corrects it: the paper should be neutral, so whatever tint it shows belongs to the light and is divided out. Depth groups and the undertone cut-offs (by the hue of the skin color) are working values that have not yet been checked against a makeup artist’s eye.
Skin close-ups. A face photo puts a pore across one or two pixels, too few to measure. Close photos taken about 20 cm away give 10 to 20 pixels per millimetre. Scale comes from the iris in each photo, and each area is resampled to 10 pixels per millimetre, so close-ups taken at different distances are measured alike. Pores are found as small, round, dark spots at several sizes; fine lines as long, narrow dark ridges; brown spots by the red channel’s absorbance, which tracks melanin; redness by the ratio of red to green, in patches against the skin around them. On synthetic skin with known pores and lines, pore counts matched exactly, pore sizes came within about 10%, and fine-line length within 10%. None of this has yet been checked against a dermatologist’s grading or on repeat photos of real skin, so treat the numbers as a way to follow your own skin.
Face shape. Upper-forehead, cheekbone and jaw widths and face length are compared with the average face the mesh is built on. Shape words are a styling convention; most faces sit between two.
Treatment previews move points of the face mesh and retouch the skin under it on your device. Each slider’s full range is a modest, commonly seen change, such as about 2.5 mm at the brow tail or 1.5 mm of lip height. They show the kind of change, not your result.
Before and after. The later photo is turned and scaled so the bridge of the nose, the eye corners and the sides of the face line up with the earlier one. A reading only counts as changed when it moved by more than the amount it moves between two photos of an unchanged face. Different light moves shadow, fold and skin readings, so compare photos taken the same way.
What it cannot see
- Depth. Temples, cheek projection, jawline, jowls and chin projection need an angled view or a depth sensor.
- Movement in conversation. The photo set captures held expressions, not how your face moves as you talk.
- Cause. The scan sees a shadow under the eye; it cannot tell a hollow from pigment from thin skin.
- Makeup, filters and retouching change every skin reading. Heavy hair across the face blocks the comparison.
- The scan has been tested on computer-generated faces only. It has not been validated against clinicians or on a wide range of real faces.
Sources
- MediaPipe Iris, Google Research: iris diameter of 11.7 ± 0.5 mm as a scale reference.
- Discriminative thresholds in facial asymmetry: a review of the literature: the differences at which observers notice asymmetry.
- Estimating skin tone and effects on classification performance in dermatology datasets: the typology angle formula.
- Individual typology angle, Haut.AI documentation: the six tone groups from very light to dark.