How AI face analysis apps actually work
A face analysis app finds landmark points on your face in a photo, measures the distances, angles and symmetry between them, and compares those measurements with a reference group. The result is only as good as the photo, the reference data and the scoring rules, which is why a trustworthy app explains all three.
Step 1: finding landmarks
A face-detection model looks at the photo and places points on recognizable features: the edges of the eyebrows, the corners of the eyes, the nose, the lips, the jaw and the outline of the face. MaxxLook’s detector maps up to 478 landmark points on your face. Where the model doesn’t provide points, such as the ears and neck, they are estimated, marked as estimates and weighted lower.
Different photos do different jobs. In MaxxLook, the front photo provides the landmark geometry and the smile photo provides the smile measurements, while the profile and quarter views add skin readings and let the app check that a measurement repeats across photos. See what a MaxxLook scan captures.
Step 2: turning points into measurements
Once the points are placed, geometry does the rest. The app computes width and height ratios, how far left and right features differ from each other, head and feature tilt, spacing, and how the face divides into thirds. MaxxLook computes 170+ measurements across 12 sections this way.
Skin measurements work differently: they are computed from pixels rather than points. Examples are texture, pore visibility, how much redness varies across the face and under-eye shadowing. These measure variation within your own face. A responsible app does not rate skin tone or how light or dark your skin is.
Step 3: comparing with a reference group
A raw measurement means nothing on its own. An eye-spacing ratio of 0.46 is neither good nor bad until you know what that ratio looks like across other people. So apps compare each measurement with a reference distribution, and three choices decide how honest that comparison is: which group you are compared with (age, sex, population group), how the reference was built, and which direction counts as better.
MaxxLook builds its reference ranges by running the exact same pipeline over a large, openly licensed face-image collection, FairFace (CC BY 4.0), keeping only near-frontal faces and computing each measurement’s distribution per demographic group. Skin and smile stay exact-value-only, because a collection of web photos can’t provide baselines for lighting-dependent or expression-dependent measurements. The full derivation is on the transparency page.
Step 4: turning comparisons into a score
The last step is a set of scoring rules, and they are a design decision, not a fact of nature. In MaxxLook, each measurement becomes a 0 to 100 reading against its reference range, an area score pools the readings in that area, and the overall score is a weighted blend in which symmetry and proportion count most. An area with nothing measured is left out instead of being padded with a made-up middle value.
Where accuracy breaks down
- Camera distance and lens. Close-up selfies with a wide lens stretch the middle of the face and change proportions.
- Head tilt and turn. A slightly rotated head changes symmetry measurements immediately.
- Expression. A smile lifts the cheeks and changes the mouth and eye measurements.
- Lighting. Uneven or colored light changes skin measurements the most.
- Reference data. No collection of photos represents everyone, and most are neutral-faced, so some measurements have no defensible baseline.
- Scoring rules. Two apps can measure the same face and score it differently because they made different choices in step 4.
What a good app should tell you
If an app shows you a number but never explains how it was made, you can’t judge it. Use the nine-question checklist for choosing a face analysis app before you scan.
Common questions
Are AI face analysis apps accurate?
They are consistent about geometry and much less certain about meaning. Landmark measurements from a well-taken photo are repeatable, but camera angle, lighting and expression all move them, and any score depends on the reference group and scoring rules the app chose. Look for apps that publish their method and report changes below their noise floor as “within noise”.
Is face analysis the same as face recognition?
No. Face analysis measures features in a photo, such as proportions and symmetry. Face recognition matches a face to an identity. They are different tasks, so check what any app stores and for how long before you scan.