AI face analysis: the complete guide

AI face analysis is the automated measurement of facial geometry from a photograph: software detects landmark points on the face, then computes distances, angles, ratios, and symmetry from those points. It measures structure — it does not judge worth.

What AI face analysis is

Every measurable claim about a face — how symmetric it is, how wide the jaw sits relative to the cheekbones, whether the eyes are close-set or wide-set — reduces to geometry: points, distances, and angles. For most of the last century those measurements belonged to anthropometry, taken by hand with calipers in clinics and research labs. AI face analysis automates the same measurements from a photo.

The pipeline has two stages. First, a computer-vision model locates landmark points on the face — the corners of the eyes, the tip of the nose, the border of the lips, the angle of the jaw, and hundreds of points in between. Second, ordinary geometry turns those points into measurements: the distance between the inner eye corners, the angle of the jawline, the ratio of midface height to lower-face height. The AI does the locating; the analysis itself is arithmetic you could check by hand.

That distinction matters. A landmark position is an objective, repeatable measurement. What a measurement means — whether a given ratio is "good" — is an interpretation, and honest analysis keeps the two separate. The rest of this guide walks through both halves: the measurement pipeline that produces the numbers, and the limits of what those numbers can support. It is written to be useful whether you are simply curious about your own face, comparing analysis tools, or preparing for a consultation and want to arrive with vocabulary and measurements rather than vague impressions.

What AI face analysis measures

A full analysis covers the face region by region. Each of these areas has its own set of landmarks, measurements, and reference ranges:

• Face symmetry — how closely paired left and right landmarks mirror each other across the facial midline, measured per region rather than as one number.
• Golden ratio and proportions — how facial ratios compare with classical reference values such as phi (1.618), and what that comparison can and cannot tell you.
• Face shape — the overall contour classification (oval, round, square, heart, oblong, diamond) derived from width measurements at the forehead, cheekbones, and jaw.
• Eye analysis — canthal tilt, intercanthal distance, eye width and height, and how eye spacing compares with the classical one-eye-width guideline.
• Nose analysis — nasal width, length, tip projection, and the nasofrontal and nasolabial angles visible in profile.
• Jawline analysis — gonial angle, jaw width, chin projection, and the definition of the mandibular border.
• Lip analysis — upper-to-lower lip ratio, mouth width relative to the nose and eyes, and philtrum length.
• Cheekbone analysis — bizygomatic width, the position of maximum cheek projection, and midface fullness.
• Side profile analysis — the sagittal view: facial convexity, chin and nose projection, and the relationships orthodontists measure in cephalometry.
• Facial harmony score — a composite summary of symmetry, proportion, and balance, built so every contributing number is traceable to specific landmarks.

How landmark detection works

The core technology is a landmark-detection model — a neural network trained on large sets of annotated face photos to predict where specific anatomical points sit in a new image. Refrakt uses a 468-landmark face mesh, which maps the face as a dense grid of points, each with x, y, and z coordinates. Here is the pipeline step by step:

1. Face detection. The model first locates the face in the frame — a bounding region that tells the next stage where to look.
2. Mesh fitting. A second network predicts the positions of all 468 landmarks inside that region: 3D coordinates covering the eyes, brows, nose, lips, cheeks, jawline, and the full face contour. Because the mesh includes depth (z) estimates, it captures projection — how far the nose or chin extends forward — not just a flat outline.
3. Pose normalization. Head tilt and rotation are estimated from the mesh and corrected, so a slightly turned head does not register as asymmetry.
4. Scale normalization. Pixel distances mean nothing on their own — a face 50 cm from the lens spans more pixels than one at 80 cm. Measurements are therefore expressed as ratios of face size or converted to real-world units using a known reference.
5. Measurement. With a normalized mesh, the geometry is computed: distances between landmark pairs, angles at feature corners, ratios between regions, and left–right differences across the midline.

On a good frontal photo, modern landmark models place points within a few pixels of where a trained human annotator would. That is accurate enough for reliable ratios and angles — and it is repeatable, which hand measurement with calipers never fully was.

Measurement categories at a glance

What it can tell you — and what it can't

What it can tell you

• Where your measurements sit. Objective numbers for symmetry, proportions, and feature geometry, and how they compare with population reference ranges.
• What drives an impression. If your jaw reads as wide or your eyes as deep-set, the analysis can point to the specific measurements behind that impression.
• What would change. Because measurements are geometric, they can be recomputed under a simulated change — the basis of 3D facial simulation.
• Consistency over time. Repeated scans under controlled conditions can track changes from orthodontic work, weight change, or procedures.

What it can't tell you

• Whether you are attractive. Attractiveness is a judgment made by people, shaped by culture, era, and taste. Geometry correlates with it; geometry does not define it.
• Anything about character. Claims that facial structure reveals personality or ability are physiognomy — a field discredited over a century ago. No landmark measures kindness.
• How you look in motion. A static scan misses expression, animation, and voice — much of what actually registers when people meet you.
• Medical conclusions. Measurements are informational. Diagnosis and treatment decisions belong with clinicians.

Getting reliable results

Because the analysis is only as good as its input, a few capture habits make the difference between numbers you can trust and numbers you should ignore:

• Distance. Shoot from roughly arm's length or further. At 30 cm, perspective distortion enlarges central features (nose, mouth) relative to peripheral ones (ears, jaw width); at 1 m and beyond the effect largely disappears.
• Pose. Face the camera squarely with a level head. Even a 5-degree head turn shifts apparent left–right proportions, and pose correction can only compensate so far.
• Lighting. Even, diffuse light from the front. Strong side lighting casts shadows that landmark models can misread as contours.
• Expression. Neutral and relaxed. A smile changes mouth width, cheek projection, and eye height — real geometry, but not the resting geometry most reference ranges describe.
• Consistency. If you want to compare scans over time, capture them the same way. A measured change is only meaningful when the capture conditions did not change with it.

This is why Refrakt uses a guided scan rather than accepting arbitrary uploads: the app checks distance, pose, and lighting before measuring, so variation in the numbers reflects faces, not photography.

Privacy considerations

A face scan is biometric data, and it deserves more caution than an ordinary photo upload. Before using any face-analysis service, it is reasonable to ask four questions:

• Transmission. Is the photo encrypted in transit (HTTPS/TLS)?
• Retention. How long is the image stored, and can you delete it? Some services keep uploads indefinitely; others process and discard.
• Training use. Is your face used to train the company's models? This should be disclosed, and ideally opt-in.
• Sharing. Is facial data sold or shared with third parties? Under GDPR, biometric data used for identification is special-category data with strict processing requirements.

Refrakt's position: scans are processed for your analysis, facial data is not sold, and deletion is available on request. Details are in the privacy policy.

Frequently asked questions