Step 1: the browser finds facial landmarks
After you select a photo, a face-landmark model identifies a mesh of 478 points around visible facial structures. On AI Face Analyzer, the free preview loads this model only after you choose an image, and the analysis runs inside your browser. The image does not need to be sent to our server.
Landmarks are coordinates, not identity. We do not compare them to a face database, name a person, or use them for authentication. The goal is to measure relationships within this one image.
Step 2: pose and image quality are checked
Measurements are unreliable when the head is turned strongly, tilted, too small in the frame, blurred or badly lit. A responsible tool should reject poor inputs instead of producing confident-looking output from weak evidence.
Our quality gate requires one adult, front-facing face with sufficient detection confidence. It also checks image dimensions, tilt, yaw-related asymmetry, brightness and contrast. Warnings explain what to change before trying again.
Step 3: coordinates become ratios
Raw pixel distances depend on image size, so they are not useful by themselves. The engine normalizes distances against stable references such as face width or height. This creates scale-independent ratios for eye spacing, face length-to-width, nose-to-face width and vertical thirds.
For bilateral balance, corresponding left and right points are compared around an estimated midline after basic pose normalization. The result describes the photograph—not an immutable property of the person.
Step 4: measurements are translated into plain language
The free tools use deterministic rules to explain what each ratio means. For example, the face shape analyzer looks at length-to-width, forehead, cheek and jaw relationships, while the symmetry test focuses on bilateral landmark differences.
Labels such as oval or square are approximations. Real faces often sit between categories, so the result is best treated as a styling shorthand rather than a diagnosis.
Step 5: the paid report adds visual context
A landmark mesh cannot describe everything visible in a photograph. With explicit consent after payment, the paid report sends the image once to a vision model through zero-data-retention routing. The model receives the deterministic measurements as evidence and returns a strict, validated report structure.
The report is constrained to six appearance and presentation dimensions. It cannot discuss protected or sensitive inferences, medical topics, procedures or identity. We retry transient generation errors twice; if the report still cannot be delivered, the system attempts to return the payment automatically.
What “zero data retention” means here
AI Face Analyzer does not store the original paid photo on its server. The provider request is restricted to endpoints that accept zero-data-retention routing and deny data collection. The browser may temporarily keep the selected file in IndexedDB for up to two hours so a Stripe redirect can resume. It is deleted after successful upload or expiry.
Derived measurements and the generated report are stored for 30 days so you can revisit the result and download a PDF. After that, a protected cleanup task removes those facial derivatives while financial order records are retained for accounting and dispute obligations.
Where uncertainty remains
- Landmark placement can vary with occlusion, makeup, hair, glasses and image quality.
- Two-dimensional ratios do not describe depth.
- Face-shape categories overlap and depend on chosen thresholds.
- A visual model’s appearance judgment remains subjective.
- A report based on one photograph cannot represent how a moving, expressive face appears in person.
A better mental model
Think of AI face analysis as an annotated portrait review: part ruler, part photography critique, part styling brainstorm. It can expose useful relationships and suggest experiments. It should never be treated as biometric truth, medical guidance or a ranking of people.