Facial analysis methodology: what is measured and how evidence is graded.

Quick Scan and the complete three-view LookRange are available in Beta on supported devices. The current engines are LR-QS-BETA-1.0.0 and LR-MV-BETA-1.0.0. They produce deterministic Beta estimates from accepted inputs.

Browser and contract tests show that the pipeline can reject unsupported inputs, reproduce the same numbers for the same decoded evidence, and generate the documented result fields. Real-participant calibration and fairness evaluation are still required before those results can support population-level conclusions.

Swipe sideways to view the full table.

Selected photos are decoded in the browser and passed as in-memory bitmaps to an isolated worker. MediaPipe supplies landmark and expression-related signals; LookRange applies its own quality checks, normalized measurements, scoring rules, and result language.

MediaPipe supplies landmark inputs only. LookRange applies the scoring and interpretation rules, while deterministic code—not a language model—sets every numeric result.

• 1. File preflight. Confirm a real JPEG, PNG, or WebP file, then enforce size and pixel limits before full analysis.
• 2. Face and photo checks. Require exactly one usable face and check framing, pose, sharpness, exposure, occlusion, and expression risk.
• 3. Relative geometry. Convert selected landmark relationships into ratios normalized by face width, face height, or eye relationships, avoiding raw-pixel comparisons.
• 4. Single-photo result. Return a Photo Score, Provisional Structure Score, confidence, component readings, and retake guidance.
• 5. Three-view result. Validate Front, Left 45°, and Right 45° independently, then combine accepted evidence into Structure Score, LookRange, confidence, and a variance explanation.

These are current Beta v1 engineering thresholds, not universal standards. Internal sharpness and luminance values are normalized implementation signals and should not be read as camera specifications.

Swipe sideways to view the full table.

Quick Scan compares normalized geometry with frozen, deliberately broad engineering reference bands. Its Provisional Structure Score combines facial harmony, symmetry signals, and basic proportions. Photo Score then combines 57% provisional structure with 43% photo-condition support.

The multi-view engine normalizes selected horizontal measurements for view projection, weights each accepted view by confidence, photo conditions, and angle support, then uses weighted medians and bounded outlier handling. LookRange is centered on structural and photographic evidence, has a minimum half-width of 4.5 and a maximum half-width of 11, so the total interval can never exceed 22 points.

These formulas describe current software behavior. Their reference bands and weights are Beta design choices, not biological laws or proof of an objective beauty scale.

Swipe sideways to view the full table.

Calculated synthetic contract fixture. It contains no real person, training record, or customer feedback. The numbers below are calculated at build time by LR-MV-BETA-1.0.0 from frozen geometry and photo-condition inputs with controlled variation across the three views.

The main variance driver is sharpness. Confidence means evidence support inside this contract fixture; 97.4/100 is not an accuracy percentage or a probability that the appearance estimate is correct.

Swipe sideways to view the full table.

Score-contract stress test · synthetic input. One otherwise accepted view is assigned a Photo Score of 5 at the score-contract layer while the highest view is 97.1. A raw minimum/maximum approach would therefore report 5–97.1.

The current robust engine returns 91.7–100, within its 22-point maximum width. This demonstrates bounded outlier behavior at the software-contract layer and provides no accuracy evidence about photographs or people.

Protocol frozen; real-participant evaluation not started. Cohort sizes below are minimum targets, not enrolled or completed sample counts. Participants must explicitly consent to research use and remain separated by person between development and blind holdout sets. Eligibility is defined in the Terms of Use.

A failed blind holdout cannot be repaired and rerun under the same version while still being called independent validation. The engine must remain Beta, document the failure, change version, and use a new holdout.

Swipe sideways to view the full table.

The analysis cannot establish identity, health, ethnicity, personality, compatibility, employability, social status, or personal worth. It is not a medical, psychological, forensic, biometric-identification, or cosmetic-treatment service.

A single photo remains provisional. Even a valid three-view result describes only the supported images and the current Beta formula. Human attractiveness judgments include substantial individual preference, while camera distance and perspective can change photographed facial configuration and social judgments.

Stop repeated checking if the tool increases distress or compulsive comparison. Another score is not mental-health care.

The LookRange Method & Safety Lead is the accountable internal role for the facial-analysis protocol. The role owns input-quality rules, landmark normalization, the frozen scoring contract, calibration design, prohibited-use boundaries, public limitations, and corrections to this record.

The public work product is the versioned methodology, deterministic fixtures, threshold tables, safety language, and review history on this page. LookRange publishes a named biography only after the individual and credentials can be verified.

• Applied computer vision review. Reviews face detection, pose gates, landmark-derived ratios, projection limits, and browser-local processing boundaries.
• Measurement and calibration review. Owns fixture reproducibility, cohort separation, holdout gates, failure reporting, and version changes.
• Safety and content review. Maintains the eligibility terms, prohibited high-impact uses, neutral language, distress stop signals, and corrections channel.

Swipe sideways to view the full table.

Copy edits, algorithm updates, software tests, and population evidence remain separate change types. The public record keeps those categories distinct so the page date remains meaningful.

Swipe sideways to view the full table.

These sources explain underlying landmark capabilities, image-quality variables, perspective effects, and individual differences in face preference. They are not endorsements, and they do not supply LookRange scores, reference bands, weights, or validation results.

• Google AI Edge · Face landmark detection guide for WebDocuments the MediaPipe web task and its landmark, blendshape, and transformation outputs.
• NIST · Face Analysis Technology Evaluation: QualityDescribes capture-quality factors such as focus, illumination, distortion, pose, and expression in a face-recognition context.
• Bryan et al. · Perspective distortion from interpersonal distanceReports that camera distance and perspective can affect photographed faces and related social judgments.
• Germine et al. · Individual aesthetic preferences for facesShows substantial individual variation alongside shared patterns in face preference.