Measure the geometry.
Understand the photograph.

A measured feature is evidence.
An aesthetic preference needs validation.

A face can be described precisely without assuming there is one ideal face. Our method makes three separate claims: a geometric construction can be reproduced; an AI can describe the appearance of a photograph; and a predictive model can be evaluated against a particular panel of human raters. Each claim needs its own evidence.

We do not treat symmetry, a golden-ratio fit or a community PSL label as a universal definition of beauty. Nor does learning from human ratings remove subjectivity: it learns the preferences, context and possible biases of the people and images represented in those ratings.

Our current public description is research-informed facial geometry and photo analysis. Human-aligned attractiveness prediction is a development objective. We will use “validated against human ratings” only for a named model version and evaluation that we have actually completed and published.

Different outputs. Different standards.

Photo suitability comes first

Use one clearly visible face, sufficient detail and even exposure. Front-view measurements need a front-facing photograph; profile observations need the appropriate side view. Pose, expression, blur, occlusion, glare, camera distance and clipping may limit individual readings. Checks vary by tool: a successful face detection does not certify every feature as measurable.

Our existing tools use face landmarks and image-quality checks on your device. They may reject an unsuitable image or withhold an affected result. We do not claim a clinical 3D reconstruction, a dedicated detector for every possible obstruction, or a statistical reliability percentage from these checks. An optional AI interpretation can still miss a limitation.

A scale must name what it measures.

Version 1.1 preserves these existing specialist formulas and introduces a separately labeled AI reference score for single-portrait general analysis. Geometry tools keep their own definitions. Subjective products follow the evidence and interpretation rules below.

Explore the actual constructions: golden ratio, symmetry, eye-corner angle and facial ratios.

Which tools use which method?

The homepage, Full Face Report, Face Rating, AI Attractiveness, qualitative Hunter Eyes review and portrait comparison use the six-dimension assessment. Eye angles and other geometric readings keep their measured units.

PSL Scale, Camera PSL, True Adam, True Eve, golden-ratio fit, symmetry and outline classification retain their independently defined scales or constructions. The standalone hairstyle changer is a visual styling simulation, and Roast is entertainment; neither delivers a validated attractiveness evaluation. Styling recommendations within a face report follow the framing and capture dimensions.

Your geometry score out of 100.

The free score summarizes measured symmetry and proportions. It is calculated entirely on your device and is available on your sharing card. It describes agreement with our disclosed geometric references; the AI assessment separately considers visible presentation across six dimensions.

These reference values and weights are our authored product rules. Pose, perspective, expression and landmark detection can affect the result. They are not learned human-rating weights. A high geometric fit does not establish a universal beauty ranking. AI adds photo evidence for surface appearance, framing, expression and capture, alongside relationships and contours. Its six-dimension score is a separate assessment rather than a numerical correction of this index; greater accuracy against human ratings has not yet been independently established.

Six dimension scores.
One result out of 100.

The free scan computes geometric measurements on your device. Choosing an AI report adds six photo-based dimension scores and, when every dimension is assessable, an overall Get Face Report Score. This is an AI reference assessment of the photograph, using our authored rubric. It has not been calibrated to an independent human-rating panel.

1. Measure on the device. Detect the supported landmarks and calculate the applicable ratios and angles. These readings provide evidence for the AI interpretation.
2. Review the confirmed photo with AI. For a chosen AI report, the model receives the confirmed images, relevant measurements and report preferences. It evaluates each of the six dimensions using the criteria below.
3. Assign six integer scores from 0 to 100. Each score needs a visible observation and evidence. Limited findings explain the uncertainty; genuinely unavailable dimensions receive no score.
4. Calculate the total in code. Total = round((relationships + contours + surface + framing + expression + capture) / 6). All six dimensions have equal weight. The server computes this value rather than accepting a separate overall number invented by the model.
5. Explain what the reader can test. Each dimension includes its limitation and a practical comparison. A photo change can affect the score without implying an anatomical change.

A worked example

These numbers are illustrative and show the calculation, not an accuracy result.

The shared rating anchors

These anchors describe the visible presentation within each dimension. An image that cannot be assessed is different from an assessable image with a low score. The numbers are not percentiles, probabilities or a universal ranking of people. Photographic presentation is explicitly part of the total, so different light, focus or framing can change the result.

Historical v1.0 reports retain their original observation format. Two-person comparisons do not merge two people into one total. The independent human-rating calibration and repeatability evaluation below remain planned work.

Six dimensions for a useful photo analysis.

For judgments without a fixed measurement definition, we organize observations into six dimensions. These are our editorial assessment categories. They are authored scoring categories for the photograph, rather than independently validated traits or learned model-attribution coefficients. Existing tool-specific category layouts may group them differently.

The current product total uses the disclosed equal-weight mean. A future human-rating predictor would need separate training and calibration against a disclosed target panel. Our current report identifies what works and what can be compared without treating distinctive features as defects.

Observation → evidence → limitation → comparison.

A useful explanation connects a claim to something the reader can inspect. Measured claims refer to supplied anchors and calculations. Photo observations identify the visible feature or condition. Advice proposes a reversible experiment, with a way to decide whether it helps.

Reports distinguish measured findings, visual observations and advice. Missing evidence should produce a limited or unavailable finding. A frontal photograph cannot support an invented nose-projection angle; a lighting shadow cannot establish skin health.

Attention maps indicate where a particular model attends. Component decomposition describes how a particular prediction is assembled. Neither, by itself, establishes why people find a face appealing. Our current VLM narrative is not a computed component-attribution map.

Two separate concepts must also stay separate: human disagreement is variation between raters; model uncertainty is uncertainty or error in the prediction. We currently publish neither a predicted human-rating distribution nor a calibrated uncertainty interval for an individual photo.

What must happen before a stronger claim.

The following is our proposed validation protocol, not a completed study or an accuracy announcement. We plan to compare the existing geometric baseline, an off-the-shelf VLM baseline and a dedicated attractiveness predictor such as FPEM, where permissions allow.

1. Define the task and permissions. Evaluate perceived attractiveness of an adult’s photograph under a stated rating question. Obtain appropriate consent and commercial rights for photos, annotations, model weights and training data. A citation or repository code license is not a substitute.
2. Build independent data. An initial planning target is 500 consenting adult participants, two photos each and at least 20 independent ratings per photo. These are recruitment targets, not achieved sample counts or a power justification. Final size must reflect the measured variability and planned subgroup comparisons.
3. Prevent leakage. Keep all photos of one person in the same partition. A proposed 60/20/20 identity split separates model development, calibration and final testing; check external model training overlap where possible. Lock the final test set, prompts, versions and selection rules before comparing candidates.
4. Keep the human ratings independent. Randomize presentation, use the same anchored rating question, hide model outputs, include quality checks and report rater composition and agreement. Human panel averages describe that panel, not universal human preference.
5. Report complementary metrics. Publish Pearson and Spearman correlation for agreement and ranking, MAE/RMSE and calibration plots for score error, and identity-level bootstrap intervals. Compare against baselines and rater agreement rather than advertising one large correlation as “accuracy”.
6. Test disagreement separately. If training a distribution head, compare its predicted spread with actual individual ratings and evaluate proper distribution scores such as negative log-likelihood or CRPS. A spread estimate is not an error bar for the predicted mean. Any error interval needs its own coverage test.
7. Check stability and generalization. Measure same-input repeatability and same-person variation across photos; examine controlled pose, distance, expression and light changes. Investigate errors across consented, self-reported evaluation groups with adequate sample sizes; do not infer protected traits from faces to create those groups.
8. Publish before making the claim. Report sample sizes, exclusions, dataset rights, protocol, model/checkpoint and prompt versions, preprocessing, calibration, errors and limitations. A new model version needs reevaluation. We have not yet set empirically justified numeric release thresholds or completed this evaluation.

A human-rating predictor, a “human consensus” indicator, a population percentile and a numerical contribution chart remain unavailable until their respective evidence exists. Production release also requires a separate check of cost, reliability, privacy and suitability for the users the model will actually serve.

What informs the method.

These primary sources inform our evaluation design. Their benchmarks belong to their authors’ models and datasets. They do not transfer to Get Face Report through citation. Review date: 30 October 2026.

H-ARN reports a Pearson correlation of 0.9333 with average ratings on SCUT-FBP5500 and 0.838 agreement with rater-disagreement statistics. DisFace3DNet reports 0.8904 ± 0.0063 across five folds. These are different evaluation results, not interchangeable accuracy percentages or our product’s performance. H-ARN’s text also varies between variance and standard deviation terminology; any reproduction must establish the exact target and computation.

Research photo datasets may have research-only terms or additional restrictions. Our plan does not assume commercial permission for SCUT-FBP5500, MEBeauty, LiveBeauty or a downloaded checkpoint. We will verify the applicable terms and portrait permissions before any use.

A useful result has a clear boundary.

This service does not identify people or determine their health, personality, intelligence, trustworthiness, ethnicity, gender, hormones or precise age. Styling images are generated concepts, not evidence of future physical change. Medical and irreversible treatment decisions are outside this method.

Current free geometric scans remain on-device. Choosing a paid AI report sends the confirmed photos for processing under our privacy policy. The research plan does not authorize collecting training photos or human ratings from ordinary scans, and does not introduce a new data-retention policy.

Method owner: William Zhang, Billions Tech. Questions or corrections: contact@getfacereport.com.