How Face Rating works

MediaPipe Face Landmarker finds 478 facial points in your browser. The tool measures proportions and angles, compares them with our chosen references, and returns an overall score and five component scores on a 1–10 scale. Free scoring keeps the photo on your phone or computer.

The measurements come from the picture. The weights, scoring curve, and reference ranges reflect choices we made. A closer match to those references does not establish how attractive someone is.

Each measurement gets a closeness value. It is 1 inside its reference range and falls along a bell-shaped curve outside it. Within a component, the closeness values are combined using the shares in the table below. The component score is then:

Component score = 1 + 9 × (combined closeness)^1.6

Overall score = 0.25 × Symmetry + 0.25 × Proportions + 0.20 × Eyes + 0.15 × Nose & lips + 0.15 × Jaw & face shape

The overall score is rounded to one decimal place. These are the component weights:

The display groups totals into 1–3.9, 4–5.9, 6–7.9, and 8–10. These bands describe the tool's results; they are not population percentiles.

Ratios and angles are measured from the photo. We chose the weights, curve, and reference ranges, adjusting them with MediaPipe example photos. The equal-thirds center of 1 and the golden ratio were not invented by us. Choosing to use them here was still our decision.

These are aesthetic references, not scientific standards. The golden ratio is not a scientific standard of beauty. A precise measurement does not make the judgment attached to it universal.

We ran our own comparison using the Face Research Lab London Set: 102 adult, front-facing photos with neutral expressions, including 49 women and 53 men, ages 18–54. The dataset has ratings from 2,513 people, each of whom rated all 102 photos on a 1–7 scale. We compared the tool's scores with the mean human rating for each photo.

For all photos, Spearman correlation was 0.32. The total score explained about 15% of the variation in mean human ratings (0.388² = 0.151).

The tool's scores occupied a narrow, high part of its scale in this dataset. Its numbers should not be read as human ratings converted to a different scale.

We ran the analysis ourselves, without third-party review. It covers one dataset of studio photos taken in London; we have not tested phone selfies in this comparison. The human ratings are subjective. The combined correlation is higher than either group result, partly reflecting differences between the groups. Correlation does not establish that the tool measures beauty.

Dataset attribution (CC BY 4.0): DeBruine, Lisa; Jones, Benedict (2017). Face Research Lab London Set. figshare. Dataset. https://doi.org/10.6084/m9.figshare.5047666.v5

The dataset's photos are not displayed on this page.

We scored the same 102 photos twice on the same machine using the same browser engine. All 102 results were identical; the largest difference in total score was 0.

That test used identical images. We have not measured how much the score changes when the same person uses a different photo. Light, angle, and lens can change the proportions visible in a picture, so the repeat test does not establish consistency across different photos.

Every face uses the same formula. Scoring reads ratios, without reading gender, age, or race. The grouped results for women and men are shown above. We have not tested results by age or racial group. Using one formula for everyone does not establish that it works equally well for every group.

The score leaves out skin, hair, makeup, height, voice, clothing, and personality. It cannot measure your value, character, or popularity. It estimates geometry in a photo and does not supply a population ranking or a promise about how people will respond to you.

Free scoring happens in your browser and does not upload the photo. For a paid report, only after purchase and clicking Generate does the photo go to our server, then through fal and OpenRouter to an Anthropic model that reads it and writes the report.

We send X-Fal-Store-IO: 0 with each request to tell fal not to store the request history. See fal's request-history documentation.

OpenRouter says it can send requests to providers with a zero-data-retention policy, and that some providers keep data, for example to scan for abuse. See OpenRouter's zero data retention guide.

Anthropic says its API deletes inputs and outputs within 30 days by default, with exceptions such as other agreements. See Anthropic's data retention terms.

We have not confirmed that the actual route through fal, OpenRouter, and Anthropic uses zero retention, so we do not promise it. The fal request setting does not answer that question for the whole route. Read our privacy policy before generating a report.

Face Rating launched on 2026-11-03. Free scoring uses MediaPipe tasks-vision 0.10.35 and face_landmarker_float16_v1, running on the CPU. Paid reports use anthropic/claude-haiku-5.5.

These sources document what we used. We chose the reference ranges ourselves; they do not come from these sources.

• DeBruine, Lisa; Jones, Benedict (2017). Face Research Lab London Set. figshare. Dataset. https://doi.org/10.6084/m9.figshare.5047666.v5
• MediaPipe Face Landmarker: MediaPipe tasks-vision 0.10.35 and face_landmarker_float16_v1.
• fal's request-history documentation.
• OpenRouter's zero data retention guide.
• Anthropic's data retention terms.