A LITTLE TRANSPARENCY
What goes into your harmony score.
The short version
FaceRater detects 478 facial landmarks in your browser with Google MediaPipe. It measures 17 proportions in the uploaded photo and compares them with a balanced reference set of adult faces. MediaPipe does not provide a beauty or attractiveness score, and neither do we.
The harmony score
Each proportion is compared with 2,560 front-facing, relaxed-expression adult photos from FairFace, a public research set (CC BY 4.0) balanced across seven perceived race groups and two genders. For 12 proportions we measure how close your value sits to the most similar group's typical value, so a proportion that is common in any group is not treated as unusual. Three of them count only in one direction: a slimmer jaw or chin and a more lifted outer eye corner than typical are not marked down, because raters consistently find them attractive; in a check against 2,513 people's ratings, this made the score agree noticeably better with them. Eye openness scores higher the more open the eyes are, because larger, open eyes are consistently rated as more attractive. For four symmetry measures, smaller differences rank higher. The proportions form five parts: facial thirds (25%), facial fifths (25%), eye area (20%), jaw and chin (20%) and symmetry (10%). Each part and the combined result are ranked against the same reference photos. The 10-point score places that rank on a curve where a typical photo reads 7.5, the top tenth 8.4 or more and the top hundredth 8.9 or more.
What the percentile means
“Ahead of 68%” means this photo's combined result is higher than that of 68% of the reference photos. The reference is a set of public research photos, not FaceRater users and not a representative sample of everyone. Once enough people have scanned, we will recalibrate against anonymous counts of FaceRater scores and say so here.
How well it tracks attractiveness
We checked the score against the London Face Set: 102 adults photographed in a studio, each rated for attractiveness by 2,513 people. The harmony score correlates with their average rating at Spearman 0.29 (95% interval 0.09–0.47); our earlier symmetry-only rating correlated at 0.05. That is a weak link: proportions explain only a small part of how attractive people find a face, which is why we call it a harmony score. Smiling lowered the score by 0.8 points on average in the same set, so use a relaxed expression.
Why not the golden ratio
There is no good evidence that a single ratio such as 1.618 defines a beautiful face, and classic “ideal” proportions come from one narrow tradition. Harmony uses the common range of real adult faces instead. Being typical is not the same as being attractive: striking faces often differ from the average in distinctive ways.
Known limits
In the reference set, average scores differ between perceived groups by up to 0.5 points; East Asian and Black faces average lower, mostly because the low-resolution research photos give noisier landmarks for some groups. Adults over 60 average lower because proportions change with age. Head turn and nod are corrected using the landmark depth, but larger angles still change the result. We will repeat these checks with FaceRater data.
Why photos give different results
Camera distance, angle, expression, lighting and occlusion change the geometry. Every photo with one measurable face is scored. A score is marked approximate when the head is turned more than 5° or tipped more than 10°, the expression is not relaxed, or the face is small, cut close, very dark or bright, or soft; beyond a 12° turn, a 20° tilt, closed eyes, an open mouth or a wide smile it is labeled a rough guide. We only stop when there is no single face to measure or the image is under 320 × 320 pixels.
What the other measurements mean
Ten further measurements describe the photo and do not change the harmony score. Symmetry describes similarity to mirrored landmark positions. Eye widths and corner angles describe what is visible in the photo. Length and width ratios refer to mesh points, not anatomical or hairline measurements. We do not infer age, ethnicity, personality or health.
Face shape and jaw readings
The face shape, hairstyle and jawline tools reuse your local landmark scan. After levelling the eye line, every width is divided by cheek width (mesh points 234–454). Length runs from the mesh forehead point to the chin: under 1.19× reads as round and 1.28× or more as oblong. Lower-face breadth combines jaw width (172–397) with the 2D contour angle at each jaw point; a broad lower face of balanced length reads as square. The upper outline (forehead sides and temples) is compared with the lower outline to identify heart and triangle shapes, and temples at or under 0.91× cheek width suggest diamond. When a small change in one measurement would produce another shape, the result names it as also close. These thresholds were set by checking synthetic test portraits and our sample photos; they have not been validated against a representative population. Hairstyle suggestions match each style’s visual effects with what usually balances the outline. None of these readings is a beauty score. The harmony score uses some of the same jaw measurements, but compares them with the reference set instead of sorting them into shapes.
Color analysis
The color analysis reads skin from five cheek and forehead patches, the irises, and a band of hair above the forehead, ignoring the darkest and brightest pixels in each area. It corrects white balance by comparing the visible whites of your eyes with a typical eye-white color, and compares skin brightness with them to estimate depth. Temperature combines skin hue with your quiz answers, which carry most of that weight; depth combines skin, hair and eyes; clarity combines hair-to-skin contrast, eye saturation and skin saturation. The season follows the trait that stands out most, then the warm–cool lean, using the 12-season system. These weights were set by checking synthetic test portraits; they are not validated against professional draping, so the result shows a confidence level and a second season when the call is close.
The full report
The free score uses a fixed formula. The AI stylist and the full style report build on your saved measurements with AI-assisted explanations and a practical plan; the report adds 44 descriptive measurements across seven regions and generated styling concepts. None is presented as an ideal proportion. Our public sample uses a synthetic adult photo and is labeled as an example.
Your original score stays put
A report and a share card use the saved analysis snapshot. The report generator cannot replace its score. Scans made before October 2026 used a symmetry-only rating (photo-balance-beta-1); on the device where they were made, they are re-scored from their saved landmarks. Current method version: harmony-1.