How our tools work

Every tool on this site follows the same four steps: find your face, check the photo is usable, measure, then explain the result. This page covers each step, what every tool measures, and the limits you should know about.

Your photo stays on your device

The face analysis runs inside your browser. When you start a tool, your browser downloads a face-mapping model (Google’s open-source MediaPipe Face Landmarker, licensed under Apache 2.0) from our own site. From then on, everything happens on your phone or computer:

• your photo or camera frame is analyzed locally and is never uploaded;
• measurements and results are calculated locally and are never sent to us, and never included in analytics or ads;
• when you leave or start over, the image is discarded.

Our site’s security settings only allow the page to talk to our own server, which adds a technical safeguard on top of the promise. The model files are only downloaded when you press start, so they don’t slow down the rest of the site.

Step 1: finding your face

The model places 478 reference points (landmarks) on your face: around the eyes and irises, eyebrows, nose, lips, jaw and the outline of your face. It also estimates how your head is turned.

Step 2: checking the photo

Before measuring, each tool checks that:

• exactly one face is in the picture;
• your face is large enough in the frame;
• your head isn’t turned, tilted or tipped too far (we tell you which way to adjust);
• the lighting isn’t too dark, too bright or too uneven.

If something’s off, you get a specific fix, like “Turn your head slightly left”, instead of a poor result.

Step 3: measuring

Measurements are ratios, not pixels, so the size of your photo doesn’t matter. We divide distances by a stable reference on your face (the distance between the inner corners of your eyes). Where software can’t see something well, we don’t guess:

• Hairline. The face map stops above the brows, so face length starts from an estimate that you can drag to your real hairline.
• Eyelid creases. A photo can’t reliably show whether a crease is hidden, so the eye tool asks you three quick questions.
• Nose profile. A front photo can’t show the bridge from the side, so you choose your profile from simple drawings.
• Hair color. You confirm or pick your natural hair color from swatches.

Step 4: explaining the result

Classification uses clear, published rules, not a black box. Each tool shows:

• the most likely result and how confident it is;
• the second most likely result (many faces sit between two types);
• a plain-language explanation of why;
• the key measurements, drawn over your photo so you can see what was measured.

What each tool measures

Every tool also works without a photo, using tape-measure values or a short quiz.

Face shape: the method in detail

The face shape detector measures six things, all in the face’s own upright frame so a tilted head doesn’t change them:

The hairline starts as an estimate: we place it above the brows at about 0.85 times the distance from the brows to the base of the nose, which matches measured adult averages. You can drag it to your real hairline.

Comparison with an average face. Each measurement is divided by the value for Google’s average (canonical) face model measured the same way, so 1.00 means average. Tape-measure mode uses typical tape-measure proportions instead, because a tape measures slightly different points.

Scoring. Each of the seven shapes has a few criteria (for example, heart: forehead at least as broad as average, jaw narrower than average, face narrowing from forehead to jaw, narrow chin). Each criterion scores from 0 to 1 and is weighted. The shape with the highest total wins. Neighboring shapes deliberately overlap, so a face halfway between oval and oblong gets a close runner-up and a low confidence rather than a falsely certain answer.

Confidence is high when the winning score is strong and well ahead of the runner-up. It’s lowered when the photo had quality warnings, capped at medium while the hairline is still an estimate, and capped at medium for the quiz.

We tested the rules on synthetic faces made by reshaping the average face into each shape, and checked them on test photos. Because face shape is a styling convention rather than a medical measurement, there’s no official dataset to test against, so we’ll keep refining the thresholds.

Eye shape: the method

The eye shape detector measures, for each eye: height divided by width (round vs almond), the angle of the line from the inner to the outer corner (upturned, straight or downturned), and the gap between the inner corners divided by eye width (close-set or wide-set). Each is compared with the average face. If the iris landmarks show white below the iris, that also points to round.

A photo can’t reliably see a hidden eyelid crease or judge depth, so three questions cover the crease, the visible lid and how deep the eyes sit. No visible crease means monolid; a covered crease or no visible lid means hooded; eyes set back under the brow bone mean deep-set. The corner angle is only used to describe shape. It is never scored or ranked.

Nose shape: the method

The nose shape detector measures three things from the front: the width across the nose wings divided by the distance between the inner eye corners, the length from the bridge to the base divided by the lower face, and the tip width divided by the nose width. Each is described as narrow, medium or wide against the average face. The side profile (bridge line, tip direction, tip size) is chosen from drawings, because a front photo can’t show it. Each nose type has a points table for those three answers, and the best match wins. Names like Roman and Greek come from classical art; we never link nose types to ethnicity or nationality.

Lip shape: the method

The lip shape detector measures lower lip height divided by upper lip height, total lip height divided by mouth width (fullness), mouth width divided by the distance between the inner eye corners, and the depth of the cupid’s bow dip divided by the upper lip height. Shapes such as full, thin, heart, bow, wide, round, top-heavy and bottom-heavy each score against those values, and we show one or two. If your lips are parted or you’re smiling, the result is marked low confidence, because both stretch the lips.

Color analysis and undertone: the method

For the color analysis and undertone test, we sample color from both cheeks, the colored ring of each iris and a patch above the forehead for hair. We drop the brightest and darkest pixels in each area, which removes shine, shadow and stray hairs, and take the median of the rest. Colors are converted to CIELAB (D65), the international color space built around human perception.

• Temperature: mostly the skin’s hue angle (pinker reads cooler, more golden reads warmer), plus hair and eye warmth.
• Depth: mostly your confirmed natural hair color, plus skin and eye lightness.
• Clarity: the contrast between hair and skin, plus how vivid your eyes are.

Each of the 12 seasons is a point on those three scales, and the nearest one wins. Because lighting and camera white balance can shift color more than the differences we’re measuring, the photo is always blended with the quiz. It counts for 40% normally, 50% if you correct the light with white paper, and 25% if the photo looks strongly tinted. For undertone, the photo counts for 30% and the quiz for the rest, since olive undertones in particular are hard to see in photos.

The numbers behind each result

Most face shape and feature tools don’t say how they decide. These are the actual rules our tools use, copied from the files that run them. A result scores fully at the values below; values partway there score partly, which is how a close runner-up appears.

Face shape, in tape-measure terms

All values compare your measurements with your cheekbone width, except where the forehead is named. The photo mode uses the same rules, measured against an average face from the face mesh.

Confidence is high when the winning shape scores at least 0.75 out of 1 and leads the runner-up by 0.20, and medium at 0.55 with a lead of 0.08. Anything less is shown as low confidence.

Eye shape

Each value is compared with the same measurement on the average face (1.00 = average).

Hooded, monolid and deep-set come from your answers to the three lid questions, not from numbers. The corner angle only describes shape. It is never scored or ranked.

Nose shape

The front photo gives three descriptions, each compared with the average face: width (nose wings compared with the gap between the inner eye corners; average 0.96) reads narrow below 0.90 times average and wide above 1.10; length reads short below 0.92 and long above 1.08; the tip reads refined below 0.92 and full above 1.06 times average.

The type comes from your three profile answers. Each answer gives points to the types it fits: a straight bridge gives Greek 3 points, a gentle bump gives Roman 3, a strong curve gives aquiline 3, a scooped bridge gives snub 2.5; a full tip gives fleshy 3, a small tip gives button 2.5. The type with the highest share of its possible points wins.

Lip shape

Heart and round lips combine several of these: heart needs a defined bow with a fairly full upper lip, round needs fullness on a narrower mouth with a softer bow.

Undertone and color season

Skin color is read as a hue angle in CIELAB. Below 52 degrees it votes cool (pinker), above 60 degrees warm (more golden), and between the two neutral. Olive needs a yellow-green lean with little redness (redness less than half the yellowness) and moderate color strength. The photo is 30% of the undertone result.

For the color season, each of the 12 seasons sits at a point on three scales from -1 to +1: temperature (cool to warm), depth (deep to light) and clarity (soft to bright). For example, light spring sits at warm 0.5, light 0.9, clarity 0.2, and deep winter at cool 0.5, deep 0.9, clarity 0.2. Your answers and photo place you on the same scales, temperature counts 1.25 times as much as the other two, and the nearest season wins.

Limits we want you to know about

• Phone cameras distort. Close-up selfies make the middle of the face look bigger. Hold the phone at arm’s length or further, at eye level.
• Categories overlap. Face shapes, eye shapes and seasons are styling conventions with fuzzy edges. That’s why we show a second result.
• Light changes color. Color results are only as good as the light. Daylight from a window, without direct sun, works best.
• We never score attractiveness. Tilt, proportions and colors are used only to describe your features for styling, never to rate them.

Read the full disclaimer for more.