Understand a changing result

Why Your Face Symmetry Score Changes

Two photos of the same face can give different scores. Head turn, expression, distance and estimated point placement all change the input. Use this guide to check the likely source and compare repeated captures with a consistent method.

In this guide
  1. Where a changing symmetry score comes from
  2. Check these photo conditions first
  3. What the math cancels - and what it leaves in
  4. Run a controlled 3-to-5-photo check
  5. Calculate the variation across your photos
  6. Read the median and spread together
  7. Why a five-frame camera scan is a different check
  8. When the headline changes by a point
  9. Questions about changing symmetry scores
  10. Sources and further reading
Read the measurementStart with mismatch

A rounded headline hides detail. Use the percentage of eye spacing beside it.

Check the inputMatch the conditions

Keep the camera, distance, pose, light and expression consistent.

Inspect the evidenceLook at the points

Find which region changed and whether its landmarks follow the visible feature.

Inspect a photo in the checker Explore your photo variation

Where a changing symmetry score comes from

There are several steps between a face and a number. A photograph projects a three-dimensional subject into two dimensions; a model then estimates points from the pixels. Our calculation aligns those points and converts their reflected distances into an index. A difference at an earlier step can travel through the rest of the pipeline.

Four stages between a portrait and its symmetry index A vertical flow diagram. First, capture conditions determine the photo. Second, the detector estimates facial landmarks. Third, the calculation levels the eye line, reflects sixteen pairs and divides their distances by eye spacing. Fourth, the display converts the average mismatch to an index and rounds it. Capture and detector changes can affect the input; rounding can hide or exaggerate a visible step in the headline. 1 Capture the photo Pose · distance · expression Light · visibility · processing 2 Estimate landmarks The model places image points. Check their overlay positions. 3 Calculate mismatch Level eye line; reflect 16 pairs. Mean distance ÷ eye spacing. 4 Display the result 100 − mismatch % (floor: 0). Round index to a whole point.
Original diagram of this checker’s measurement pipeline. A changed headline can begin with a changed photo, a changed landmark estimate or a rounding boundary. The equation stays the same for fixed inputs.

The index uses the image’s x and y coordinates. Although MediaPipe also returns depth-related landmarks, expression scores and transformation matrices, those outputs do not make this calculation a calibrated 3D facial scan. An estimated pose check helps screen inputs; it does not reconstruct a straight-on view.

Check these photo conditions first

Common reasons two photos can produce different measurements, with a practical check for each
What changed?Why it can affect the resultWhat to check
Head turn or chin angleYaw turns the face left or right; pitch tips it up or down. Both change the projection of points at different depths.Look straight toward an eye-level lens. Compare nose and cheek visibility before comparing numbers.
Camera distance or positionMoving the camera changes perspective. Nearer parts of the face can project larger relative to farther parts.Use a stable camera position and repeat the same distance. Keep the whole face clear in the frame.
ExpressionA smile, squint, open mouth or raised brow changes visible landmark positions.Use the same relaxed expression. If the question concerns a smile, compare similarly expressed photos in the smile symmetry test.
Light or obstructionA strong shadow, hair across a feature or a reflection on glasses can make a boundary harder to locate.Use even light and keep the relevant features visible. Inspect the estimated overlay rather than assuming the detector found every point correctly.
Blur, filters or reshapingBlur and image processing change the pixels the detector receives; reshaping can alter the geometry itself.Compare sharp originals with equivalent processing. A screenshot or social media copy may differ from the original file.
Tool or model versionA new detector, selected point set, reference line or formula can change what is measured.Use the same tool and method. Keep the model and method identifiers from the measurement export with your capture notes.

Weiss and colleagues’ 2020 research evaluated asymmetry methods using simulated photographs derived from 3D face models and found that head pose could confound traditional scores. That supports checking pose; it does not provide a correction factor or a validated tolerance for this checker.

Ward and colleagues’ 2018 mathematical model shows why short camera distance changes apparent nasal proportions. Perspective and lens distortion are different effects: stepping back changes the viewpoint, while zooming or cropping from the same viewpoint changes framing. The study did not measure this site’s symmetry index or show that every distant photograph is accurate. See how selfie distance affects facial proportions for a separate perspective experiment.

What the math cancels - and what it leaves in

For each image, the checker defines two eye centers from the inner and outer eye corners. It centers the coordinates at their midpoint, levels the line between them, then reflects 16 selected landmark pairs across the perpendicular centerline.

For the aligned landmark coordinates:
dᵢ = √((−x-left − x-right)² + (y-left − y-right)²)

Overall mismatch:
M = 100 × (d₁ + … + d₁₆) ÷ (16 × D)

D is the distance between the corner-defined eye centers. M is a percentage of that eye spacing. The display index is max(0, 100 − M).

Properties of the calculation when the landmark coordinates are transformed exactly
Coordinate changeEffect on the calculationLimit in a real photo
TranslationMoving every point by the same horizontal and vertical amount cancels when coordinates are centered.A new crop or framing can still change the detector’s estimates or remove a feature.
Rotation within the image planeRotating every point equally cancels through eye-line leveling. This is the mathematical correction for roll.A real head movement can include yaw or pitch as well as roll, and image rotation can change the pixels presented to the detector.
Uniform scalingMultiplying all coordinates by the same positive factor multiplies both residual distances and eye spacing, so their ratio stays the same.Resizing an image may change available detail and estimated points. Stretching width and height by different amounts is not uniform scaling.
Yaw, pitch or camera perspectiveThese do not generally move all projected points by one common 2D translation, rotation and scale.Eye-line leveling and normalization cannot remove them or recover hidden parts of the face.

These are properties of fixed coordinates, not promises about the detector. Normalization removes a particular mathematical scale effect. It does not subtract landmark-placement error or make two unmatched photographs equivalent.

The reference also matters. Leveling the two eye centers absorbs their relative vertical difference into the alignment, so this index cannot independently report every eye-height difference. The manual eye symmetry test measures corner-defined eye height relative to the original photo axis instead. Read the full methodology or the score formula guide for the point set, region weights and rounding rules.

Run a controlled 3-to-5-photo check

A small set can help you see how much the result varies in one setup. This is a troubleshooting exercise, not a validated accuracy study or a test for physical change. Decide which conditions to keep fixed before you begin.

  1. Fix the camera setup. Use the same camera, lens setting, position and distance. Place the lens at eye level and keep enough detail to see the features clearly. Use soft, even light and equivalent image processing.
  2. Take separate photos. Capture 3 to 5 frontal photos over a short session. Relax and face the lens again between captures, while keeping the setup fixed. Reusing one file measures the same input repeatedly rather than testing how you reproduce a photograph.
  3. Inspect every overlay. Check the region views and individual pairs. If a feature is obscured or a point is visibly misplaced, record that reason and retake the photo. Do not drop a usable result simply because its score is lower.
  4. Record mismatch and conditions. Note the percentage of eye spacing, capture order, relevant warnings and method version. The main display rounds mismatch to one decimal place; the measurement JSON export retains more detail.
  5. Summarize the set. Enter the mismatch values below. Read the median alongside the minimum, maximum, range and median absolute deviation. Keep the photos and notes so you can inspect an unusual observation.

Then, if useful, deliberately change one condition - such as camera distance - and record it. That experiment asks how the photo measurement responds to that condition. It does not identify the cause of a difference in facial anatomy. For a side-by-side overlay, use Compare Face Symmetry and its capture checklist.

Explore your observations

Calculate the variation across your photos

Enter the mismatch percentages of eye spacing shown beside your results. Start with 3 to 5 separately captured, matched photos. The calculator accepts up to 12 observations and summarizes only the numbers you enter.

Use 3–12 nonnegative numbers, separated by commas, spaces or new lines. Use a decimal point and leave out the % sign. Enter mismatch values, rather than 0–100 index scores.

Compare one other photo’s mismatch with this set’s median. For example, enter 3.6 to see a difference of +0.5 percentage points for the sample set.

The prefilled values are an illustrative example. Calculate to summarize your own matched photo measurements.

The chart shows values in entered order and a median line. Its vertical scale follows this set’s minimum and maximum; compare the printed values when switching sets.

Read the median and spread together

The median is the middle entered value after sorting. With an even count, it is the average of the two middle values. It describes the center of this set without giving a single extreme observation as much influence as an arithmetic mean.

The range is maximum minus minimum. The median absolute deviation (MAD) is the median of each value’s absolute distance from the set’s median. Here it is the raw, unscaled MAD - not an estimated standard deviation. Both spread measures use percentage points of eye spacing. These definitions follow the NIST statistics handbook.

A worked example with one higher observation

Take the illustrative mismatch values 3.0%, 3.2%, 2.8%, 3.1% and 4.5%. Sorted, they are 2.8%, 3.0%, 3.1%, 3.2%, 4.5%, so the median is 3.1%. The minimum–maximum is 2.8–4.5%; the range is 4.5 − 2.8 = 1.7 percentage points.

Distances from the median are 0.1, 0.1, 0.3, 0 and 1.4 percentage points. Their sorted middle value is 0.1 percentage points, the MAD. The small MAD describes clustering near the middle; it does not erase the 4.5% observation. Inspect its photo and capture notes.

An additional photo at 3.6% differs from the median by +0.5 percentage points. The sign means greater mismatch under this method. It does not establish deterioration or identify a cause.

A zero MAD is possible when enough entries equal the median even if other values differ. A zero range means the entered values agree; rounded inputs may conceal smaller differences. No cutoff here labels a set “stable,” “accurate” or “normal.” An observed range is not a confidence interval, and a comparison inside or outside it is not a significance test.

Why a five-frame camera scan is a different check

The checker’s camera scan summarizes five accepted frames: it shows their median index and observed mismatch range. Its regional details come from the frame nearest the median mismatch. The frames share a short capture, camera setup and detector, so they can also share the same projection or placement error.

A narrow range within that scan shows agreement among those observations. Separately taking and repositioning for several photos asks a broader question about your capture process. Testing on another day or device changes more conditions again. NIST’s measurement terminology distinguishes agreement under the same conditions from agreement under changed conditions; neither alone establishes accuracy against a suitable reference.

When the headline changes by a point

The headline is rounded to a whole index point. For example, mismatch of 3.49% gives an unrounded index of 96.51, displayed as 97. Mismatch of 3.51% gives 96.49, displayed as 96. The headline changes by one while mismatch changes by only 0.02 percentage points; both mismatch values display as 3.5% when rounded to one decimal.

For a close comparison, use the exported mismatch values and the same method. Each of the 16 pairs contributes equally to the overall mean, so a small change in one pair can be diluted, while a reference change can affect several pairs together. The six regions have different pair counts; averaging their percentages equally does not reproduce the overall result.

Questions about changing symmetry scores

Why do different photos of my face give different symmetry scores?

They can differ in perspective, head pose, expression, lighting, image processing and estimated landmark positions. Start with matched photos and inspect the overlay. The index describes those estimated photo coordinates and cannot tell which difference reflects anatomy.

Should I keep the highest score or average all my scores?

Keep the observations that met your planned photo conditions, including lower scores, and note any exclusions with a visible reason. For troubleshooting, summarize their mismatch percentages with the median and spread. Picking only the highest result does not characterize the set, and an average cannot correct shared projection or detector error.

How much score variation is normal?

This checker has no validated normal variation range. It depends on the capture process, selected points, detector and method. The calculator describes your entered observations without assigning a normality threshold or deciding whether a change is significant.

Can the exact same photo get a different score?

The calculation is deterministic for fixed landmark coordinates and the same formula. Identical pixels do not by themselves establish that every detector version, configuration or runtime will return identical estimates. If a result differs, compare the actual file, processing, landmarks, model identifier and method version first.

Why is my score different in another app?

Apps can use different landmarks, reference axes, denominators, weights and final scales. A 95 on one site does not necessarily mean the same measurement as 95 here. Use the same disclosed method for a photo comparison; agreement between apps is not an accuracy study.

Does flipping a photo remove its asymmetry or distortion?

A horizontal flip preserves the picture’s distances and perspective while reversing its sides. A detector may estimate different points in the reversed image. Flipping does not repair perspective or turn a mirrored composite into a measurement of anatomy. Explore the inverted filter guide for the visual distinction.

Can a lower result prove my face became less symmetrical?

No. A difference can arise from the photograph, landmark estimates or the method, and the tool cannot separate those from physical change. Persistent concerns about movement or function belong in an appropriate professional assessment. A score does not diagnose a condition or select a treatment.

Related questions: what an inverted filter changes, what symmetry research says about attractiveness, and the broader facial symmetry guide.

Sources and further reading

The equation, pair count, exports and scan behavior describe this site’s current implementation. The sources below explain detector outputs, projection, pose and descriptive statistics. None validates this product’s index or supplies a normal variation band.

Privacy settings

Photo analysis works without analytics. You can optionally allow Google Analytics to measure basic page visits. We do not send photos, filenames, face measurements or scores to analytics.

Analytics is off unless you allow it.

Your choice is stored in this browser. Turning analytics off after allowing it reloads this page; save any wanted result first. Read the privacy policy.

Check my face