What is combined across frames
For each usable frame, the scanner calculates face length divided by cheek width, forehead divided by cheek width, jaw divided by cheek width, the average jaw corner angle, and the chin taper angle. The photo and video paths use the same pixel-space geometry function. Its distances and angles are invariant to image-plane rotation; turning the head can still change them.
The result classifies the median of each of those five features. It does not average raw coordinates or vote on labels. The optional JSON download also includes a mean that discards the lowest and highest 10% of values for each feature.
Which frames are used
The beta checks that there is one face with a visible forehead and chin, sufficient frame coverage, near-frontal pose, a neutral expression and adequate light and edge detail. Pose uses an eye-line roll, a nose-offset proxy and the model's face transformation. Expression checks use the model's blink, jaw-open and smile coefficients. These checks are heuristics; they are not an independently validated occlusion detector.
The scanner targets 50 accepted frames. After 10 seconds of capture it can finish with at least 30, provided they span at least 1.8 seconds. Duplicate video timestamps are not sampled, only one frame is processed at a time, and long gaps or multiple faces reset the accepted window. Model preparation is timed separately. These are beta settings, not promises about every device's speed.
What “scan variation” means
For each feature we calculate the interquartile range: the 75th percentile minus the 25th percentile. A scan has low variation when every feature is within the scale below, moderate variation when every feature is within twice its scale, and high variation otherwise. These scales are authored operating thresholds, not measured population statistics.
| Feature | Low-variation IQR scale |
|---|---|
| Length / cheek width | 0.025 ratio units |
| Forehead / cheek width | 0.025 ratio units |
| Jaw / cheek width | 0.025 ratio units |
| Jaw corner angle | 3° |
| Chin taper angle | 3° |
A low-variation scan can still be consistently wrong because of hair, pose, camera perspective or the classifier's reference profiles. Match weights describe similarity to seven profiles; usable-frame counts describe acceptance; variation describes repeatability within a scan. None is a calibrated accuracy percentage. Video frames are correlated and must not be counted as independent participants.
Collect a useful repeatability comparison
- Use your own face or an adult who has explicitly agreed. Keep the same device, camera distance, expression and lighting for the first set.
- In the camera tool, enable measurement download before scanning. Download each result and start a new scan, rather than duplicating a file.
- Repeat at least three times where possible. Record device/browser, camera orientation, lighting and any rejected or failed attempts separately.
- Compare the files below. The baseline is always the midpoint accepted frame of each scan; it is not chosen for being unusually bad.
- Test changed lighting or pose as a separate condition. Do not pool different people or conditions into one repeatability claim.