30 landmark tests: how stable is a face score?

What this experiment shows

The combined score spans 81–87 in these 30 configurations. Compare rows with the same jaw shift to isolate rotation; compare rows with the same rotation to isolate the moved point. The combined score also includes photo presentation.

Inputs and method

We use one hand-authored 478-position coordinate fixture in a 900 × 900 image. We combine five in-plane rotations (−12°, −6°, 0°, 6°, 12°) with six left-jaw shifts (0 to 18 pixels). Brightness, contrast and sharpness inputs stay fixed. All 30 configurations run through the production geometry and overall-score functions. These are coordinate tests, not 30 photographs.

Calculation version: ARF calculation experiments 1.0. Engine: ARF Local Geometry v3.0.

Calculated results

Calculated from the published inputs. The full values and configuration details are in the table below.

30 calculated configurations. Scores use the site's 0–100 scale; match shares and enlargement are percentages.
Rotation (°)	Jaw shift (px)	Symmetry	Geometry	Photo	Overall
-12	0	100	83	75	82
-12	3.6	100	83	75	82
-12	7.2	100	83	75	82
-12	10.8	99	82	75	81
-12	14.4	99	82	75	81
-12	18	99	81	75	81
-6	0	100	85	80	85
-6	3.6	100	84	80	84
-6	7.2	100	84	80	84
-6	10.8	99	83	80	83
-6	14.4	99	83	80	83
-6	18	99	83	80	83
0	0	100	86	85	87
0	3.6	100	86	85	87
0	7.2	100	86	85	86
0	10.8	99	85	85	86
0	14.4	99	85	85	85
0	18	99	84	85	85
6	0	100	85	80	85
6	3.6	100	84	80	84
6	7.2	100	84	80	84
6	10.8	99	83	80	83
6	14.4	99	83	80	83
6	18	99	83	80	83
12	0	100	83	75	82
12	3.6	100	83	75	82
12	7.2	100	83	75	82
12	10.8	99	82	75	81
12	14.4	99	82	75	81
12	18	99	81	75	81

Download this table (CSV) · Reproduce the calculation

How to use this when reading your result

Repeatability has two parts: identical inputs should give identical calculations, while changed inputs can give changed results. Alignment can correct an image-plane rotation, but it cannot recover hidden geometry from a head turn or determine whether a landmark shift came from expression, camera perspective or detection error.

Test your photo consistency →

What this experiment cannot tell you

The fixture has hand-set coordinates only for the landmarks needed by the calculations; unused coordinates are at the centre. It is not an anatomically validated face. No image decoding or MediaPipe inference is tested, so this experiment does not establish photo-to-photo reliability or demographic accuracy.

Sources and calculation choices

MediaPipe Face Landmarker documentation — Google AI Edge. Documents the model and its 478 estimated landmarks. Our scores, shape profiles and style suggestions are our own calculations, not MediaPipe outputs. A landmark detector is not an attractiveness model and does not validate our scoring weights.

See the full source notes and methodology. To report an error, contact us with the page URL and the input or calculation that appears wrong.