MediaPipe Face Landmarker documentation
Google AI Edge
How we use it. Documents the model and its 478 estimated landmarks. Our scores, shape profiles and style suggestions are our own calculations, not MediaPipe outputs.
What it does not establish. A landmark detector is not an attractiveness model and does not validate our scoring weights.
Nasal Distortion in Short-Distance Photographs: The Selfie Effect
Ward et al., JAMA Facial Plastic Surgery (2018)
How we use it. Background for why camera distance changes relative feature size in a perspective image.
What it does not establish. This paper does not test AI Rate Face. Our camera simulation uses its own stated dimensions and is not a reproduction of the study.
New “golden” ratios for facial beauty
Pallett, Link & Lee, Vision Research (2010)
How we use it. An example of a study of particular feature arrangements and preference judgments.
What it does not establish. Its measured arrangements are not a general endorsement of 1.618, our four selected ratios, or a universal beauty score.
The role of the golden proportion in the evaluation of facial esthetics
Rossetti et al., The Angle Orthodontist (2013)
How we use it. Context for the limits of treating selected facial ratios as an attractiveness standard.
What it does not establish. Results from one study and one measurement protocol should not be generalized to every face, culture or photo.
Which numbers are our own design choices?
The overall 30/25/25/20 score weights, the seven face-shape profiles, ratio tolerances, confidence labels and styling suggestions are not fitted to a published population dataset. They express how this version of the product summarizes a photograph.
The selected golden-ratio references include φ ≈ 1.618 for three ratios and 1.00 for eye gap / eye width. Facial thirds and fifths are proportional drawing conventions. Our photo-based fWHR, midface and profile ranges provide interface context; they do not establish clinical normality, personality or an attractiveness threshold.
Where a calculation uses a chosen reference rather than a measured population estimate, read it as a comparison with this tool's target. We do not attribute our exact cutoffs to a paper that did not test them.
Original calculations and data
The Camera Lab publishes five computational studies. Their inputs, code and outputs are available to download. That is evidence of how the calculation behaves on those inputs, not independent validation of the product.
Read the methodology for formulas and the editorial policy for source and correction standards.