Face Attractiveness Analyzers: What the Underlying Research Actually Supports

A face attractiveness analyzer is supposed to tell you something a mirror cannot: what specifically is driving how a face reads, backed by more than one person's opinion. Most don't. They return a score and imply the model behind it understands attractiveness, without saying what the model was trained on or which research it reflects. The controlled academic literature on facial attractiveness is decades deep and reasonably consistent on what actually predicts it. A legitimate analyzer should be built on that research and able to point to it. Here is what that research says, and what to check for before trusting an analyzer's output.

Definition

What the research actually supports

1. Averageness

Langlois and Roggman's 1990 study in Psychological Science digitally averaged photographs of multiple faces into single composite images and had them rated for attractiveness alongside the individual faces that made them up. The composites, mathematically closer to the population mean on nearly every dimension, were rated as more attractive than the individual faces averaged into them, and the effect got stronger as more faces were blended in. Averageness is not a synonym for "unremarkable": it is a specific, measurable proximity to population-level facial proportions, and it is one of the more consistently replicated findings in this literature.

2. Symmetry

Little, Jones, and DeBruine's 2011 review in Philosophical Transactions of the Royal Society B, a standard reference in the evolutionary psychology literature on this topic, summarizes decades of controlled work and identifies facial symmetry as a second consistent predictor of attractiveness judgments, alongside averageness and sexually dimorphic features. Symmetry is measurable directly: it is the degree to which landmark positions on one side of the face mirror the other, not a subjective impression of balance.

3. Sexually dimorphic features

Perrett and colleagues' 1998 study in Nature manipulated the sexual dimorphism of face images, exaggerating or reducing masculine and feminine shape cues, and found attractiveness ratings tracked those manipulations in both British and Japanese samples, with a preference in that data for feminized faces beyond the population average in female raters. Dimorphism is a set of specific shape differences (jaw width, brow ridge, cheekbone position, among others) that can be measured on a face and are not evenly weighted; the Little, Jones, and DeBruine review notes that some dimorphic cues are frequently outweighed by symmetry in the same judgments.

Why a single score cannot represent this

These three factors, averageness, symmetry, and dimorphism, are independent measurements with independent research behind each one, and they interact rather than summing neatly. An analyzer that returns one number has already decided, invisibly, how to weight all three (and whatever else it may be tracking) before you ever see the output. There is no way to check that weighting because there is nothing to check it against. A breakdown that reports each factor on its own, by contrast, can be compared directly to the studies above: if a tool claims to measure averageness, it should be measuring proximity to population norms the way Langlois and Roggman did, not producing a number with the same name attached to a different process.

What to check before trusting an analyzer

• Ask which factors it measures by name. Averageness, symmetry, and dimorphism are the three with the most consistent research support. If an analyzer cannot name what it is measuring beyond "attractiveness," it is not measuring these.
• Ask for a citation per factor, not one citation for the whole tool. A legitimate analyzer can point to the specific study behind each measurement, the way the three above can each be looked up directly.
• Check whether the output is a single number or a breakdown. A single score has already merged averageness, symmetry, and dimorphism together with a weighting you cannot see or verify.
• Upload the same photo twice. A deterministic, landmark-based analyzer returns the same measurements each time. A model trained on unlabeled preference data frequently does not.

What this looks like in practice

Facet is built on the measurement side of this research. A scan runs 468 facial landmark points through a deterministic scoring engine, then checks each resulting measurement, including proximity to population averages, bilateral symmetry, and specific dimorphic proportions, against a threshold drawn from a specific peer-reviewed source, the same standard cited above. The output is a breakdown across 10 modules with a citation attached to every threshold, not a single attractiveness number standing in for all of them. That structure, and how it differs from the subscription-based clinical alternative on the market, is covered at [Facet's homepage](/) and in the direct comparison at [Facet vs QOVES](/alternatives/qoves).

The fastest way to see what an attractiveness analyzer's single number is actually standing in for is to look at the measurements underneath it. A complete example scan, every module and every cited threshold filled in, is posted at [/sample-report](/sample-report). Run your own photo through it to see averageness, symmetry, and dimorphism reported and sourced separately.