Skip to content

Measuring Beauty

Anaface analysis of Michelangelo's David

Anaface analysis of Michelangelo’s David

I’ve come across a paper that was using “beauty” as one of the predictors. To measure beauty, the authors used

I don’t trust metrics without trying them on a gold standard first. So, I tried how well Anaface does on something that the arts world considers as one of gold standards of beauty – Michelangelo’s David. My annotation might be imperfect, but David only gets to be only a good 7: his nose is too narrow and his eyes are too close.

Of course, I applaud the use of interesting predictors in studies, and Anaface is a better tool than anything I’ve seen before, but maybe we need better metrics! What do you think?


  1. Ashok Rao says:

    Ryan Gosling is only a 7.79. Obama is a 6.67.

    I’d imagine this is as good as it gets, but I would agree – better metrics (or some more rigorous way of… well… measuring your metric) are necessary!

  2. jonathan says:

    Sears used to be run – and was run into the ground – by senior management known as “the tall men” because you had to be well over 6′ feet tall to reach that level there. That speaks to a peculiarity of that corporate culture, but the idea is similar.

    I got an 8.46. My eyes are too small for the distance between them, my face is too narrow and my ears are too long. I am not a CEO, never wanted to be a CEO and actively avoided that.

  3. Rahul says:

    How sensitive is this to the person coding the dots? Try having the same photo re-coded by three of your friends. I suspect variation levels that may swamp the signal.

    Even worse, try having it rate different photos of the same person. This may be influenced more by quirks of photographic geometry (e.g. how perfectly straight was the person looking) than by the underlying data.

  4. Quartz says:

    It’s not that easy, and it might even be an ill-posed problem. In an evolutionary perspective certainly extreme deviations from a “healthy” region are signals of bad health and thus not attractive, while on the other hand inside that region there must be variability to allow for a mechanism also to avoid excessive inbreeding. If beauty was a target, a point (or even an area) then we would genetically get trapped in a fixed point, which we know luckily didnt happen and on the contrary our biology tries to avoid.

    Also remember how culturally contingent is the concept of beauty: nowadays tan is attractive just like once it was pale skin. Classical beauties do have curves, while today’s media spread an anorexic model. Distant ethnic groups are not equally attractive.
    And think how boring some charming actresses look when no professional photographer is at work etc…

    Imho human “beauty” is not so easy to measure, it’s part a process, it’s relative, it’s cultural, and it’s in the eye of the beholder (trite sayings need not fail to withhold).
    Sure you can measure some geometrical relationships of a face, and often you hear of golden ratios and similar relations hinting at harmony or other “important” characteristics, but from there to judging the overall composition is a larger leap than it seems.
    This topic is also related to zen gardens, which achieve a certain kind of visual harmony by avoiding structures, which is extremely hard (as anyone of you who designed a pseudorandom generator knows) the more elements you get (and our faces have many more than it might seem at first).

    Fascinating topic, and someday we’ll have some interesting adaptive measure, but nowadays its still a geek kind of quackery imho, just like the “formula of love” and similar catchy pursuits.

    • Quartz says:

      On symmetry, again it’s not that easy: large asymmetries look ugly (since they hint at health issues) but perfect symmetry has also been found to be less attractive than our real slight asymmetries. But that doesn’t mean that ANY small asymmetry is attractive etc…

    • Rahul says:

      Rather than one target metric of beauty it might be better to think of it as one cluster of features that another cluster of people likes. Ergo there may be multiple such feature clusters.

      Also, people take a more aggregate approach to beauty; the body type, hair, posture, skin color, skin texture and such metrics are ignored by a purely facial symmetry analysis.

  5. Michael says:

    We ran a simulation to see if students would select more attractive people for managerial positions in a (simulated) company. To determine physical attractiveness, we used Mechanical Turk, which worked well for our purposes. It would be interesting to see how well results from Anaface matched with results from a Mechanical Turk test.

    The Mechanical Turk test indicated that age and weight were strong predictors of attractiveness for both males and females. Would Anaface work with those characteristics?

  6. Lord says:

    The horizontal asymmetry is obviously the slight tilt in photo for which there is no way to correct, how open or closed the mouth may have an effect, and there may be a discrepancy between natural and model. It is unclear what 7 out of 10 or 78% means or who would fit their model of perfection or how aging may or may not affect this as I expect it probably affects our impressions more. Is this a ranking, some normal distribution measurement, or multifactorial product with some power law? Ranking doesn’t seem right as it seems too stringent. Some photo groups of different numbers would help. Anyone come up over 9?

  7. Lee Sechrest says:

    A long time ago, someone defined a unit of beauty called a “Helen,” which was the amount of beauty required to launch a thousand ships. Probably not even super models come close to I Helen.

  8. Steve Sailer says:

    Great example to test the measure upon: The David really is the most awe-inspiring work of art (painting or sculpture) in Europe.

  9. “David” was carved to be viewed from below and has many features adjusted by the genius sculptor to allow/compensate for this… Consequently your reference photo with a horizontally level standpoint is a flawed benchmark…

Leave a Reply