Here is a novel approach to generating more attractive faces using a computer, though the morphing methodology used leaves much to be desired. The article is by Wong et al.(1, pdf)
The genetic algorithm used
Take a sample of faces and have them rated for attractiveness by judges on a 1 – 10 scale. Then each of these faces is randomly assigned a number between the minimum attractiveness score and maximum attractiveness score. If a face has an attractiveness rating that is greater than the random number associated with it, it “breeds,” i.e., it is morphed with another randomly selected face that satisfies the same requirement to get a progeny (child). Using this approach, because of chance, it is possible that some very unattractive faces get to “breed” and very attractive faces do not get to “breed” but the bias is that more attractive faces are more successful at “breeding” or carrying over to the next generation.
The shortcoming of the study was using only 30 faces per generation. Small sample sizes can result in chance playing a greater role in determining who gets to “breed.”
Morphing
The morphing methodology or way to blend two faces did not adjust for face size or more precisely centroid size. The authors adjusted for face height and indicated that next time they would adjust for distance between the eyes, but as explained previously, neither method adjusts for face size (see how to adjust for face size).
The Parental sample
Each of the following faces is a blend of two faces having the same ethnic background. The reason that original faces were not used for the study is that the authors had many faces and could not obtain permission for using the photographs from each participant who was photographed.
The Parent faces.
The faces are derived from whites, Latinas, East Asians and Middle Eastern individuals. The progeny generations were obtained by morphing without any regard for ethnic background, and the judges were ethically mixed, too. Hence the study is bound to provide some interesting results.
The most and least attractive faces in the parental and progeny generations
Most attractive (top row) and least attractive faces in parental generation (P), first generation of progeny (F1), second generation of progeny (F2), third generation of progeny (F3) and fourth generation of progeny (F4), respectively. The attractiveness ratings are listed at the bottom.
The most relevant finds are the elimination of 1) the more masculine faces, 2) the more-distant-from-white Latinas/their Middle Eastern counterparts and 3) stereotypical East Asians in the progeny (see the pdf for more pictures). In the first generation offspring, the most attractive face is predominantly European but with slight East Asian mixture. In the second generation, the most attractive face is European. In the third and fourth generations, the most attractive face for each generation is ethnically unclassifiable (e.g, European nose, East Asian width).
The third and fourth generation results are a little odd since it appears that the European element in the most attractive face has gone down though the most attractive face in the second generation would rank among the more attractive faces in the fourth generation. This could be an artifact of chance, as explained above, given the small sample size, but then it could also be argued that the results of the first two generations are artifacts of chance. However, take a look at all faces in the third and fourth generation:
Faces in the third (D) and fourth (E) generations.
We note that the faces in the third and fourth generations generally lean toward Euro-Mediterranean norms. European faces are more overall derived and there is a preference for more derived faces. Therefore, the results of this study are broadly consistent with a preference for more feminine faces in women as well as a preference for more derived facial features.
Comments about the study
The morphing methodology is poor. Cosmetologists is spelled as cosmotologists. There is a mistake in the statement, “The multiple correlation coefficient for this model was 0.12, indicating that 12% of the variability in average facial attractiveness score was explained by the regression on these predictors.” It should be 1.4%, not 12%.
References
- Wong BJ, Karimi K, Devcic Z, McLaren CE, Chen WP. Evolving attractive faces using morphing technology and a genetic algorithm: a new approach to determining ideal facial aesthetics. Laryngoscope. Jun 2008;118(6):962-974.