Side Quest: The Average Model Face

Article / 20 April 2020

During an earlier flash sale, I grabbed the LoveChild software at a steep discount. It's a photo morphing software that is meant for facial portraits. Its advertised use is for you to combine two portraits to create a new "person" that you can use to generate 3D faces with the developer's other software (which I do not own).

As part of my workflow for creating characters, I definitely prefer to pick and choose features that I want. And "averaging" people tends to make them less distinctive. But I love playing around with software, experimenting, and having fun.

Have you seen composite photographs that show the "average" human face, or "average" human female face from [insert continent here] or "average" human male face from [insert country here]? Like this? They usually end up being fairly attractive no matter what. 

So what happens when you average a gaggle of models who are attractive to begin with? I chose to experiment with photographs from Miu Miu's Fall/Winter 2020 runway show, since I'd have fairly consistent lighting and a fairly large sample of models (66 with no repeats--well, I didn't fully verify this, but neither Bella nor Gigi Hadid walked twice, and those two tend to be the most likely suspects). Of course, since LoveChild only processes two photographs at a time, as opposed to averaging out the entire set all at once, a few photos would have to be weighted higher/lower. That turned out not to matter much, though, because by the time I whittled this down to a few photographs, they were all looking pretty similar anyway.

The result? See above. What do you think? Is it close to what you would have expected?

Some notes on using LoveChild: 

  • As you can see in the product promos, there are definitely artifacts when you combine two photos. But as you can see from my photo shared above, as you keep averaging photos, the result is blurrier but smoother. (The result may be better if higher resolution photos were used, but I cropped the heads from full-body photos, so the originals weren't of the highest quality.) 
  • The program looks for faces. If another model appeared walking behind the main model being pictured, I did have to crop out the background faces or else the program would produce an error.
  • Some of the models wore sunglasses. Most did not, so the glasses mostly faded away by the end result (though I admit I did a tiny, tiny bit of touch-up deleting a faint line).

Some thoughts on further experimentation:

  • This tool could be used as a visual representation of how any brand has cast more diverse models over the years. Or, how they compare to each other in any given year.
  • We could also see any given brand's "taste" in runway models, how that has evolved, or how it has changed over the years. Probably only minimally useful, since any individual model could vary greatly from the average, but if a show has a distinct preference for long noses, maybe that would be reflected in the average.

If you're on my website and want to comment, please visit this post on the Artstation site: