Face Averaging and Art History

Earlier this month, I blogged about face averaging as a historical technique: combining face averaging software with a time base to illustrate patterns of change over time.  Since then, I’ve tried some further experiments applying the principle not to particular groups of people as I did last time—Indiana University students, United States Senators, Miss America contestants, and so forth—but to particular genres and periods of art.  As before, I’m using the deluxe version of Abrosoft’s FaceMixer to achieve these effects.

Here are the average faces of Fayum mummy portraits—a genre of paintings created in Roman Egypt during roughly the first three centuries AD—based on 181 total examples (ninety-seven female portraits and eighty-four male portraits), most of which I scanned from a few books borrowed from Indiana University’s Fine Arts Library:fayum-faces-averaged

The female portrait on the left easily passes muster as a “Fayum face.”  I think anyone familiar with the Fayum mummy portrait genre would recognize this picture not only as an example of it, but as a quintessential example of it, arguably more typical in its way than any real Fayum mummy portrait could be.  The male portrait on the right is a little more problematic.  It too presents a recognizable “Fayum face,” but the five-o’clock shadow represents the average of paintings of men with and without facial hair.  A few Fayum mummy portraits depict men with stubble, but typically a subject would either have a beard or not rather than something midway between.  The facial hair issue makes the averaging of male faces a bit more conceptually troublesome than the averaging of female faces.  With that in mind, I’m going to focus here for the moment on female faces.

The averages shown above cover the whole time span of Fayum mummy portraiture, so they give us a glimpse at the genre as a whole, with no indication of how it may have changed over time.  However, we can also create averages for more limited periods, such as these (the number in white indicates how many different source images were averaged):

fayum-female-50year-groupsDoing this isn’t as straightforward as it might seem.  The dating of Fayum mummy portraits can be a bit controversial, and even where there’s a reasonable consensus, dates are usually given as ranges (such as “AD 138-192”) rather than as a single year.  What I’ve done here is to include all examples with a date estimate that overlaps the target range for each average—that’s why the numbers in white add up to more than ninety-seven.  The results suggest this was a valid approach; at least, it’s easy to see a change over time, and one that doesn’t seem to be a meaningless artifact of the process.  Early Fayum portraits really do look more like the first average, and late Fayum portraits really do look more like the third average.  So not only does face averaging let us create averages of whole genres; it also lets us see their development over time in a way that would be more challenging to perceive from looking at specimens individually.  Moreover, we can arrange the averages sequentially in a timeline to create an animation of the historical development of the genre, like this:

fayum-animatedThe historical evolution of the Fayum mummy portrait—while visually striking and potentially interesting—may admittedly not be the hottest topic in art history.  But we can use the same technique to illustrate patterns of change in better-known and more intensely-studied traditions.  Here are averages of female faces as depicted in paintings produced in Catholic Europe decade by decade during the late fourteenth and early fifteenth centuries (the number in black shows how many source images were averaged in each case):


This time, to make things more conveniently manageable, I’ve assigned a single date to each of my source images, usually taking the average or median of an estimated date range; for instance, I’ve treated “ca. 1440-1450” as 1445, and “last quarter of 14th century” as 1387.  I’ve also done my best to select only female faces, which are harder to isolate from beardless male faces and genderless angel faces than you might think.  Otherwise, I’ve drawn indiscriminately on images of saints and secular portraits (the Virgin Mary is grossly overrepresented), on both young and old women, on both main subjects and figures in larger scenes, and on varied geographic regions and artistic styles.  As long as I could find a decent-resolution color image of an identifiably female face oriented not more than around 45 degrees away from the viewer and dated within a range of fifty years or less, I threw it into the mix.

There’s a gradual but perceptible change here from the 1370s to the 1450s: foreheads become more exposed, eyes become more downcast, and eyebrows become more elevated.  But there’s also some unevenness in the data.  I had more trouble finding images for some decades than others, with the 1410s being particularly poorly represented.  I suspect this may have less to do with actual production, survival, and selection rates than with how art historians go about estimating the dates of paintings.  The 1410s seem to coincide with a moment of transition between two relatively distinct norms, for example, so perhaps there’s no “look” commonly associated with that decade, leading art historians to favor the 1400s or 1420s instead when assigning dates.  Apart from that, the transitions from average to average are choppier than with my Fayum animation, partly because each source image is averaged into one and only one of them, with no data carrying over from one point to another to smooth things out.  Grouping data strictly by decade is arguably a bit misleading, too, since many of the source images are datable only to a range of twenty, thirty, or even fifty years (e.g., “mid-15th century,” which I treat as 1450).

So here’s another experiment in which I’ve averaged the faces in overlapping thirty-year groups while still shifting forward in time at ten-year increments:

30-year-groups-1370to1449This approach has the advantage of drawing on a greater number of examples for each average, never dropping below 100.  It illustrates the same overall trajectory as before, but more smoothly, enabling us to create a nice animation (drawing here on data from the full hundred-year period 1360-1459):1360s-1450s-animatedThe dark triangular regions in the upper and lower left corners are artifacts of cropping and rotation, but I don’t find them too distracting.  Extending this animation to cover a longer span of time—say, 1250 to 1850—will only be a matter of time and work.  The display could easily be slowed down or sped up as desired.

Overall, the combination of face averaging with a time base seems to offer a wonderfully vivid means of animating the history of certain interconnected traditions of art, culture, and fashion, within whatever chronological, generic, or geographic limits we choose to set.  I’m trying to think of a good name for it.  The term prosopography (prosop- = “face”) has been used to refer to the writing of historical collective biographies, which has some vague affinity to what I’m trying to do here, but apart from being already “taken,” it doesn’t explicitly invoke the element of change over time.  So I’m leaning towards prosopochrony to reflect the concept of representing faces as a function of time; that would make me a prosopochronist, and these would be prosopochronic animations.  Anyone care to weigh in?

10 thoughts on “Face Averaging and Art History

  1. The images are magnificent! I found the blog by Googling Fayum portraits (a long time fascination of mine), and ended up meandering happily through your blog. Delightful.

  2. Pingback: “Say Cheese!”: Using Face Averaging to Track the Rise of the Photo Smile | Griffonage-Dot-Com

  3. Pingback: What is Eduction? | Griffonage-Dot-Com

  4. Pingback: Historical 3D Fun With Anaglyphs | Griffonage-Dot-Com

  5. Pingback: The Fashionable Face: A Work in Progress | Griffonage-Dot-Com

  6. Pingback: My Fiftieth Griffonage-Dot-Com Blog Post | Griffonage-Dot-Com

  7. Pingback: Averaging Faces in Profile, and Other Things | Griffonage-Dot-Com

  8. Pingback: Improvements in Face Averaging | Griffonage-Dot-Com

  9. Pingback: Time-Based Averaging of Indiana University Yearbook Portraits | Griffonage-Dot-Com

  10. Pingback: Upside-Down Audio Revisited: A Step Forward in the Quest to Turn Music on its Head | Griffonage-Dot-Com

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

This site uses Akismet to reduce spam. Learn how your comment data is processed.