Amid the FaceApp backlash, an AI portrait generator has largely flown under the radar. Its unique selling point is its ability to turn any selfie into a 15th century European portrait. In order to properly generate portraits resembling the works of old masters like Rembrandt or the more garish works of Van Gogh, “tens of thousands of paintings” were used to teach the machine learning model. They ranged from the Early Renaissance to as late as the Contemporary Art period. All a contemporary internet surfer has to do is upload a selfie to embed their likeness into the annals of art history.
Would be selfie uploaders of color may immediately point out that the subjects of European painters were primarily white. Data scientists call this flaw a bias. Due to the available data set, the AI will be biased towards a certain visualization of the human anatomy. The auteurs of this machine learning model acknowledge their model’s bias and claim that it is one of the most notable lessons to take away from their project. For example, since smiling was considered too comical for a portrait, you will be hard pressed to find a 15th century portrait featuring a face-distorting smile. Rather, portraits of the time were meant to be a stark, albeit slightly idealized, representation of the patron. The model reflects this style be converting a full smile to a faint closed-mouth smile. In this way, the team behind AI Portraits want you to learn something about Art History by experimenting with their bias.
Still, however good their intentions may have been, it has raised eyebrows among those of color. Morgan Sung, a writer for Mashable, recently demonstrated the ways in which Portraits misrepresented certain physical features. She quoted Coin Center’s Director of Communications who said that his portrait made him look “more Italian than Indian.” And later on pointed out that BTS’s Jungkoo’s monolids were replaced with lidded eyes. Darker skin color, as is to be expected, was lightened.
Credit: Getty Images
A rudimentary understanding of the process in which a model is trained allows one to realize that the above biases are the direct result of the limited data set. The solution could be to broaden the data set to include paintings from other cultures similar to those that the AI Portrait team referenced in their brief tour through the history of portraits.
credit: AI Portraits
The team doesn’t directly state the reason for not choosing other data sets, though they mention that according to art historians like Joanna Woodall, Shearer West, and John Berger 15th century European art is the “inflection point” of portraiture. The implication here is that there is a wide swath of data available in this period to train a model. To compensate for the lack of cultural diversity, there is a wide range of styles that do potentially reduce the color bias somewhat. For example, Van Gogh often used dark colors in his composition.
credit: AI Portraits