The Virtual Black Hair Experience: Evaluating Hairstyle Transfer Generative Adversarial Networks on Black Women

Abstract

There’s a history of Black Women being left out, reduced, or undermined as users in recent technological advances. The goal of my project is to analyze this history connected to today's cutting edge Hairstyle Transfer Generative Adversarial Networks by centering Black Women as subjects of my work. I first generate images from Figaro’s hair type dataset using MichiGAN software. I hypothesize that since MichiGAN was trained on an image dataset that underrepresented Black Women, the Black Hairstyles it generates would be perceived as less real than White generated hairstyles due to the models training as well as because of racially constructed bias. What I found, however, after collecting results from two Amazon Mechanical Turk experiments is that workers perceived generated Black hair styles as more real than White ones. At first glance, this result seems to disprove the existence of a negative racial connotation between the model’s execution and subsequent testing. However, with a deeper dive into the data, MichiGAN performed well due to the restrictions on and boxing of Black hair to certain categories, thus detailing another layer of Black Women reduction in data and technology.