Facebook researchers have figured out how to fix bad photos by replicating people’s facial features, and the results are disconcertingly accurate.
In a paper published on Facebook’s research site, and spotted by Motherboard, Brian Dolhansky and Cristian Canton Ferrer show how an “Exemplar Generative Adversarial Network” (ExGAN) can be trained to accurately retouch images, all while avoiding the “uncanny valley” effect.
In order to work properly, it first needs to look at reference pictures. In this case, those reference pictures need to show you with your eyes open.
For the study, Dolhansky and Ferrer trained the ExGAN on around two million 2D-aligned images of around 200,000 individuals. There were at least three pictures of each individual.
Impressively, but also very very eerily, people struggled to differentiate retouched images − aka pictures featuring digitally created eyeballs − from real photos.
“In order to further verify our method, we performed a perceptual A/B test to judge the quality of the obtained results,” the paper says. “The test presented two pairs of images of the same person: one pair contained a reference image and a real image, while the other pair contained the same reference image and a different, in-painted image.
“The photographs were selected from our internal dataset, which offered more variety in pose and lighting than generic celebrity datasets. The participants were asked to pick the pair of images that were not in-painted. 54% of the time, participants either picked the generated image or were unsure which was the real image pair.
“The most common cause of failure was due to occlusions such as glasses or hair covering the eyes in the original or reference images. We suspect that with further training with more variable sized masks (that may overlap hair or glasses) could alleviate this issue.”
Not creepy at all, Facebook. Not creepy at all.
What’s the last thing Facebook did that didn’t creep you out? Share your thoughts with us @TrustedReviews.