Artificial Intelligence is getting really, freakily good at recreating the human face, as several viral videos of "Tom Cruise" using low-cost, fairly basic deep fakes have recently shown.
One cool new area that generative adversarial networks (GAN) have been applied to in photography is to restore old photographs and footage, upscaling the quality as well as depicting the world more similarly to how it would have appeared at the time.
A team of researchers from the University of Washington, UC Berkley, and Google Research has demonstrated their own AI program that does just this, showing historical figures such as Mao Zedong, Ruth Bader Ginsberg, Andrew Johnson, Marie Curie, Thomas Edison, Huiyin Lin, Benjamin Disraeli, and Mahatma Gandhi as if they were photographed by cameras we have today.
"Our understanding of [Lincoln']s appearance is based on grainy, black and white photos from well over a century ago," the team explains in their paper, published on pre-print server Arxiv.
"Antique photos provide a fascinating glimpse of the distant past. However, they also depict a faded, monochromatic world very different from what people at the time experienced."
One way that old photos are distorted is that the negatives of Lincoln's time were only sensitive to blue and UV light, which made cheeks appear darker than they were, and "overly emphasizing wrinkles by filtering out skin subsurface scatter which occurs mostly in the red channel."
Because of this, Lincoln appears in these old photos looking a lot more wrinkled (and a lot less sexy) than he would have in real life.
Recreating what historical photographs would have looked like has a number of challenges – including faded photographs, the limitations of cameras of the day, and how the film was developed at the time. A traditional way of restoring them is to apply digital filters such as noise removal. This is tricky, as from the original photograph, you can't fully tell what has been distorted by the aging of the photograph and other problems such as color sensitivity and development processes.
"Instead, we propose to project the antique photo on the the space of modern images, using generative tools like StyleGAN2," the team wrote in the paper. "In our method, we first generate a 'sibling image' by projecting the black and white photo into the StyleGAN2 latent space, which resembles some characteristics of the input but often has a different identity. We can create the interesting effect of morphing the 'modern' sibling to the output for the 'historical' figures," they add in the description of one of their videos.
The "modern sibling" here is another model with similar facial features, recreating the pose of the historical person. Once the historical face has been projected onto the sibling face (which has better resolution and more realistic coloring), it is then morphed again to look like the original face.
The result is that you can see historical figures in high quality, looking much closer to how they would appear on modern cameras. Check them out below.