In 2015, a team from the University of Tübingen published really neat algorithm that uses a deep neural network to copy a painting style and apply it to your favorite photo.
A bit like Google’s trippy Deep Dream project but “with the craziness toned down a little bit” is how Quartz reported it. Instead of reinterpreting everything as dogs at a psytrance festival on a beach in Goa, it copied stylistic elements like brushstrokes and lines.
I thought it was cool, but it didn’t look like it had many applications beyond stylizing your photos. Then in 2017 user vic8760 posted this incredible image on Reddit.
User vic8760 had repurposed the German team’s style transfer algorithm. They didn’t apply a striking art style to a target image, but used two paintings that were fairly similar in style: a neoclassical portrait of Napoleon, and a High Renaissance painting of a crowd scene (Raphael?). Instead, they transposed the content by replacing most of Napoleon and his steed with drapery.
The result is an image that doesn’t just caricature a well-known art style—it pioneers an art style of its own. One that’s distinctly algorithmish.
I started experimenting with this content-transposition approach on the cool free platform deepart.io (go try it for yourself!) I soon discovered the results could be deeply unsettling—not so much an uncanny valley as an endless, Lovecraftian abyss.
We maybe need a name for this kind of stuff that is not only scary but differently scary from human-created imagery — I’ve suggested algo-horror. Check out my Twitter account @algohorror where I post and retweet similar kinds of stuff. I’ll be writing a post on the aesthetics of this genre soon.