The Twitter Bot That Turns Your Photos Into Art History
The most fascinating thing on Twitter right now is @DeepForger an AI bot that replies to photos you send it with a re-rendering “forgery” in the style of a classic artist. For example, here is the “wtf ????” guy meme in the style of John Singer Sargent, in his circa 1881 style:
I know what you’re thinking: this is the same as those photobooths at Chuck E. Cheese in the ‘90s where you can select if you want your photo in the “Leonardo da Vinci” or “Monet” style. Well, friend, you’re wrong. It’s way more complicated.
The bot is the work of Alex J. Champandard of Vienna, Austria, who is an artificial intelligence consultant and organizer of the nucl.ai Conference for developers in creative industries on artificial intelligenct and machine learning. He previously worked as an AI programmer at Rockstar Games Champandard created the bot as a sort of sandbox to play around with new research in his field. It builds off the work of a Cornell University academic paper called A Neural Algorithm of Artistic Style that described how machine learning could analyze artistic visual patterns.
“It looked pretty impressive,” Champanard told BuzzFeed News in an email. “And given how fast things were moving in the field of machine learning, I thought: Someone's going to implement this as described in the paper, how can I take it further? I realized the best would be to help the community explore the space of computational art (as made possible by this algorithm) then finding out what works and what doesn't.”
Just tweet a photo at @DeepForger. You can request a certain artist’s style by adding their last name. To improve your results, you can also add tags like #landscape, #portrait or #people to give it more hints. The bot searches for classic artworks to mimic through the Metropolitan Museum of Art’s online collection and its Twitter account that tweets out random art images. Then, it implements an algorithm from the research paper to detect patterns like color and style, as well as things like brush strokes or grain textures. Finally, a second process combines the style of the classic painting with the photo you fed it, and voila – a forgery.
Of course, it’s slightly more complicated. Okay, more than slightly.
And it’s getting really popular. As of this writing, the bot has a wait time, because each forgery takes between 3-8 minutes to process, and requires 4GB of memory. (There are ways, however, to speed it up: you can quote-retweet your request, for example).
Some of the earlier forgeries did not turn out so well, like this ghoulish portrait of Marilyn Monroe in Lichtenstein style:
Yet Champandard is encouraged by how the bot has been improving. (Take, for example, the portrait below of BuzzFeed’s Mat Honan.) It’s learning.
“It's hard to understand what the implications of the algorithm are,” he said, “but it could change everything about the way artists interact with their fans. Imagine digital artists being able to provide fully customized paintings of people's cats or dogs in their own style? It's also helping artists discover new possibilities; game developers submitted screenshots of their game to see what came out—as a tool for experimentation and creativity.”