Machine Churns Out AI-Generated

In honor of Halloween, researchers at the Massachusetts Institute of Technology Media Lab launched the Nightmare Machine website, which allows visitors to vote on AI-generated horror images created via an open source deep neural network algorithm developed last year.

Scientists from the Australian Commonwealth Scientific and Industrial Research Organization (CSIRO) collaborated on the project.

The Nightmare Machine features some 200,000 images of normal human faces that were fed into the neural network by MIT researchers and CSIRO’s Data61 digital and data innovation group.

The neural network algorithm transforms images of normal-looking human faces into those with the characteristics of cinematic zombies, such those featured in the popular AMC TV show, The Walking Dead.

To date, the Nightmare Machine has collected more than 300,000 individual votes, along with user responses suggesting that the digitally modified photos can be quite terrifying indeed.

Portraits and Landscapes

In addition to faces, the Nightmare Machine is able to take postcard perfect images of famous landmarks — New York City’s Statue of Liberty, Germany’s Neuschwanstein Castle, or Paris’ Louvre Museum, for example — and transform them into utterly terrifying settings.

The Nightmare Machine transformed this image of the Statue of Liberty into the spooky rendering shown above.

The way the Nightmare Machine transforms the images is explained in the paper, “A Neural Algorithm of Artistic Style,” published last year by researchers Lean A. Gatys, Alexander S. Ecker and Matthias Bethge — all of the Werner Reichardt Centre for Integrative Neuroscience and Institute of Theoretical Physics at the University of Tubingen, Germany.

The photos can be manipulated to evoke an emotional response based on utilizing predetermined filters, the researchers noted.

Traditional photo-manipulation and editing tools, such as Photograph, offer an array of stylish tools. Likewise, the Nightmare Machine features its own stylish flare, highlighting the not-so-subtle differences between a “haunted house” versus a “fright night” or “toxic city” treatment.

Each is something that would be worthy of a lingering bad dream, and all highlight the way an algorithm can adjust its nightmarish effects in very different ways to induce particular feelings in viewers.

“This technique for using deep neural nets to analyze what constitutes a scary scene and then apply a ‘scariness filter’ to other photographs can be used in a wide range of other applications,” said Paul Teich, principal analyst at Tirias Research.

Gothic Architecture

However, the key to what constitutes something that is truly scary may be in keeping it subtle — perhaps by employing a thus a suspenseful build-up evocative of Stephen King or H.P. Lovecraft rather than depicting a zombie-inspired bloodbath.

“The scariest of these applications is applying just a little horror to everyday scenes or even news items — not enough to identify what is scary, but enough to sway the emotions of a viewer,” Teich told TechNewsWorld.

“These techniques can apply to real-time video as well, so a scene can be made progressively more or less scary to direct a viewer’s or player’s attention and interaction with scenes,” he added.

Not Every Day Is Halloween

The researchers at MIT and CSIRO used Halloween to show off the ability to transform the beautiful into the horrific, but the technology’s designers also laid out how it could be used in other creative ways.

The frightful images are a Halloween-inspired application of a “general purpose AI machine designed to elicit human reactions,” said Roger Entner, principal analyst at Recon Analytics.

“I would not be surprised if on Feb. 14, there will be a press release with the MIT AI Love Machine designed to make people feel romantic,” he told TechNewsWorld.

However this technique could be used in other emotional contexts as well, “so that the entire human emotional range can be manipulated subtly and in real time while viewers are watching their newsfeed, advertisements, or playing a game,” suggested Tirias’ Teich.

“It will make immersive content that much more engaging, sticky — and perhaps much more addictive,” he observed.

The applications for this could this could be many fold, “especially for the entertainment industry — ranging from movies to VR,” said Entner. “All in all, we are early on, and this is designed to drive attention to the work MIT is doing.”