COCO CHANEL had it. Yves Saint Laurent, too. And Gianni Versace without doubt. Originality put these designers into the fashion history books. A team at Facebook hopes to use artificial intelligence to take fashion in bold new directions as well.
Previous AIs have created music, artworks and poetry. Now the Facebook researchers have coaxed three AIs into designing clothing. Overall, they produced around 1000 items, including handbags, jumpers and T-shirts. The aim was to come up with truly original creations that could then inspire human designers.
鈥淭echnology can amplify our creativity,鈥� says Camille Couprie at Facebook鈥檚 AI research lab in Paris. 鈥淚t can take a lot of brain power to think about new ideas, so if AI can help with that and accelerate that process it would be good.鈥�
Advertisement
The team has been talking with a well-known fashion brand about tapping into the method, but the resulting images are currently too small to be useful for a real-world fashion house.
The designs were produced by algorithms known as generative adversarial networks (GANs). These AIs pit two neural networks against each other, one generating ideas, the other judging them. Through thousands, and sometimes millions, of iterations, both sides master their skills. Eventually, something is created that suitably satisfies both the AIs and humans.
鈥淥ne of the AIs had the strange idea of creating some trousers with an extra pair of legs鈥�
But GANs usually produce close imitations of the images they were trained on. True originality, as many fashion connoisseurs would agree, is trickier. To inspire a more creative edge, the team introduced two disruptive functions, which they termed 鈥渃reativity losses鈥�, into the networks. In essence, these confused the AIs enough that they were forced to deviate from existing styles and towards more original content.
Three GANs were trained on around 4000 images of existing fashion items created by humans, learning the importance of texture and shape. Two were off-the-shelf GANs. The third, dubbed StyleGAN, was constrained in the shapes it created so the fashion items would actually be wearable. However, it was given carte blanche on other design elements.
To judge how the AIs had done, the team showed 800 of the images to people for them to review. Almost two-thirds of the designs were judged as being created by humans, and they were mostly considered original too. StyleGAN had the best eye for fashion as rated by the reviewers ().
However, certain designs were a little too novel. For example, one of the AIs had the strange idea of creating some trousers with an extra pair of legs.
It is unsurprising that algorithms can design clothes just as well as humans, says Stevan Harnad at the University of Quebec in Montreal, Canada. But he wonders whether people might eventually become bored of AI designs. 鈥淭he human behavioural side of this study did not get far enough to see whether human observers would eventually have detected something mechanical and repetitive in the designs,鈥� he says.
Read more: Artificial intelligence tells nightmare-inducing tales of terror; Google鈥檚 neural networks invent their own encryption; AI artist conjures up convincing fake worlds from memories
This article appeared in print under the headline 鈥淔acebook鈥檚 AI fashion designer鈥�
