The T-shirts sold by Cross & Freckle, a New York–based fashion upstart, don’t look revolutionary at first glance. They come in black or white, they’re cut for a unisex fit, and they sell for $25. Each of them has a little design embroidered into the cotton that references staples of New York City life: pigeons, dollar pizza slices, subway rats.
Sarah McBride, the brand’s cofounder, likens the T-shirts to “the New York City summer uniform.” But McBride didn’t actually design the shirts. Neither did anyone else. They were designed instead by a neural network, which crunched doodle data from millions of people and spit out the original art that makes up the embroidery.
Cross & Freckle isn’t the first company to use AI to generate art—people have been doing that for years. But this project offers a glimpse into the nascent world of AI-generated fashion, where designers use machine-learning models to remix and riff on old designs. Cross & Freckle doesn’t just use AI to create its designs; it also got the brand’s name and logo from a neural net, called the Hipster Business Name Generator and used an AI text generator to create the mumbo-jumbo marketing copy on the company’s website. It’s a new model for a brand that relies entirely on AI.
The embroidered designs look like something a child might have scribbled with a crayon, or something an adult might have drawn with a computer mouse on Microsoft Paint. That’s actually not far off: The graphics came out of a variational autoencoder trained on data from Google Creative Lab’s game Quick, Draw, which has collected over a million doodles from people around the world. Quick, Draw works like Pictionary: Draw a giraffe or a light bulb or a slice of pizza, and the computer guesses what it is. In aggregate, the drawings make up what Google calls the “world’s largest doodling data set.”
Paul Blankley, who handles Cross & Freckle’s technical side, created the autoencoder and fed it doodles of pigeons, rats, pizza, and dogs from Google’s data set. “There are hundreds of thousands of drawings for each of the different categories, from countries all over the world,” he says. “So you get this cool mix of what does the world think it means to be a doodle of a pigeon, or a doodle of a piece of pizza.” The autoencoder riffed on those drawings, creating its own original designs.
“What’s great with generative AI is that it will churn out a bunch of different options so you can choose what might be the best version of a puppy or pizza slice for a T-shirt, or which hipster name is the most random but still real-sounding,” says McBride. “So there was still an element of human selection and curation involved.”
The fashion industry already relies on artificial intelligence to fine-tune various parts of clothes-making, from managing inventory to predicting trends to offering recommendations to customers. “Virtual stylists” help people find their ideal size when shopping online, and data-driven recommendation engines help surface things that fit into someone’s personal style. Brands big and small are searching for ways to incorporate machine prowess into the process. Just last week, Amazon introduced StyleSnap, a tool that uses computer vision to find items of clothing similar to the styles in a photograph—the latest push to get people buying clothes on Amazon.
But AI also offers new opportunities for designers who are looking not just to forecast the trends, but to create original designs. Glitch, a company created by a pair of MIT graduates, sells dresses designed by deep-learning software. And last year, just in time for New York Fashion Week, artist Robbie Barrat designed an entire collection for the fashion house Balenciaga using a neural net. The program scraped images from Balenciaga’s lookbooks, catalogs, and runway shows, then remixed them to create original styles. The resulting images are strange; the models’ faces, in particular, smear into each other in the AI-generated photos. But the designs themselves are cool and push the boundaries of fashion in interesting ways.
One image shows a runway model sporting a three-component coat made of black satin and bright pink fur. Another reimagines a windbreaker in the shape of a classic button-down. They’re weird, but so is high fashion—and you could imagine many of these looks appearing on a real Balenciaga runway.
“The most interesting aspect of using AI to generate media is that it allows us to create shapes that have never been created before, while also giving us a look into an AI’s interpretation of what a dog, pizza slice, pigeon, and subway rat should look like,” says McBride. “The AI becomes our cocreator.”
With the help of AI, designers shift into a role that’s more curatorial and maybe even more creative. Whether we choose to wear those creations is up to us.