An algorithm for style: How AI is reimagining fashion design
New fashion line Glitch uses deep-learning algorithms to reinvent the little black dress
Two computer scientists in the United States have developed an algorithm for style and are reprogramming how some of the fashion industry's timeless garments are made.
Emily Salvador and Pinar Yanardag, co-founders of a new fashion line called Glitch, are retiring the need for hand-drawn sketches in exchange for artificial intelligence-based designs.
The sentient software takes over the initial design process, Salvador explains, and is reinventing the eternal style of the little black dress.
"We thought it would be a great idea to try to recreate that using artificial intelligence," she said.
This article of clothing has been widely considered a fashion staple since it was popularized by world-renowned designer Coco Chanel in the mid-1920s.
How it works
AI is typically used to power everyday web and smartphone applications from voice, image and facial recognition to language translation.
The technology, however, is increasingly being used in more complicated tasks, such as generating art, creating text and diagnosing cancer from images.
Fashion is no exception in this sweeping revolution of machine-driven intelligence, according to Yanardag and Salvador.
Everything from the bell sleeves to the diagonal hemline on a little black dress made by Glitch is designed by a machine-learning algorithm known as a generative adversarial network, or GAN — which generates new images based on existing patterns.
In order to draft the architecture of each garment, Salvador explains, one computer analyzes a collection of vintage dress sketches while another uses this data to determine if it's the original dress design or a "photo-realistic dress" that has been augmented.
"They go back-and-forth with one another until it produces new results that look similar to the original data set, but aren't the same," she told Spark host Nora Young.
"We were getting some dresses that were pretty bizarre. Like some of them have an asymmetrical sleeve on one side, or a bell sleeve, or a weird hemline, or maybe there's a piece cut right out of the middle," she continued.
The designs aren't completely built by machine-learning, however.
The computer, for example, can't imagine the fabric or material that will be used when the dress is built, Salvador points out, which allows for interpretation.
"The human collaborator makes an assessment of if this dress is made of a velvet material or of a rayon or cotton or what have you not," she said.
"I think it's an exciting opportunity for an AI system to generate a visual output, but then for a human to come back and interpret what it sees."
Future of AI in fashion
While AI excels at analyzing thousands of fashion designs at once, pulling out the best features of each and then quickly generating a mock-up design, Yanardag adds, the algorithm can't totally replace people.
"We have our creativity, our emotions, our experiences as a human being. So combined with this AI-powered creativity, I think the future is very exciting. We can work with AI as collaborators, not as competitors."
The pair met at MIT in Cambridge, Mass., through an initiative called "How to generate (almost) anything," where participants apply deep-learning systems to a variety of creative projects including food recipes, perfume scents, jewelry designs and fashion creations.
The idea for Glitch, Yanardag told Spark, was born from the project's AI-inspired creations.
"After doing a bunch of different designs, we decided that actually we should start a company," she said.
Glitch currently sells 100 dresses in its little black dress clothing line. The company also allows customers to upload and generate their own designs that can then be voted on by others.
"We are hoping to use this data and analyze it, and even use it to produce research where we will be investigating whether there's a cross-cultural preference over fashion designs, and maybe we can infer some different trends based on the country that these people are voting from," said Yanardag.
"Based on these preferences we can personalize what we are selling."
Click 'listen' near the top to hear the full interview
Written by Amara McLaughlin. Produced by Rachel Matlow.