Jan 26
Is Dall-E the Next Dior? How AI Is Trying to ‘Make It Work’ in Fashion
AI companies Stable Diffusion, Dall-E 2, and Midjourney are trying to be the next big thing — but while imitation is the sincerest form of flattery, it’s the lowest form of fashion design.
By Chandra Steele
Ever hip to a trend, fashion has been at the forefront of much tech innovation. Diane von Furstenberg put Google Glass on the runway. Rebecca Minkoff’s flagship location transformed into a “ connected store” complete with dressing rooms outfitted with interactive mirrors. And the metaverse is filled with luxury boutiques that are shopped by avatars wearing Gucci and Prada. But how welcoming is the industry going to be when AI takes over and starts designing fashion of its own?
That day is nearly here, as the AI art generators that have been stealing work from artists and even winning art competitions begin to move on to new industries. Fashion is a natural transition, and the signs that Stable Diffusion, Dall-E 2, and Midjourney are vying for a top spot in the industry are already there.
The eerie AI people populating social media feeds are wearing outfits that have been dreamt up by text-to-image generators, but real fashion is a step beyond just throwing something on to go out. When AI is tasked with creating clothing, it’s proving to have similar unsettling but intriguing results as it does with everything else it works on. And there are the same concerns about how much of AI’s fashion creations are original and how much are theft.
You Can’t Cop These Drops
AI is starting at the bottom when it comes to fashion design. Literally, that is, with sneakers. This sneaker does not exist scrolls through sneaker designs that you can’t buy anywhere. They’re the creation of a neural network that Applied AI Engineer Stan van der Vossen trained with images of 50,000 existing sneaker styles. You can collab with the AI by clicking on a shoe and using a slider to adjust colors as well as how futuristic and creative the design is.
The sneakers that float by on the site look much like the ones you’ll find sitting on shelves at Foot Locker. Even some brand elements remain, with Nike swoops and NBs from New Balances still apparent. It raises the idea of how easy it would be for these companies to assign the design of shoe variants to a program instead of a person.
Similarly, developer Niels Garve used a database of existing sneakers to create AI Sneaks, a collection of limited-edition NFT sneaker designs. Even though fans can’t wear these shoes, they can buy them in the virtual sense and equip their avatars with footwear. Limited-edition designs are available on OpenSea in exchange for Ethereum.
But some AI sneaker designs are destined to belong to everyone and no one, driving hypebeasts crazy, since they can’t stand on long lines waiting for a drop or fight the bots for them on StockX. Social media is filled with sneakers that are just there to be scrolled past: lacy pale-coral Nike; Adidas and Asics enhanced by pearls, pom-poms, and a bit of sparkle. Frustrated commenters post repeatedly, begging to be able to buy them: “Are these for sale,” “Can I orderrrrr🔥🔥😢,” “I don’t get why they post them but no way to place an order. On top of that nobody responds when people ask how to place an order.”
Not Creative Director Material
The account churning out these in-demand, nonexistent sneakers is @ai_clothingdaily on Instagram. You can watch some of the process in posts, with Stable Diffusion morphing John Galliano’s S/S 2005 collection into something new or riffing on Hussein Chalayan’s already alien-looking S/S 1998 runway.
The reworked runways are the work of Field Skjellerup. Skjellerup told us he likes to blend classic silhouettes with materials that might not be usable or functional in reality, so he experiments with keywords that combine these concepts until he finds an image worth sharing.
Skjellerup is self-taught. He said his exploration of AI and fashion comes from a strong personal interest in both. While he does not have the capabilities to produce the garments that commenters are clamoring for, Skjellerup told us he hopes to collaborate with a large fashion house or athletics brand.
“Since the public AI systems are all trained on the same data set, the aesthetics will be shared by multiple users. That’s why you have to use it as starting point in the creative process, not the end.” — Paul Trillo
Subscription-box styling service Stitch Fix has put AI to work as everything from a stylist to an assistant designer. After customers fill out a style questionnaire or the Ask a Stylist form, their information is vetted first by AI and then humans. Instead of being siloed, the AI is more of a team member, both expanding its knowledge while it observes the choices made by stylists and sharing its skill of objectivity by giving feedback on customer reactions and behaviors.
The company has used Dall-E 2 to turn customer input into images that stylists can then use to search Stitch Fix’s inventory for matching items. For this process, Stitch Fix uses GPT-3, which is a program that can understand the semantics of human speech and generate conversational text (think ChatGPT) to come up with what customers are looking for; it then inputs those words into Dall-E 2. Using the terms high rise, skinny, great color, red, fun color, great fit, stretchy, without distressing jeans, Stitch Fix was able to create an image of the desired item, which resulted in sourcing its real-world counterpart.
Stitch Fix is also using AI to help design its exclusive brands, 41 Hawthorn, Goldray, and Market & Spruce. In its introduction to the brands, the company says, “They’re the brands that we brainstorm, design, (sometimes) collaborate with our data scientists and bring to life—all under our very own roof.”
AI is not leading any fashion houses yet, but a Hong Kong-based AI lab, AiDLab has developed software meant to aid fashion designers. AiDA, which stands for AI-based Interactive Design Assistant, was employed by 14 designers to collaborate on 80 outfits that hit the runway at the Fashion X AI show in Hong Kong last month.
AI is not leading any fashion houses yet, but a Hong Kong-based AI lab, AiDLab has developed software meant to aid fashion designers.
Human designers input their own sketches, the materials they want to use, and a color palette to a “mood board,” and like a dutiful design assistant, AiDA pitches its designs. A promotional video from the lab stresses that AiDA is a tool and not the ultimate designer: “AiDA empowers fashion designers to work with AI for speedy ideation, while at the same time ensures signature design elements of particular fashion brands are realized in the new designs.”
Fashion brand Finesse employs AI as a trend forecaster. Finesse was founded by Ramin Ahmari, who has undergrad and graduate computer science degrees from Stanford University, where he specialized in AI; the company has a production process that relies on computers and consumers.
Each potential design is posted on the Finesse website in three versions, and site visitors vote for their favorite. Once the votes are tallied, the winning item goes into production and is ready to be sold in two weeks. AI is used to gauge how many of each piece should be produced, something touted by the company as making it “sustainable AF,” since it cuts down on waste.
Even the copy on Finesse’s About page seems like it was designed by AI to capture the zeitgeist or, more likely, crawling search engines:
“Finesse is Zendaya, shaking up the Hollywood Belle stereotype with her beautiful brown skin. Finesse is Bella with the lollipop and GiGi with the headscarf. Finesse doesn’t give a f*ck about opinions, because Finesse isn’t afraid to be themselves. Finesse is sustainable because people messed up this earth for long enough. Finesse makes you break up with your girlfriend because she’s bored. Finesse doesn’t believe in gender. Finesse clearly doesn’t care about your preconceived notions of beauty, sis. Finesse is Hunter Schafer and Barbie Ferreira on a neon night out.”
The cultural references just flash right by in that brand statement, much like the Dall-E-designed outfits in a video that filmmaker Paul Trillo created, in which a model’s clothing changes many times over (scroll down to his tweet to see the video). To make the video, Trillo said that he shot an image sequence and used Dall-E to generate a new outfit for each frame. After generating hundreds of outfits, he curated the ones he liked and strung them into a sequence.
Trillo said that as a creative, he sees the potential in AI as a tool for brainstorming ideas and exploring new routes in the creative process. “The efficiency of trying out ideas is unmatched,” he said.
But Trillo has caveats about the limits of AI’s creativity. “The obvious downside is that larger companies that see efficiency may also think they need less workforce,” he said. AI and humans are not interchangeable as designers, though. While humans can go out into the world and imagine something new, AI is stuck with the same pool of inspiration made up of data sets.
“I think people need to be careful of is the potential homogeneity of work. If you’re using AI image generators as the final output for something, it’s going to look just like the output of someone else using the tool. Since the public AI systems are all trained on the same data set, the aesthetics will be shared by multiple users. That’s why you have to use it as starting point in the creative process, not the end.”
Everything Old Is New Again (and That’s the Problem)
Because AI is remixing, not designing from scratch, it brings up thorny moral and costly legal issues around who should be credited and compensated for its designs. But nothing is truly new in fashion, including this dilemma.
Fashion designers are influenced by or steal from one another all the time. Sometimes it’s a dialogue through the decades, sometimes it’s outright theft, and sometimes it’s a fast-fashion knockoff. Often the courts are left to decide—both the actual courts and the courts of public opinion, which heavily influence the industry.
Another way that AI is derivative is how it is yet another way fashion has capitalized on the prescience of Dapper Dan, who started his career in the 1980s by ripping up logoed couture and crafting an entirely new street-style look. He was sued into near-oblivion. But decades later, fashion houses brought him into the fold to introduce new ideas into the industry.
Deep Objects, a “decentralized design studio” that uses AI in its work, has released a series of what it calls white papers that are more like manifestos about AI and fashion. In Issue 001, it posits that maybe we have already been displaced in the grand design: “Through deep learning, machines have found their muse. By training AI on centuries of things we’ve thought and made, what we’ve unwittingly done is turn the lens back on ourselves.”
But AI is just that, a lens that examines what is already there. Despite ideas that echo through the timeline of fashion history, there is genuine progress. Coco Chanel freed women from corsets, Stella McCartney is revolutionizing mycology and accessory production, and Iris van Herpen has turned 3D-printed filaments into couture. AI does not have the capacity to go beyond what is input into it. It does not answer to social and political influences, develop new materials, or create technologies of its own. Its lack of imagination is where it comes apart at the seams, and for now, it is closer to a needle and thread than to a designer cut of whole cloth.
Originally published at https://www.pcmag.com.
