The fashion exec’s guide to generative artificial intelligence

This article on generative AI is part of our Advanced Membership package. To enjoy unlimited access to The Long View from Vogue Business and bi-monthly Market Insights Reports and webinars, sign up for Advanced Membership here. Generative AI has quickly captured the attention of fashion and beauty executives who are excited by its potential to aid in both creative efforts and in more practical ways, including building imaginative ad campaigns, improving customer service interactions and streamlining the process of writing product descriptions. AI in fashion, beauty and retail is not new. However, the range of generative AI tools available now are far more advanced, and far more accessible, than previously possible. In 2022, consumer-facing image-generation tools, including Dall-E 2 (created by OpenAI) and Stable Diffusion (created by Stability AI) became widely used, followed by the release of OpenAI’s chat interface ChatGPT, which gained 1 million users in five days. Since then, an improved version of OpenAI’s GPT (or “generative pre-trained transformer”) model, GPT-4, became available, and OpenAI launched an API in March that enables businesses to integrate it into their apps, websites, products and services. Google released its conversational AI service called Bard, and companies including Microsoft, Salesforce and Amazon have released generative models. Amazon’s cloud customers can now use generative AI to produce text and to improve web-search personalisation, while Shopify merchants can use ChatGPT tech to write product descriptions. While it’s still early days in terms of quality, the more that brands and consumers use and refine the tools, the better they become. Fashion and beauty experimentation has begun: Revolve and Prada Beauty have created marketing imagery, Kering and Zalando have introduced customer-facing chatbots, G-Star has designed a denim collection, Levi’s has created e-commerce models and Adore Me is using it to write product descriptions. “We’re in this phase where everyone’s expectations are so high and I’m trying to bring people back down to Earth,” says Adore Me VP of strategy Ranjan Roy. The global market size of generative AI is expected to grow from $7.9 billion in 2021 to $110.8 billion by 2030, according to Acumen Research and Consulting. Generative AI could add between $150 billion and $275 billion to the apparel, fashion and luxury sectors’ operating profits in the next three to five years, McKinsey says. “I never thought I’d see something like this in my lifetime,” says Rupert Breheny, a Google technical specialist in AI augmented creativity, who has experimented with creating never-before-possible fashion imagery. These recent developments have led to a wider interest in AI in general. Mentions of AI in earnings calls overall have risen almost 79 per cent in the past year, while 22 per cent of the startups in the winter 2023 cohort of technology startup accelerator Y Combinator were focused on generative AI, according to CB insights. Creative industries such as entertainment, fashion and marketing will be the most likely to feel the impact of generative AI first, MIT Technology Review predicts. And, 74 per cent of retail and e-commerce companies plan to increase their spending in AI in the next three years, according to AI data training and software firm Scale AI. It found that while every industry is planning to increase its AI budgets in the next three years, retail and e-commerce were among the more eager sectors. Companies can incorporate generative AI in multiple ways, ranging from creating their own generative models (which enables more control but is more most challenging and costly) to using open-source models (which still require expertise) or cloud API tools such as GPT-4 (which are faster to use but can be less accurate to specific business needs). It’s also early days in terms of best practices; already, brands and creators are navigating new ethical, legal and technical complexities. Some of these are more conceptual; the term “AI” has long conjured fears of a dystopian future in which robots gobble up jobs and become out of control. Other considerations are more immediate, spanning copyright, identity and the needs to fact-check seemingly expert responses. Experts tend to agree that a major shift is already underway. “We should really be careful because it is the start of something,” says Alessia Glaviano, head of global PhotoVogue and director of the PhotoVogue Festival. “The possibilities are endless, and the only limit is good taste and creativity.” Table of contents: Images and videos Text and chat Regulations and considerations Generative art largely refers to images and videos that are created via text-based prompts. It can be used to generate variations on a provided theme — essentially remixing and reimagining photo prompts — or be used to create myriad combinations of a set list of elements. In fashion and culture, some of the most popular examples include PFP collections, which generated 10,000 versions of avatars based on a set number of supplied traits, or apps that ingested a person’s submitted selfies to create portraits. More recently, publicly available tools including Dall-E 2, Midjourney and Stable Diffusion have enabled people to create what appear to be expert-quality paintings and photography with just a few words. The uses in fashion are still expanding, but already, people are building out full clothing collections, generating diverse ranges of models and creating ad campaigns and photoshoots. Generative photoshoots Generative art can be used for creating imaginative photoshoots and films that are faster, more efficient and more imaginative than what would normally be possible in traditional photography. With budgets shrinking, it can be less expensive to replace a traditional photoshoot with AI-created elements, PhotoVogue’s Glaviano says (costs would largely be focused on the creative team and the software, rather than costs of flying to destinations and location fees). Retailer Revolve displayed a range of billboards created with generative AI that showed models among larger-than-life, mushroom-like landscapes. “We can’t build Chanel runway sets and spend millions of dollars doing these crazy things, but we thought, we can do something really special here,” said Revolve CEO and co-founder Michael Mente. These types of projects can still be complex, Glaviano notes, referencing a May cover of Vogue Italia that portrayed Bella Hadid posing against a background that was created using generative AI programme Dall-E. That process, she says, was quite elaborate, because the team wanted to create something in the most ethical way possible, in terms of creating artificial humans or creating content that is positioned as photography. This included collaborations between photographer Carlijn Jacobs and AI artist Chad Nelson. Similarly, in December, fashion photographer Nick Knight unveiled 8,000 one-of-a-kind surrealist artwork NFTs, made using more than 200 traits encompassing fashion, makeup, hair and nails. In April, Prada Beauty debuted a social media campaign for five of the brand’s fragrances that manipulated traditional images taken by photographer Johann Besse. Hybrid approaches like these, which combine “traditional” methods of creativity with fantastical, generative AI-created elements, might be smart as brands and creators navigate ethical boundaries, Glaviano advises. She adds that AI-generated images might also be better considered illustration, rather than photography. As Knight said, “It is up to the artist to make this digital art as soulful as a painting.” Clothing design Generative art can be used to design, or aid in designing clothing, which can both open up the field to people without traditional skillsets, speed up the prototyping process and inspire creatives by generating unexpected elements. While not every result is perfect, the speed of creation makes it easier to experiment. “Fifty to 70 per cent of the results might be trash, but 30 per cent are not — and it did it in three minutes,” says Anthony Lupo, chair at law firm ArentFox Schiff, which advises clients ranging from Victoria’s Secret to YSL, Diesel and Lacoste on technologies including NFTs and generative AI. Glaviano compares it to the advent of the iPhone, but even more impactful. In January 2023, supply chain startup Cala became one of the first to integrate Dall-E technology into its product offering. People can input text-based prompts into Cala’s proprietary dashboard to generate and modify design suggestions, preview them as illustrations or photorealistic depictions, then proceed to produce them physically. Similarly, 3D product creation company Threekit recently introduced ThreekitGPT, which enables brands and consumers to customise products via AI chat. Tools like these can be constrained to a brand’s product catalogue, meaning that eventually, customers could personally generate garments that are still a fit for the brand’s look and feel. Already, G-Star launched a 12-piece denim collection with cape-like garments designed using AI. Revolve plans to physically produce a small selection of garments that were recently submitted to the first AI Fashion Week, which invited aspiring designers to create collections of 15-130 pieces using AI. The competition included a public presentation in New York in April and a team of expert judges including Tiffany Godoy, Vogue Japan’s head of editorial content; Natalie Hazzout, Celine head of men’s casting; and Erika Wykes-Sneyd, VP of Adidas’s Three Stripes Studio. Not all generated designs necessarily need to be constrained to the physical realm. Avatar platform Ready Player Me allows people to customise avatar garments using an AI generator, and in February, Vogue Runway illustrated menswear trends by commissioning an 11-look digital capsule. Models and people Generative AI can be used to create images of fictional characters, to manipulate images of existing people and to animate a person’s imagery to create videos. This is one of the most intriguing uses for fashion, because it stands to change how brands work with models, but the usage of people makes it inherently more complex. Various companies, including Lalaland.ai and Vue.ai enable brands to generate fictional models who can be digitally dressed in clothing for use in e-commerce and advertising imagery. During the pandemic, the reasoning behind this process became more urgent; brands looked to AI to help regionalise their e-commerce pages, and to relatively quickly generate a broader range of model likenesses. In March, Meta partnered with AI-focused retail automation company Vue.ai to enable brands to generate AI-powered models in ads; this was closely followed in April by Levi’s, which announced that it would begin testing technology from Lalaland that would allow the retailer to display on-model photography on a wider range of models, by digitally dressing fictional people created with AI. The London College of Fashion’s Innovation Agency created an entirely fictional catwalk, complete with AI models and AI clothing, based on a recent LCF MA fashion show. It’s not all fictional personalities. Recently, supermodel Eva Herzigová digitally captured her face, body and movements, with the intention of using the data to recreate her likeness in future ad campaigns and photoshoots. And, beauty intelligence startup Haut.ai’s “SkinGPT” tool uses generative AI to show people what their skin might look like in the future, based on their skincare regime. Read more: AI menswear, Generative art, AI Fashion Week preview, AI Fashion Week review, Eva Herzigová’s digital twin, IoDF wearables The ability for generative AI to create human-like, conversational and believable text has prompted a surge in experimental uses, both customer-facing and internal. Customer-facing tools have largely focused on next-gen chatbots that provide advice, answer questions and help in more free-flowing formats than traditional chatbots. Internally, brands are using tools such as ChatGPT to generate marketing copy, product descriptions and more. Clienteling and customer service OpenAI’s ChatGPT tool opened up to developers in March 2023, and retailers immediately began integrating the technology for use in customer service and personal shopping capabilities on apps and websites, because the tool can be limited to a brand’s parameters and trained to improve its responses. Kering has an experimental site called KNXT, where it added a personal shopping chatbot called Madeline (using ChatGPT), which can recommend what to wear for any occasion, including suggesting products from Kering-owned brands. In a statement, the company, which owns brands including Gucci and Balenciaga, said that this was part of Kering’s innovation strategy to identify and develop potentially disruptive business models, and KNXT was specifically made to test “innovative digital experiences with the young generation”. Similarly, German e-tailer Zalando is testing a “fashion assistant”; peer-to-peer marketplace Mercari launched a shopping assistant called Merchat AI; multi-brand retailer Ssense uses ChatGPT to generate outfit ideas shared on Instagram; startup Sociate created a multimodal chatbot to help resale e-commerce shoppers find products; and payments company Klarna added a ChatGPT integration for product recommendations. And, in a hardware integration, smart eyewear maker Innovative Eyewear introduced 15 styles of smart glasses enabled with ChatGPT that enables a spoken word version of ChatGPT for wearers. Marketing and more Many are experimenting more quietly with ChatGPT behind the scenes. In addition to Shopify and Amazon’s recent business-to-business integrations, e-commerce personalisation provider Bloomreach (which counts Neiman Marcus and Benefit Cosmetics as customers), integrated OpenAI into the company's content management solution (CMS). Lingerie brand Adore Me (which was acquired by Victoria’s Secret in January) has been experimenting with generative AI for the past year, including training product and marketing teams, and starting with using the technology on SEO-optimised product descriptions. It has found that it saves up to 30 hours a month on copywriting, which allows employees to spend their time on more creative parts of their role, VP of strategy Roy says. This has led to a 40 per cent increase in website traffic. “The way we started with our copywriters is asking, ‘What do you hate doing the most?’ and they said product descriptions, SMS, email subject lines. That’s good, because when something's repetitive and structured, it can work,” Roy says. “Now, instead of 30 hours, it takes a copywriter about 30 minutes in a month to put together 120 product descriptions.” Still, he adds, there is human oversight: copywriters are served three choices and able to make changes before hitting publish. Read more: **Digital clienteling, AI in secondhand shopping, Shopify integration These news tools present unprecedented opportunity and risk. These include ethical questions (“Who controls a person’s likeness after they die?” and “Should brands disclose if part of an image has been generated using AI?”) to questions around copyright, trademark and accuracy (“Who is responsible if a machine copies trademarks?” and “What if the chatbot is wrong but believable?”). There are also concerns about AI taking jobs and spreading disinformation. “There is nothing that will happen in our lifetime that is more existential than this, and people are really grasping with what to do with this,” says ArentFox Schiff’s Lupo. The space is additionally complicated because it’s moving very quickly, he adds. “In six months, we would have a whole different conversation.” Ethics A number of conflicts have already played out. In response to Levi’s announcing that it would be adding more model diversity via AI-generated models, many wondered if this would remove job opportunities for human models. In response, Levi’s clarified its positioning, saying that it was physically impossible to photograph every individual stock-keeping unit on a wide swath of people, and that it would not decrease the number of human models it hired or its efforts at diversity and inclusion throughout the company. There are additional questions regarding representation, as many models ingest and amplify the human biases that proliferate online. The Institute of Digital Fashion’s digital wearable campaign, which was designed in part using AI, had to manually add diversity to digital models, while Asics launched an AI training programme after finding that prompts related to “healthy exercise” generated unrealistic and harmful images. More broadly, people are still worried about the impact these tools will have on jobs; 77 per cent of consumers think AI will lead to job loss, Forbes reports. Others, particularly younger generations, are more optimistic about AI’s potential for job-creation, per Atlantic Recruiters, with possible new roles including prompt engineers and empathy trainers. Mente, of Revolve, says that computer-aided design won’t replace jobs, while PhotoVogue’s Glaviano has found that the skills are transferable, with talented photographers turning into talented prompt engineers. Intellectual property There’s also the question of intellectual property. “The copyright issue is going to become central to the debate,” Adore Me’s Roy says. “How people are attributed, whose works are created, and how, has to be addressed. Creating a painting in the style of Van Gogh is fine because Van Gogh is not here, but contemporary art is definitely something that needs to be solved.” To that end, Stability AI, Midjourney and DeviantArt, which enable people to “create art in the style of” specific brands and artists, have faced lawsuits saying that they are stealing artists’ work. And, many brands are closing watching artists who generate clothing designs that heavily reference trademarks and logos; Balenciaga (in which an artist digitally dressed the Pope, using Midjourney) and Nike are often key points of inspiration. Brands still have trade dress protection, and whether something was designed by a creative agency or a computer doesn’t impact how the law is interpreted, ArentFox Schiff’s Lupo advises. However, he adds, brands might use AI to help design something that is similar without infringing. TikTok is said to be developing disclosures for AI-generated video, although this could be tricky territory if content is a blend of fictional and authentic, the same way that Photoshop is used today to alter imagery. Lupo supports the idea of a “AI watermark”, which could be used to label if a video is fictionalised. Regulations In March, Italy temporarily banned ChatGPT due to concerns over lack of transparency and data inaccuracy; this is one reason why a new clienteling chatbot created by GameOn for Valentino did not use GPT tech. The ban has since been lifted, after parent company OpenAI agreed to inform users how it’s processing data, give the option to opt out of data-sharing and put limits on usage by minors. New guidelines are in development, with Italy, the US, China and the UK exploring regulation. The US Copyright Office has advised that AI images generated from text can’t be copyrighted in the US, and the National Telecommunications and Information Administration (NTIA) has requested input on policies that shape AI accountability. The Cyberspace Administration of China (CAC) is considering requiting generative AI service providers to be responsible for the validity of data. The EU’s AI Act establishes a “risk-based” framework, while the UK’s AI White Paper hopes to “make the UK a science and technology superpower by 2030” through five guiding principles. More is necessary, say high-profile tech executives, including Tesla CEO Elon Musk, Apple co-founder Steve Wozniak and Meta president of global affairs Nick Clegg, who have called for pauses on experiments and global regulation. Tech regulation has been a long time coming, and these types of considerations and concerns are seen as a natural byproduct of innovation — even if the pace has intensified. “We had the same type of discussions in the industrial age,” Lupo says. “In my personal opinion, Pandora's box is open, and we need to figure out how to use it.” Read more: AI regulation, Model representation Additional reporting by Madeleine Schulz To receive the Vogue Business newsletter, sign up here. Comments, questions or feedback? Email us at feedback@voguebusiness.com.

Make better business decisions

Make better business decisions