AI models are here. Can they actually improve fashion representation?

To receive the Vogue Business newsletter, sign up here. When Levi’s announced plans to experiment with AI-generated models to increase diversity on its e-commerce channels, it received swift backlash. Creating an avatar of a person of colour — rather than hiring a real person — seemed like a way to profit from the appearance of diversity without having to invest in it. It also raised a debate over whether AI could be used to improve representation in fashion in a way that benefits marginalised groups. Michael Musandu, the CEO and founder of digital model studio Lalaland.ai, which partnered with Levi’s on the AI models, says his goal is not to replace traditional photography and physical models, but rather supplement them by allowing brands to be more time-efficient and offer a personalised shopping experience, and that inclusion must play a role in the process. “We believe human models will continue to play a vital role in the fashion industry, establishing genuine connections with consumers; our technology aims to support this. And, yes — we need more of them to come from underrepresented groups if fashion companies are serious about inclusion efforts,” he says. Part of the backlash to the Levi’s experiment stemmed from the discomfort surrounding the use of AI to replace humans. Brands and AI companies need to tread carefully, even when the intentions of diverse AI model representation is ostensibly to allow customers to see what they look like in clothing. The use of AI models raises questions around who is profiting, what representation is actually “real” versus manufactured and the intention of those powering the technology. Through a different lens, AI could open up opportunities for brands to fight biases if they’re on board to help provide data to programme algorithms with inclusivity from the start, and for models (particularly from underrepresented backgrounds) to be paid for their digital likeness. “Empowering diverse models by giving them the means and protection to capitalise on their digital twins through licensing agreements is essential,” says Dale Noelle, founder of True Model Management, which has a 3D body scanner at its New York headquarters which it uses to create digital versions of its models. Musandu sees his technology as a way to create a new stream of income for BIPOC communities. “As a boy growing up in Zimbabwe, I rarely saw anyone who looked like me featured in photos, advertisements or the runway. This deeply impacted me and was a motivating force in my desire to utilise technology — ethically and responsibly — to drive real and sustainable change in representation.” Musandu says he welcomes critical conversations about AI as it will only make the end product better and more considerate. (Levi’s, in response to the backlash, said in a statement, “We do not see this pilot as a means to advance diversity or as a substitute for the real action that must be taken to deliver on our diversity, equity and inclusion goals and it should not have been portrayed as such.”) There are many ways that brands can employ digital models or influencers to show their clothing using AI, and many services that provide technology to support this. The problem, historically, has been when these projects are associated with diversity. Digital representation has often been criticised as a form of appropriation: controversy has surrounded dark-skinned digital model Shudu Gram because she was created by a white man, and questions around true representation of women of colour have surrounded digital influencer Lil Miquela, who is designed to be mixed race Brazilian-Spanish. Some say that the best approach is to base AI models on physical humans from the start, especially if the intention is to allow consumers to see themselves represented and understand what an item of clothing would look like on them. Lalaland’s AI models are 100 per cent computer generated, but Musandu says he is able to achieve the hyperrealistic look because he licences images from a variety of communities (particularly those that are underrepresented) to bulk up his images database. Christine Marzano, founder of fashion-tech platform Bods, agrees that the biggest challenge for AI representation is a lack of data on certain demographics, and less data means the AI is going to be less accurate. Like Lalaland, Bods had to create its own image database. Bods uses AI and machine learning in a different way, to pick out a consumer’s measurements from photos submitted, creating a 3D avatar that fits the consumer’s exact size specs, with options to customise musculature or curviness, positioning of breasts and skin tone. Fashion is at a pivotal point, experts say, to train AI systems to improve representation. That doesn’t mean replacing human models with generated avatars — experts agree that the most equitable forms of AI will include a diverse team behind the scenes and involve real models from underrepresented backgrounds who are paid fairly. “We pay models for photoshoots like they would a traditional photoshoot. We brief them on how their images will be used prior so they’re well-informed, which we find works well,” says Brian Harris, director of R&D, new initiatives and digital transformation at Vue.ai, which provided AI models for Meta’s Creative Shop. Brands have a duty to be transparent when an image of a human is generated by AI. When experimenting with AI-created models, they should ensure they have a robust DE&I strategy already in place, so that their diverse AI models are a continuation of existing efforts. “If there are brands that are already well known for using diversified people across their campaigns and behind the camera, then AI is going to be an extension of what they already do,” says Shereen Daniels, managing director of racial equity advisory firm HR Rewired. “If you're a brand that already has challenges around representation and racial equity, tacking [AI] on is always going to look problematic, like you’re trying to take a shortcut to ‘look better’.” In its initial statement, Levi’s said it has been working towards improving DE&I over the past year, and that AI model diversity is not its sole solution. “Over the past year, we’ve been focused on ensuring that not only is our work diverse but those working on the content both in front of and behind the camera are reflective of our broad consumer base, and we’re continuing to do just that,” the statement says. Brands that want to succeed with equitable AI innovations need to ensure their DE&I team is not siloed from the decision-making process and are involved in early discussions. Daniels encourages brands to make decisions about new innovations, products or ways to market through an “equity lens” to examine the intended or unintended impact on marginalised communities — especially given fashion’s history of profiting from exploitation and appropriation. Tommy Hilfiger has developed a department to bring together its innovation, DE&I and sustainability teams to “bring these teams and expertise together for increased collaboration and knowledge sharing”, according to a representative. The brand has not yet explored AI-generated models, but it awarded Lalaland.ai as one of the winners for the Tommy Hilfiger Fashion Frontier Challenge in 2022. “While Tommy Hilfiger does not partner with Lalaland.ai to create AI models, we recognise that this is an important technological evolution for which the fashion industry has a responsibility to understand the full set of opportunities, diverse perspectives and implications.” Follow the money Who is profiting? The biggest red flag, experts say, is when brands are actively avoiding paying people from underrepresented groups, whether that’s models, photographers, developers or even people whose likeness has been used to develop a certain dataset. For now, human models are still needed — both as a reference to create accurate photorealistic AI images and to capture the “human magic” that fashion loves. AI companies are finding ways to weave some of the individuality and personality lost in fictional, digital models by working with real models. “Brands that are trying to stay true to inclusivity and diversity are going to want to feature that magic they see in real model imagery, so I think you’ll see them leaning towards the side of more of a digital twin, where a model can be digitised and they can update that imagery and be paid every time,” says Harris. However, diversity in technology for digital twins still needs to catch up. British talent incubator On|Off hosted a phygital fashion show during London Fashion Week in February which featured physical models alongside avatar counterparts wearing digital twins of the garments. The physical models were mostly gender non-conforming models, however, the avatar versions of the models were gendered. “When we did the digital fashion show with the avatars next to the real life models, I realised the models were probably more diverse than their avatar equivalents,” says founder Lee Lapthorne. “I don’t feel that the AI programmes have evolved so much to have proper representation of individuality. We were limited by the mechanics of the avatar framework.” On the design and photography side, Revolve co-founder Michael Mente sees generative AI as an opportunity to democratise fashion because the tools lower the barrier to entry for aspiring designers. Revolve partnered with Maison Meta at AI Fashion Week to seek out new talent working with AI, supporting one designer by producing their collection physically and selling their products. “With proper education, people around the world regardless of race or income can participate on an equal playing field,” says Mente. However, some experts caution the true cost of free open generative AI tools: though they may be more democratic in terms of access, the image databases are scraped from all over the internet so the data is not given voluntarily. “Legislation and guidelines around AI must evolve as quickly as the technology to remain equitable,” says Noelle. The problem is these paid programmes are inaccessible to the marginalised groups. Fashion Institute of Technology assistant professor Jennifer Lee says that less than 10 per cent of her 3D garment visualisation course is made up of minority students due to funding issues, which is why she volunteered to teach the programme for free in 2021 to a cohort of minority students. Experts say funding from institutions and brands will be needed to give marginalised people a seat at the AI table — which brands would benefit from, as diversity behind the scenes could minimise some of the missteps in representation. Training AI for better representation Many inclusive brands are excited about the prospect of being involved in a solution to override biases that have been built into algorithms. “If you looked at millions of images across the internet you’d never know that plus-size women are the invisible majority,” says Nadia Boujarwah, CEO of plus-size speciality retailer Dia & Co. “The worst danger would be for AI to pick up on all the biases that are so deeply entrenched in what we do, but if we consciously reverse those biases systematically, that’s the most exciting outcome of all.” Nonetheless, the brand is taking a cautious approach given the controversy and has not yet experimented with image-based AI. Lingerie brand Adore Me found that the image-based generative AI it experimented with isn’t advanced enough to accurately represent the fit of items such as bras and corsets, but the brand has created its own large language AI models for text-based tasks, which VP of strategy Ranjan Roy says was a fairly simple process. Roy thinks brands that want to ensure their image-based AI is diverse enough should look into building their own customised models and datasets, because AI ingests, and can magnify, the same biases that exist in human-created imagery on the internet. “The datasets that are going to drive business for the next five to 10 years are being built now, and the training sets we’re using now are biassed or built by a very specific population. It’s really exciting and important that brands can train image models with diverse models rather than using AI to replace diverse models.” Brands could also use generative AI as a way to create test campaigns and gather feedback from consumers and employees on the campaign concepts, says Fashion Institute of Technology assistant professor Ramona Dunlap. They can then take those insights and feedback and work with agencies and photographers to find physical people to create an optimised version of the campaign that truly speaks to people from various underrepresented communities. With the acceleration of AI activations, brands that want to be successful will need to have an intimate understanding of what their audience truly wants and values. “Oftentimes at companies, we might sit at a table and think we know what’s best and we don’t put it in the hands of consumers — and that’s when you get backlash,” says Dunlap. “That leads to the company having to spend the money twice because they’ve gone out there and invested in something that their audience doesn’t want at all.” Correction: This article was updated to reflect Maison Meta's involvement in AI Fashion Week and the AI Fashion Incubator, for which Revolve is a partner. (25/4) Comments, questions or feedback? Email us at feedback@voguebusiness.com. More from this author: Ferragamo sales decline 4% as it navigates ‘transition period’ Gucci among fashion firms targeted in EU antitrust raids Fashion can now report its plastic use. Will reduction follow?

Make better business decisions

Make better business decisions