To become a Vogue Business Member and receive the Technology Edit newsletter, click here.
Google has introduced a virtual try-on tool using generative AI, with the goal of making it easier for people to find the right fit as they’re shopping online.
In Google search, the new tool will appear in results to give people the option to see how shoppable items look on a range of models. Google photographed 40 men and 40 women to capture a range of sizes (from XXS to 4XL), body shapes, ethnicities, hair types and skin tones. Shoppers can scroll through and select images of various models they want to see; the item is then digitally and photorealistically portrayed on the model’s still image.
The first roll-out will only include women’s tops, and is available to any retailer set up in Google’s Merchant Center. It’s available to global brands but only viewable to US consumers to begin. Brands don’t need to do anything to onboard their inventory to be featured in the tool.
The sizes are not standardised, meaning that the label of the model’s size does not necessarily correlate to the brand’s alphabetical or numeric size. Since the fitting process is not informed by text prompts, it doesn’t read material makeup to interpret drape or sheen, but rather bases it all off combining the original brand-supplied, on-model photography with the model images taken by Google.
Lilian Rincon, Google’s senior director of consumer shopping product, says that this is a very early example at how this technology could work. “We feel a lot of responsibility about being thoughtful. There are so many angles of diversity, and you can imagine in the future that these might increase.” She adds that the features and quality will improve over time. Future considerations also include the option to add each model’s height, and plans to expand into additional product categories.
Google also published a research paper today about the project, written by the University of Washington researchers who produced the technology. They noted that a key challenge was the ability to accurately recreate the garment on different model poses. In objective user studies conducted to evaluate “TryOnDiffusion” (as they named the new tech), raters selected it as the best for 93 per cent of the options. A second study, which added more challenging model poses, the raters selected the new method as best of 96 per cent of the inputs.
In the past year, Google consulted 18 brands, ranging from mid-sized direct-to-consumer to global brands, to help inform this feature; it found that a highly accurate experience and featuring human models were two of the most important points. Google has also added image recognition to help people find additional similar products.The problem of online fit is as old as online shopping, and is often to blame for high returns, which have plagued retailers. US consumers were estimated to return more than 20 per cent of all online purchases in 2022, according to the National Retail Federation. In 2020, returned inventory created an estimated 5.8 billion pounds of landfill waste and return shipping emitted 16 million metric tons of carbon dioxide, Optoro reports.
There’s also the consumer side: 42 per cent of online shoppers don’t feel represented by the images of models they see in e-commerce and 59 per cent feel dissatisfied with an item they shopped for online because it looked different on them than expected, according to a 2023 Google and Ipsos survey of US online shoppers.
Tech startups have been partnering with brands for more than a decade to attempt to solve these challenges. They have often turned to AI, which can generate realistic-looking images based on text- or image-based prompts. As early as 2017, Amazon acquired Body Labs, which makes 3D models for virtual try-on. In 2019, Meta developed tech that uses AI to analyse outfit images and make adjustments to make it more stylish. Zeekit, which also uses photographs of human models to show online fit, was acquired by Walmart in 2021. Last year, Farfetch acquired augmented reality try-on company Wannaby (which just launched a pilot with Maison Valentino ready-to-wear). Bods, which has worked with brands including Khaite, enables shoppers to build and customise avatars in their own image, then see how digital twins of the clothes fit on that avatar.
Rincon says that Google’s approach uses generative AI in a way that hasn’t been done before, and is not another version of digitally “copy and pasting” assets on to model photos using geometric warping or other tech. “This takes an image and an image and uses a diffusion model to put the asset onto our models. It’s not Photoshop.”
Challenges aren’t just technical
AI-generated content regarding humans is still seen as controversial, and complicated technically. Rincon says that Google is especially conscious to note that the models being digitally dressed are real people that Google hired and photographed, and is considering adding their first names to the user interface to further emphasise that point.
This stands to be viewed as more “ethical” as it is digitally dressed on images of existing people. A recent project by Levi’s had a similar aim to show items fitted on a more diverse range of people, working with tech startup Lalaland.ai that generates fictional people using AI. Critics responded that it was unethical to depict fictional people in place of live human models; Levi’s responded that it would be physically impossible to photograph every individual stock-keeping unit on someone of each size and shape. Vue.ai, another tech startup, has also provided this type of technology, and recently partnered with Meta to provide it to Meta’s advertisers.
Google’s approach makes this tech available to brands who don’t have the resources to hire and photograph so many models, instead encouraging retailers to upload their inventories to Google’s Merchant Center. It’s part of Google’s efforts to apply generative AI to help it become a native fashion and beauty shopping destination as it competes with Meta and Amazon for advertising revenue. Google’s 28.4 per cent share of the global digital ad market leads Meta (at 20.1 per cent) and Amazon, at 7.5 per cent, Insider Intelligence reports.
In recent months, Google has also used generative AI to add the ability for advertisers to tweak advertising imagery to make it more compelling, and introduced a Search Generative Experience (SGE) that responds to product searches with additional details such as top product considerations and product summaries with key features, pricing and stockists. Last year, it added the ability for people to visualise lipstick and foundation on a range of models. These all follow the launch in 2021 of what it calls its “Shopping Graph”, which uses AI to surface recommendations from the billions of free-to-list products, and replaces its standalone shopping app.
In an ideal world, people would be able to see how the clothing looks on their own image, and then try on various sizes to recreate a true fitting room experience. Rincon says that more is on the roadmap, and retailers have already tasked Google with helping to standardise sizing, which is a natural complementary challenge when it comes to predicting fit across multiple brands — something that multi-brand fashion-tech companies such as Rent the Runway and Stitch Fix have spent years working to solve.
Other current technical challenges reported by the researchers include portions of one image “leaking” onto the final image, difficulty preserving tattoos and muscle structure and expanding into images with complex backgrounds.
For now, even expanding beyond women’s tops stands to be a sufficient hurdle. “This sounds really simple, but it’s such a complex thing to bring into the world. Just for jeans, we have to totally rethink this,” Rincon says.
Comments, questions or feedback? Email us at feedback@voguebusiness.com.
Generative AI is coming to Google search
Virtual try-on is being hit by class actions. Should brands worry?