With her wedding three weeks away, a shopper pinches her screen to zoom in on a kurta. The model looks to be a size 38; she is a 46. However beautifully lit, the photograph tells her little about how the fabric will fall across a fuller stomach. So she does what larger shoppers have long learnt to do: skip the images and scour the reviews for someone with a body like hers.
Virtual Try On in ChatGPT: a red silk saree with a gold zari border, a selfie, then the generated try-on
| Photo Credit:
OpenAI
Retail has spent more than 15 years trying to solve this problem with technology. In the early 2010s, UK-Estonian startup Fits.me developed robotic mannequins that reshaped themselves to match a shopper’s measurements. The technology attracted attention, but its high cost raised questions about scalability. Rakuten acquired the company in 2015 before eventually shutting it down. London-based Metail took a different approach, using shoppers’ measurements to generate 3D body models.
India had its own experiment. In 2013, Myntra acquired Fitiquette, whose virtual mannequin could be adjusted to resemble a shopper, measurement by measurement. Three years later, Myntra was promoting Try and Buy, which let customers try apparel, shoes and accessories at home before paying. The most ambitious attempt to solve online fit had, in effect, delivered a familiar solution: let the customer check the mirror before committing.
When Walmart acquired virtual fitting-room company Zeekit in 2021, its apparel head, Denise Incandela, called understanding fit “one of the most difficult things to replicate online”. A Business of Fashion analysis the following year found that virtual fitting rooms had struggled to gain traction, hampered by high costs, low adoption, the difficulty of rendering garments and the inherently subjective nature of fit.
Lenskart, meanwhile, made virtual try-on work for spectacles, using a phone camera to place frames on a customer’s face. Frames are rigid. Fabric drapes, clings, puckers at the seams and moves when you do.
A picture, not a mannequin
Generative AI has changed the approach. Instead of building a body model, newer systems generate an image of the clothing on a person. In June 2023, Google introduced a diffusion-based model that could show garments on models ranging from XXS to 4XL. By December 2025, shoppers in India could virtually try on listed products using their own photographs. On October 1, OpenAI brought the idea into ChatGPT with the global launch of Virtual Try On. A Try On button appears alongside supported clothing and accessories in shopping results; shoppers can also upload a screenshot of an item they have spotted elsewhere and ask to see it on themselves. The images are generated using ChatGPT Images 2.5, OpenAI’s recently launched model.
Earlier systems struggled to look convincing. The new ones can produce remarkably realistic images, shifting the challenge to the shopper: how much can a beautiful picture actually tell you about a garment? That is the question I put to Aarush Selvan, product manager at OpenAI, in an exclusive interview for India.
Aarush describes the ambition modestly. “We want ChatGPT to be a useful place to start when you’re deciding what to buy,” he says. The try-on feature is designed to fit into a broader shopping conversation, allowing users to explore ideas, discover products and save the ones they like.
He is equally direct about its limitations. “These images may not reproduce the person or garment exactly. They should be treated as a visual reference, not a guarantee of fit or size.” OpenAI advises shoppers to check the merchant’s measurements, product details and return policy before buying.
What the image knows
For a larger shopper, that caveat begins with the body in the picture. Asked how OpenAI checks whether a try-on faithfully represents a person’s shape, proportions, skin tone and features, and guards against producing a more flattering version of them, Aarush points to testing. “Our evaluations include people of different races, skin tones, genders, and religions,” he says. The feature also uses ChatGPT Images’ existing safeguards, though he reiterates that the result may not reproduce a person exactly.
A wedding collection listing for a silk shirt with a veshti, then the same outfit generated on the shopper
| Photo Credit:
OpenAI
The garment brings its own uncertainties. In India, wedding shopping rarely begins on a shopping site. It starts in the family WhatsApp group, weeks before the wedding. A cousin shares a screenshot from a designer’s Instagram; an aunt forwards a photograph of a sari worn to another reception; a sister-in-law responds with three alternatives from different websites. By the time anyone is ready to buy, the inspiration is a camera roll full of screenshots.
ChatGPT can work directly from these images. “You don’t need to start with a product listing,” Aarush says. But where a screenshot leaves out details, the model may fill them in. “Where details are missing, the image model fills them in when it is confident in their accuracy.” These additions, he cautions, should not be treated as verified information about parts of a garment the image does not show.
Dressing for the muhurtham
Anyone who has dressed for a South Indian wedding knows how quickly one purchase becomes several. Choose a Kanchipuram sari, and its border dictates the blouse, which determines the temple jewellery, until the entire look depends on a strip of zari nobody has seen in daylight. Grooms fare no better: the angavastram must complement the border of the veshti.
A request for South Indian wedding clothing returns silk shirts paired with gold-bordered veshtis
| Photo Credit:
OpenAI
Aarush is clear about what the technology can do. Shoppers can upload an image of a wedding outfit and ask ChatGPT to find similar styles or more affordable alternatives. But the try-on feature handles only one item at a time. Visualising an entire ensemble requires a workaround. “Someone can upload images of multiple pieces and ask ChatGPT to generate an image of them together, and that could work, but it is not a supported multi-product try-on feature currently,” he says.
For looks discovered on social media, Aarush draws another boundary. “We don’t promote counterfeits or trademark-infringing replicas.” Recommendations depend on the shopper’s request, the context provided and the product information available. With Memory enabled, ChatGPT can also draw on preferences from earlier conversations, although there is no separate setting for shopping preferences.
“Product results are separate from ads. Advertisers can’t shape, rank or alter ChatGPT’s answers,” he says, adding that commercial partnerships do not influence recommendations. Results are not exhaustive, however. When suitable products are unavailable, ChatGPT may suggest styles, fabrics or colours worth exploring instead.
Prices and product details come from third-party providers, meaning updates to stock, shipping terms and prices may take time to appear. Purchases are completed on the merchant’s website, where the final price, availability, sizing and delivery details must be confirmed.
A Tamil wedding can stretch from the nalangu to the muhurtham and reception, each demanding a different outfit, often purchased weeks apart. ChatGPT’s Favourites feature lets shoppers save products for later, while folders in the Library can organise options by occasion. Aarush suggests creating a separate folder for each event. A saree saved in September need not remain frozen in time. “When you open ‘Library’, we refresh the product information when you click on a saved product,” he explains.
Setting up Virtual Try On: ChatGPT asks for a selfie, then a full-body photo, kept as reference photos for future try-ons
| Photo Credit:
OpenAI
The reference photograph used for try-ons is retained for future sessions, sparing shoppers repeated uploads. It can be replaced or deleted under Settings, Personalisation and Reference photos. Generated try-on images are saved to the Library and remain private unless the user chooses to share them.
Somewhere on Instagram is an outfit someone wants for a Diwali card party, worn by a stranger who makes it look effortless. Until recently, finding something similar meant wrestling with a third kurta in a cramped trial room, only for the mirror to confirm that the colour was never going to work. Now that discovery can happen on the sofa, using a single photograph. With Diwali celebrations giving way to wedding season, a feature launched only on October 1 offers a glimpse of where festive shopping might be headed. The technology may have changed the trial room, but it has yet to settle the oldest question in Indian wedding shopping: whether the outfit will survive the family WhatsApp group’s approval.