Background Context
In 2019, Maísa Benatti was a size-and-fit product manager at Farfetch, and her job was to help shoppers find the right size. Months into the role, after sitting through research sessions and reading through feedback forms and support tickets, she concluded she had been solving the wrong problem. Size was not the barrier. Confidence was.
The phrase she kept hearing, in one form or another, was always some version of: “I don’t look like the model.” Not a measurement issue and not even a case of checking the size chart. Shoppers simply could not picture themselves in the garment on screen, and no amount of numerical detail was going to close that gap.
Sizing and fit problems are still the main cause of online clothing returns, making up about 70% of all returns as reported by Coresight Research and Alvanon. The combined cost of online returns across retail translates to roughly $600 billion annually.
Benatti came to see it as a representation problem rather than a sizing one, and by her account, it was the most fundamental unsolved problem in online fashion. This realisation set the direction for the next seven years of her career.
The Long Road to Good Enough
Virtual try-on (VTO) technology, which renders a garment on an image of a shopper’s own body before purchase, is the industry’s most direct response to the representation problem identified. It has taken far longer than expected to become dependable, but the underlying data on returns, conversion and market growth now suggests it has reached that point.
Looking back to 2019/2020, Farfetch moved quickly to launch one of the earliest virtual fitting-room experiences. But VTO technology was not ready. And the shopping experience showed why. During the early years of this new technology, most people wouldn’t upload a photo of themselves. It felt too personal, too exposing. But even for those who did, the experience demanded a lot. The person had to wear tight clothing so the algorithm could accurately read their body. Then they had to wait. Sometimes twenty-four hours or more!
VTO was technically impressive and practically useless. The fashion industry had identified the right problem, but, they built the wrong solution entirely. Adoption was low. Business impact was limited. VTO arrived too early, with tools that weren’t ready.
For the next four years, shoppers kept asking the same question: “How will this actually look on me?” Accurate representation remained one of the biggest barriers to buying fashion online. For much of that time, the honest answer was that the solutions available just weren’t solving the problem well enough. VTO seemed like it would never go mainstream. The technical problems felt too hard, the consumer behaviour felt too fragile, and the catalogue scale felt too enormous.
When Habit Replaces Novelty
Skip ahead to 2024-2026, and AI infrastructure built for fashion content production and shopper experience finally got to the point where VTO felt close to delivering what shoppers had actually been asking for since 2019. The problem had not changed, but the technology had finally become good enough to help solve it.
Most retail technology only succeeds when it solves a real business problem and becomes something consumers actually use. The most valuable retail technologies also create behaviours that consumers want to repeat. For most of its history, VTO solved a business problem and created a novelty. Consumers tried it once and didn’t come back.
What we’re seeing now is different. Consumers are returning to the experience multiple times in a session. They are trying more products. They are sharing results. They’re incorporating VTO into their shopping experience rather than treating it as a one-and-done feature to tap and forget.
We have crossed the line from novelty to habit. And that changes everything about what VTO needs to be.
The moment consumer technology becomes habitual, it stops being just a feature. It becomes something people rely on. This is where the technology needs a higher level of reliability and performance.
Why Getting Enterprise Virtual Try-On Right Is Harder Than It Looks
It’s easy to make VTO look impressive in a demo. The challenge is making it work consistently in the real world.
One of the biggest challenges is how the model represents body measurements. A 74cm waist means something different on a 5’2″ frame than on a 5’10” frame, yet to a neural network, “waist = 74” is simply a number. Without the right data representation, AI models struggle to understand body similarity and fit.
Solutions that work at scale go beyond raw measurements, using learned embeddings to capture relationships between proportions, Fourier encoding to model nonlinear changes between body shapes, and multimodal inputs that combine measurements, images, pose estimation and real behavioural signals such as returns and confirmed perfect fits.
Body preservation is equally critical, since the shopper’s body must remain a stable reference point while garments render around it. When that breaks down, even subtly, the experience can feel wrong enough to make shoppers leave.
Making one garment look convincing on a stock model is one thing. Doing it accurately and consistently across hundreds of thousands of SKUs, different body types, and near-instant load times is a completely different challenge. At that scale, infrastructure becomes the real requirement.
From Single Garments to Complete Outfits
Early virtual try-on tools rendered one garment at a time, limited by what the technology could support rather than by how people actually shop.
When people shop, they’re usually thinking about the whole look. For example, picturing a jacket against trousers, shoes and the rest of an outfit. A buying decision is typically a decision about a full look, even when only one item ends up in the cart. A physical fitting room reflects this reality. Online shopping has struggled to replicate it, and single-item try-on, while an improvement, remained a fragment of the fuller experience.
As recently as a year ago, layering multiple garments in a single render was unreliable, with bags and shoes frequently breaking the image. Advances in generative modelling have since made it possible to render complete outfits, while preserving both the garments and the shopper’s likeness.
This is a true structural improvement.
From “Will This Suit Me?” to “I Didn’t Know I Would Love That”
Traditional cross-sell recommendations suggest products based on browsing and purchase patterns. Outfit-level try-on lets shoppers see how those suggestions work together on them. Someone considering a jacket can explore it with different trousers or shoes, making it easier to judge the combination before adding another item to their cart.
The larger opportunity is discovery, not immediate upsell.
Single-item try-ons mostly answer a narrow question, “Will this suit me?” Outfit-level try-on can produce a different response, closer to surfacing a combination a shopper would not have assembled independently. That shift, from reducing uncertainty to expanding possibility, is where the biggest opportunity in fashion AI sits, and it’s all about helping shoppers find styles they would not have found on their own, plus building an experience that inspires a decision rather than only confirming one already made.
What Enterprise Retailers Should Be Building Toward Now
It is time for retailers to move beyond asking whether VTO can produce a convincing image. It can. Now, they need to know if it works consistently across their catalogue, if shoppers return to it, and if it improves commercial outcomes. Then, the questions become: “What does the shopper experience look like when outfit visualisation is the default, not the feature?” “What does content strategy look like when every garment needs to be try-on-ready in every colour and size?” “What does the PDP look like when discovery is part of the experience, not just confirmation?”
Some of the hardest fashion AI customer problems remain unsolved. And many of us on the tech side are working on them. But for retailers, the practical step is to use pilots to test what a wider rollout will require: catalogue integration, consistent image quality, predictable response times and responsible handling of shopper photos. Those foundations matter as much as the initial results.
The Part That Was Always Missing
The question I heard at Farfetch in 2019 is still the right one: “How will this look on me?” Retailers now have better tools to answer it. The next task is making that answer reliable enough for shoppers to use in everyday buying decisions.
Done well, virtual try-on can help shoppers explore combinations they might never have considered—and feel more confident choosing something that feels like them.
