Virtual try-on in fashion: how does it work and does it reduce returns?
Generative AI can now show a garment on a photo of the shopper. That helps with style decisions, but fit and size are a different problem. What the technology does, what it needs and what the evidence says.
KEY TAKEAWAYS Summary by the editors
- Virtual try-on uses computer vision or generative AI to show how a garment would look on a shopper's own photo or on an avatar.
- Google launched AI virtual try-on in the United States in July 2025 across Search, Google Shopping and Google Images, using a diffusion model trained to understand how fabrics fold and stretch on different bodies.
- Most image-based try-on tools show appearance and style, not whether a specific size will fit, which is the main driver of fashion returns.
- Zalando reported in 2023 that items carrying its size advice showed 10 percent fewer size-related returns than similar items without it, a figure for size advice rather than visual try-on.
- Try-on tools process body images, so brands need clear privacy notices, data minimisation and a measurement plan that tracks returns, not just engagement.
Virtual try-on lets online shoppers see a garment on an image of themselves or on a digital avatar before buying. Modern systems use generative AI to render how fabric drapes on a given body. They help shoppers judge style and colour, but most do not reliably answer the question that drives returns: will this size fit me?
What is virtual try-on in fashion?
Virtual try-on (VTO) is a family of technologies rather than one product. Augmented reality filters overlay items such as sunglasses, make-up or trainers on a live camera feed. Avatar-based fitting rooms build a 3D body from measurements and dress it in a 3D garment. The newest approach, image-based generative try-on, takes a photo of the shopper and a photo of the product and generates a new image of the shopper wearing it.
| Approach | How it works | Best suited to | Main limitation |
|---|---|---|---|
| AR overlay | Tracks face, feet or hands in a live camera view and overlays a 3D model | Eyewear, footwear, accessories, beauty | Poor for loose or draped garments |
| 3D avatar fitting | Builds a body model from height, weight or measurements and simulates a 3D garment | Fit visualisation where garment specs exist | Needs accurate 3D garment data per style and size |
| Generative image try-on | A diffusion model renders the product onto a photo of the shopper | Style and colour decisions at scale | Usually does not model specific sizes |
| Size recommendation | Predicts the best size from purchase and return history and garment data | Reducing size-related returns | Not visual; depends on consistent size data |
How does generative AI virtual try-on work?
Generative try-on systems are typically built on diffusion models, the same family of models used for AI image generation. The model learns from large numbers of images of people and garments how bodies are shaped and how materials fold, stretch and fall. At inference time it receives two inputs, the person and the garment, and produces a new image that keeps the person's pose and features while replacing the clothing.
Google is the most prominent large-scale example. At its developer conference in May 2025 it introduced a try-on experiment in Search Labs that asked users to upload a full-length photo and used, in TechCrunch's report of Google's description, a new diffusion model for fashion designed to understand how different materials fold and stretch on different people. On 24 July 2025 Google launched the feature more broadly in the United States across Search, Google Shopping and Google Images product results, letting users tap "try it on" for apparel listings and share the generated looks.
Google has also tested a more personal format. Its experimental Doppl app, according to TechCrunch, uses the same generative technology but adds deeper personalisation and AI-generated videos to show how outfits might look in motion. For brands, the significance is less the app itself than the fact that try-on is moving into search engines and platforms they do not control, where the quality of their own product images and data determines how their garments are rendered.
Does virtual try-on reduce returns?
This is the central commercial question, and the honest answer is that public, independent evidence is thin. Many published return-reduction figures come from technology vendors and are not independently verified. Three points are better supported:
- Size is the core problem. Academic work on virtual try-on notes that most existing tools follow a standardised approach that limits interaction with size and fit and does not give actionable sizing guidance. A 2024 research prototype, SiCo, let users visualise different sizes on their own image, and participants in its user study reported more confidence in size selection.
- Size advice has measurable effects. When Zalando piloted a 3D avatar fitting room for jeans across 25 markets in 2023, it reported that items with its size advice showed a 10 percent decrease in size-related returns compared with similar items without guidance. That is evidence for size recommendation, not for visual try-on.
- Style confidence may help. Seeing a colour or silhouette on one's own body may reduce "looks different on me" returns, but brands should test this on their own data.
What data and integration does virtual try-on need?
- Clean product images: front-facing, well-lit images on neutral backgrounds give generative models the best input.
- Garment measurements by size: essential for any approach that claims to show fit rather than appearance.
- Category scope: generative try-on works best for tops, dresses, trousers and skirts; Google's initial experiment covered shirts, pants, skirts and dresses.
- Performance and mobile UX: generation time and camera or upload flows strongly affect adoption.
- Return reason data: needed to evaluate whether the tool changes the reasons customers give for returning items.
What are the privacy and accuracy risks?
Try-on tools process full-body photos, and some build body measurements. Brands should treat these as sensitive personal data: explain clearly what is stored and for how long, delete images when they are no longer needed, and avoid reusing them for other purposes without a lawful basis. If a tool identifies individuals from facial features, stricter biometric data rules may apply.
Accuracy is the second risk. A generated image can make a garment look more flattering, or fit differently, than the real product would. If shoppers rely on an over-optimistic rendering, try-on can increase rather than reduce disappointment. Clear labelling that the image is AI-generated and an approximation helps manage expectations.
Should a fashion brand invest in virtual try-on?
For most brands, the first priority is accurate size advice based on garment measurements and returns data, because that addresses the main cause of returns. Visual try-on is increasingly available through large platforms and search engines, so brands should first make sure their product images and data are good enough to be rendered well there. Building an in-house try-on experience makes sense for brands with high return rates in visually driven categories, sufficient traffic to test properly and the product data to support it.
Frequently asked questions
How does AI virtual try-on work?
Generative try-on uses a diffusion model that has learned how bodies and fabrics look. It takes a photo of the shopper and an image of the garment and generates a new image of the shopper wearing it. Other approaches use AR overlays or 3D avatars.
Does virtual try-on reduce returns in fashion?
Independent evidence is limited and many published figures come from vendors. Most visual try-on tools show style rather than fit, while size is a main cause of returns. Zalando reported a 10 percent drop in size-related returns for items with size advice, which is a different tool.
Is Google's virtual try-on available in Europe?
Google launched its AI try-on feature in the United States in July 2025 across Search, Shopping and Images. Brands should check Google's current documentation for availability in other markets, as rollout status changes over time.
What do brands need for virtual try-on?
High-quality product images, garment measurements by size if fit is to be shown, clear privacy notices for body photos, and a test design that compares return rates with a holdout group.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
