Spequlo
Try On

Blog

What Is Virtual Try-On? A 2026 Guide for Fashion Brands

By Divine Eyo · · 8 min read

Virtual try-on is technology that shows a shopper what a garment looks like on their own body, before they buy it, using a photo of the shopper instead of a studio model. In fashion it takes two technically distinct forms: real-time augmented reality, which works well for eyewear and makeup and poorly for clothing, and AI-generated imagery, which became commercially viable in 2023 and is what most apparel brands mean when they say "virtual try-on" today.

This guide covers how it works, what it does and does not solve, who has actually deployed it, and how to tell whether it is doing anything for your store.

Contents

How does virtual try-on work?

Modern apparel try-on is an image-generation problem. A model takes two inputs — a photo of the garment and a photo of the person — and produces a single new image of that garment on that body.

The reference architecture is TryOnDiffusion: A Tale of Two UNets, published by Google Research and the University of Washington at CVPR 2023. It uses a parallel-UNet design that warps the garment and blends it onto the body inside one network, specifically to survive significant changes in body pose and shape. Before that paper, garment try-on was largely a 3D-modelling exercise: build a digital twin of the garment, build an avatar, simulate the cloth. That approach produced accurate physics and expensive, slow pipelines that never reached most brands.

Generative try-on collapsed the cost. It works from flat product photography — the images a brand already has — with no 3D asset creation and no reshoot.

What's the difference between AR try-on and AI try-on?

They are different technologies aimed at different products, and the distinction matters when you're evaluating vendors.

Augmented reality try-on tracks a live camera feed and overlays a 3D asset in real time. Google's ARCore Augmented Faces API, for instance, provides a 468-point dense facial mesh and names glasses and cosmetics as its use cases. AR is excellent for rigid or surface-conforming items: eyewear, makeup, jewelry, watches, footwear. It runs live, on-device, with no image generation.

AR is bad at clothing, and the reason is structural. Overlaying a garment mesh cannot model drape, cling, stretch, or how a fabric folds on one specific body. A dress is not a rigid object sitting on a surface.

Generative try-on produces a still image rather than a live feed. It handles drape, folds, shadow and body-shape variation convincingly. Its weaknesses are latency and inference cost, fine detail fidelity — logos, text and busy prints are where artifacts show up — and the fact that it is a visualization, not a measurement.

A useful shorthand: if the product sits on a surface, AR. If the product hangs from a body, generative.

→ See what generative try-on looks like on a real product page

What problem is it actually solving?

Returns, and the buying hesitation that precedes them.

The National Retail Federation, with Happy Returns, projected that U.S. consumers would return $849.9 billion of merchandise in 2025 — 15.8% of total retail sales and 19.3% of online sales. One caveat worth knowing, because it is widely misquoted: the NRF does not publish an apparel-specific return rate. Any article citing "NRF says apparel returns are X%" has misattributed the general online figure.

For apparel specifically, the most credible public numbers come from two places. Zalando states that European online fashion return rates "can reach around 50%," rising to 65% for jeans, and attributes "up to half" of fashion returns to size and fit. That is a first-party retailer disclosure without published methodology, so treat the range as directional.

PowerReviews, surveying 25,452 U.S. consumers in 2023, found that 39% of shoppers who returned apparel did so because they did not like the way it fit — ahead of appearance mismatch (28%), deliberate multi-size ordering (13%) and inaccurate descriptions (13%).

One number to avoid: the "70% of apparel returns are caused by fit" figure that circulates widely in this category traces back to 3DLook, a body-scanning vendor — supplier marketing for the exact problem it sells against. Several sizing-tool blogs then cite each other rather than any primary source. If a vendor quotes it at you, ask where it came from.

Does virtual try-on show fit, or just look?

Look. This is the single most important thing to understand before buying one, and the place where the category oversells itself.

A generative try-on image tells a shopper how a garment reads on their body: silhouette, proportion, colour against their skin, whether the neckline sits where they hoped. It does not tell them whether the size 8 will be tight across the shoulders, because the model is generating an image, not measuring a body.

Fit guidance is a separate capability. It comes from comparing a shopper's measurements against the brand's own size chart, and any vendor conflating the two is describing something their software does not do. The honest framing is that try-on answers "does this suit me" and size guidance answers "which one do I order" — and a brand losing money on returns usually needs both.

Who has actually deployed it?

The pattern worth noting: Amazon and Walmart moved first with AR and avatar approaches, and the step-change came when Google shipped generative try-on in June 2023 — the same week the TryOnDiffusion paper was published, and days before it was presented at CVPR. ASOS's 2026 launch is the clearest signal of where the category landed: shoppers get a choice between their own photo and a matched model, because not everyone wants to upload themselves.

Is virtual try-on safe? What happens to the photo?

There is a real litigation record here, and brands should know it before deploying anything.

A series of class actions under Illinois' Biometric Information Privacy Act (740 ILCS 14) have targeted try-on tools since 2022, including suits against Estée Lauder, Louis Vuitton, Christian Dior, Pandora and Wella. Outcomes have split. Castelaz v. Estée Lauder was dismissed in January 2024 because plaintiffs failed to allege the facial scans made them identifiable. Warmack v. Christian Dior was dismissed in February 2023 under BIPA's health-care exemption, on the reasoning that non-prescription sunglasses are Class I medical devices. But Kukovec v. Estée Lauder survived dismissal, as did the notice-and-consent claims in Theriot v. Louis Vuitton.

Read that record carefully: the defense wins turned on statutory carve-outs and identifiability requirements, not on any finding that try-on is privacy-safe. The exposure attaches most strongly to tools that scan a live face or body. A tool that conditions on a still photo the shopper chose to upload has a different profile — not a zero one.

Practical implication: ask any vendor where photos are stored, how long they are retained, whether a shopper can delete theirs, and whether the images are used for model training.

How do you tell whether it's working?

Most virtual try-on results you'll be shown are before-and-after comparisons: conversion in the month before install versus the month after. That comparison cannot separate the tool from the season, the promotion calendar, a paid-media change or the product mix.

The method that survives scrutiny is a random session holdout — some sessions get the fitting room, some don't, assignment is random, and you compare the two groups over the same period under the same conditions. It is more work to build and it produces a smaller, more defensible number.

When you evaluate vendors, ask which one they used. A vendor quoting a large lift from a before-and-after is not necessarily wrong, but they haven't shown you evidence — they've shown you a coincidence with a hopeful reading.

Where Spequlo fits

Spequlo is a virtual try-on and fit-guidance platform built in Toronto for direct-to-consumer fashion brands. A shopper uploads one full-body photo and every garment they view is rendered on their own body; separately, Spequlo parses the brand's own size chart to give zone-level fit guidance, because — as above — those are two different jobs. It installs with a single script tag or a Shopify app embed, works from existing product photography with no reshoot, and logs every try-on against real purchase outcomes. Its analytics are designed to support a random session holdout rather than a before-and-after.

→ Add a fitting room to your product pages

Questions, answered

Is virtual try-on the same as a virtual fitting room?

In practice the terms are used interchangeably. "Virtual fitting room" is slightly older and sometimes implies a dedicated area of a site; "virtual try-on" more often describes the experience embedded directly on a product page. Neither term reliably tells you whether a tool does size guidance.

Does virtual try-on work for every type of clothing?

No, and quality varies more than vendors admit. Structured and fitted garments — dresses, tailored pieces, fitted tops, knitwear — render most reliably, because the garment's shape is largely determined by the body underneath. Flowing fabrics, complex layering, and garments with detailed prints, text or logos are harder, and artifacts show up in the detail first.

What's the difference between virtual try-on and a size recommender?

A size recommender outputs a size, usually from measurements, purchase history or return data. Virtual try-on outputs an image. They answer different questions and are frequently sold as one thing.

Do I need new product photography?

Not for generative try-on. It works from flat or on-model product images a brand already has. This is the main practical difference from the 3D-modelling approaches that preceded it, which required building a digital asset per garment.

Does it work on Shopify?

Yes — most current vendors ship a Shopify app that installs as an app embed, which means no theme-code editing. Check whether the vendor also supports your other storefronts if you run more than one.

Is virtual try-on available to Canadian brands?

Yes, though most vendors in the category are based in Europe or the U.S. If data location matters to you, ask specifically where inference runs and where images are stored, and get the answer in writing.