Virtual try-on

Two different products share the name

The single biggest source of confusion in this category is that virtual try-on describes two products with different buyers, different quality bars and different failure costs.

Production try-on is a seller-side tool. The brand runs it once per SKU to generate on-model catalogue imagery, a human looks at the output, and only approved renders get published. A bad render costs one credit and is thrown away.

Consumer try-on — often called a virtual fitting room — is a shopper-side feature embedded in the product page. The shopper uploads a photo of themselves, or picks a body close to theirs, and sees the garment on it. Nobody reviews the output before they see it, and a bad render is seen by a customer.

• Production: run by the brand, reviewed before publishing, output is catalogue imagery
• Consumer: run by the shopper, unreviewed, output is a fitting aid
• Consumer try-on carries a privacy surface production try-on does not — it takes photographs of real customers

How the pipeline works

Modern try-on is a generative image process rather than a geometric one. Older systems tried to build a 3D garment and drape it on a 3D body, which needed a per-garment digital sample and largely explains why the technique stayed niche for a decade.

The generative approach skips the 3D step. It takes two images — the garment, and the body — and generates a new image conditioned on both, learning from training data what fabric does on a body rather than simulating it. That is why it works from a single product photograph you already have, and equally why its errors look like plausible-but-wrong fabric rather than like a physics glitch.

• Garment input — flat-lay, hanger, ghost-mannequin or on-person photo
• Body input — a model from a library, an uploaded reference, or the shopper's own photo
• Segmentation — the garment separated from its background and the region of the body it should occupy identified
• Generation — a new image synthesised, with the garment's colour, print and trim carried across

What sellers use it for

In production use the job is almost always coverage rather than novelty: getting on-model imagery for SKUs that a photoshoot could not economically cover. That is most acute for long-tail catalogues, one-off inventory and colourways — thirty colours of the same tee is thirty shoots, or one shoot and twenty-nine renders.

The second common job is representation. Showing one garment on a range of body types is a scheduling and budget problem for a shoot and a repeat operation for a renderer, which is why extended size ranges tend to be under-photographed and are the clearest win here.

What it still gets wrong

Try-on is strongest where fabric behaviour is predictable and weakest where a garment's construction is the product. Reliably hard cases are worth knowing before you plan a catalogue around it, because they are the renders a human has to reject.

It is also important to be clear about what try-on is not: it is not a fit prediction. A render shows a plausible drape at a chosen body size; it does not tell a shopper whether their actual measurements will work in that garment. Presenting it as sizing advice is the main way the technique gets brands into trouble.

• Hands, and anything a hand overlaps — the most common visible defect
• Text and logos on the garment, which drift because a model treats type as texture
• Complex layering, where an open jacket over a top has to resolve both correctly
• Sheer, mesh and lace, where the fabric has to be partly transparent to read correctly
• Fit accuracy — a render is an image, not a measurement

Common questions

See also

Doing this with TinTin