How to roll out AI virtual try-on across a fashion catalogue (without a returns problem)

If you manage product imagery for a fashion catalogue, the maths on model photography stops working somewhere between a few hundred and a few thousand SKUs. Every new drop needs a shoot. Every shoot needs a model, a photographer, a studio and two to three weeks. Marketplace sellers and multi-brand retailers hit this wall hardest, because the catalogue is never finished.

AI virtual try-on for clothing changes the unit economics. You start from the product images you already have, whether flat lay, ghost mannequin or on-model, and generate realistic on-model imagery for the entire range, on a schedule that follows your inventory instead of a photographer's calendar.

This article covers what virtual try-on actually means, what goes wrong with it at scale (and how we fixed it), how Photoroom fits into the imagery workflow you already run, and how to go from a first test image to try-on on every product page.

Two kinds of virtual try-on (and which one you need)

The term gets used for two different things, and most catalogues end up wanting both.

Try-on for sellers: Your team puts a garment on an AI model to produce listing and marketing images. This is Virtual Model, available in the apps for a quick test and through the API for the full catalogue.

Try-on for shoppers: A customer uploads a photo of themselves on your product page and sees the garment on their own body. This lives inside your storefront or app, is aimed at conversion and returns, and Photoroom delivers it through the Virtual Try-On API on Enterprise plans.

The sensible order for most catalogues is seller try-on first, because it fixes the imagery you are already paying for, then shopper try-on once the catalogue is consistent. Both run on the same engine and the same integration, so the second step is an extension of the first, not a second project.

Why AI virtual try-on matters at catalogue scale

Three quarters of fashion executives say AI is already part of their operations, and product imagery is where it lands first. On Photoroom, clothing is already the most background-removed category on the platform, processed about four times as often as food and twice as often as toys. The reasons fashion catalogues move to AI on-model imagery are consistent.

• On-model sells where flat lay does not: The head of ecommerce at a multi-brand fashion retailer put it to us bluntly: "I don't think a woman would buy a dress from a ghost-mannequin image." Their location shoots beat studio flat lays in every ad test, so they stopped shooting isolated garments for apparel. We hear the same from enterprise apparel sellers repeatedly.

• Photography cost scales with SKUs, AI does not: A 70-year-old apparel retailer we work with estimated that photographing roughly 1,000 products with real models would have cost tens of thousands of dollars. With AI models the cost per image is flat and predictable, whatever the catalogue size.

• Imagery keeps up with the catalogue: New stock arrives Tuesday, listings are live with on-model images Thursday. No shoot lag, no backlog of products sold from a flat lay while they wait their turn.

• Consistency across thousands of listings: One model set, one lighting setup, one background system applied to every SKU, rather than whatever each season's shoot happened to produce.

• Diversity without casting: Show the same garment on different body types, ages and skin tones across the range, including region-specific models built for the markets you sell in.

What goes wrong with AI try-on at scale (and what we did about it)

Every tool in this category promises realism. And we have to admit that realism is actually mostly solved. The problem we see across live usage of Photoroom's AI Fashion Models is different: garment detail drift. A logo redrawn slightly wrong, a button that moved, printed text turned into nonsense. The model looks fine; the product is no longer your product. On ten images you catch it by eye. On ten thousand you do not.

That matters because of returns. For one of Europe's largest fashion marketplaces, a single percentage point rise in return rate wipes out every euro saved on image production. When returns are on you, fidelity to the garment is the economic constraint that decides whether AI imagery pays off at all.

So we built the tooling around the product, not the model. Our AI Fashion Rater checks generated images for fidelity to the original garment. Every image processed through the API returns an uncertainty score, so your pipeline can auto-publish confident results and route edge cases (reflective fabrics, multi-item shots, intricate prints) to human review before they reach a listing. Our ML team also found that training a model narrowly on virtual try-on underperformed a single generalist editing model, which is why Virtual Model runs on the same engine as the rest of Photoroom and improves every time that engine does.

Our research with sellers turned up one more thing the category gets wrong: brands put a product on a model to communicate size and fit, yet most tools frame and pose the image around the person, and every generation is an island with no consistency from one SKU to the next. We are building Virtual Model around the garment and around the catalogue for that reason.

How Photoroom fits into the workflow you already run

This is the part that decides whether try-on is a project or a feature. Photoroom is built to slot into an existing imagery pipeline rather than replace it.

• It starts from the images you have: Flat lay, hanger, ghost mannequin, on-model: whatever your current photography produces is the input. There is no reshoot to feed the system.
  

• One integration, every product: Connect the Photoroom API once and try-on applies across the catalogue. No per-product setup, and new listings inherit it automatically as they are added, which is what makes it work for marketplaces where sellers are uploading around the clock.

• It fits the QA process you already have: The uncertainty score on every image plugs into your existing review step. Set a threshold, publish what passes, review what does not. Your team spends its time on the edge cases instead of eyeballing every image.

• Output lands where your listings live: Images come back in the dimensions each channel expects, ready for your PIM, your storefront, or marketplace feeds such as Amazon and Shopify.

• Days, not months: A typical integration takes days, with onboarding and integration support included on Enterprise plans. Most customers run a scoped proof of concept on a subset of the catalogue first, compare against their current imagery, then roll out.

• Same engine, both directions: The integration that generates on-model listing images is the one that later powers shopper try-on on your product pages. You do not rebuild anything to add it.

Test it in ten minutes before you connect anything

You do not need an integration to judge quality. The fastest way to see how your garments behave is to run a handful through Virtual Model in the Photoroom web app.

1. Upload a product image: Pick a few of your harder items: a printed tee, something with a visible logo, a textured fabric.

2. Choose a model and pose: Browse the gallery and pick the look that matches your brand.

3. Choose a scene and format: Start with studio. Sellers on Photoroom pick clean studio backdrops about three times as often as location scenes like a cafe, because a controlled background keeps the attention on the garment.

4. Generate and compare: Photoroom returns several variations. Check the garment details against your original, not just the overall look.

5. Export: Download at the dimensions your listings use and drop them next to your current images for a side-by-side.

If the hard items hold up, the catalogue will. That is the point to talk to the enterprise team about running it through the API.

Where it pays off first

1. Catalogue backfill: The long tail of products still sold from a flat lay or ghost mannequin image. This is usually the largest and fastest win.

2. New drops and seasonal collections: Shoot the winter collection in July. Listings go live with on-model imagery the day stock lands.

3. Marketplace supply: For marketplaces, apply consistent on-model imagery to seller uploads without asking sellers to change anything.

4. Channel variants: One source image, every crop and format your channels need, without a separate export job.

5. Testing: Model, styling and background variations across a set of SKUs to learn what converts, at a cost that makes testing routine rather than a campaign.

But: AI models are not right for every brand. A skate shop we spoke to uses Photoroom for backgrounds and cleanup but will not put its products on AI models, because the local skaters who model for them are part of what the shop stands for. Second-hand sellers sometimes find AI models make used items look too new. If your imagery is part of your credibility, keep the humans in it and use AI for everything around them.

Virtual try-on for your shoppers: the second step

Once the catalogue is consistent, the next lever is letting the shopper be the model.

With Photoroom's Virtual Try-On API, a customer on your product page uploads one photo of themselves and sees the garment on their own body in seconds. It runs inside your own storefront or app, under your brand, from the same product images and the same integration already powering your listings.

The early numbers are why retailers are asking for it. On one marketplace customer's listings, shoppers who used Virtual Try-On converted at 3.7%, against 0.9% for shoppers on the identical listings who did not. Roughly four times higher, from removing fit uncertainty before checkout.

Because it is built on the same fidelity tooling described above, the garment the shopper sees is the garment they receive, which is what keeps try-on from turning into a returns problem. Pricing depends on catalogue size and volume.

If you run a marketplace, a multi-brand retailer, or a catalogue in the thousands of SKUs, talk to our enterprise team or read the API documentation.

A rollout path that does not disrupt your current shoots

Moving to AI try-on does not mean dropping photography overnight. Most catalogues follow the same sequence.

• Run a proof of concept on a subset: A few hundred SKUs, including the awkward ones. Compare against current imagery on the metrics you already track: conversion, sell-through, return rate.

• Backfill the long tail first: Basic product-on-model shots are the largest share of most photography budgets and the easiest to move.

• Then make it the default for new products: New listings get on-model imagery automatically through the integration; shoots are reserved for hero campaigns.

• Add shopper try-on: With a consistent catalogue in place, switch on try-on across product pages through the same API.

Further reading: How to replace fashion photoshoots with AI virtual models

Take control of your fashion imagery

With Photoroom you can produce catalogue imagery that used to need a full creative team, on a schedule set by your inventory. Every product on a model, every listing consistent, and when you are ready, every shopper able to try the garment on themselves.