Model Swap at Marketplace Scale

One of Europe's largest fashion marketplaces lists hundreds of thousands of articles from thousands of brands. A big share of that catalog is never-out-of-stock (NOS): core styles that stay online for years. Those products are photographed once, and then the imagery has to stay fresh, on-brand and relevant to very different audiences, without a reshoot every season.

That is the problem the team brought to On-Model by PiktID: how do you keep on-model imagery current, varied and personalized across a catalog far too large to reshoot?

Why a marketplace turns to model swap

Two pressures pushed them toward AI model swap rather than more photography.

The first is rights and refresh. Photographed models come with usage licenses that expire. When they do, the imagery either has to be renewed or pulled. Because On-Model swaps in generated models, the marketplace refreshes its imagery on its own schedule, with no likeness to re-clear and no shoot to re-book.

The second is personalization. A marketplace serves many audiences, and the same article can convert very differently depending on who it is shown on. The team wanted to present NOS products on a range of models and measure which resonated, the kind of localization and enrichment we cover in campaign localization and enrichment.

The hard part is doing this while keeping the product identical and the result coherent. A product page is unforgiving: faces are often cropped, models face away from camera, and the garment carries high-resolution detail (weave, trims, hardware, print) that has to survive untouched. A model swap that alters the garment is worthless for a PDP.

And consistency cuts both ways. The garment has to stay pixel-faithful, but the new model has to hold together too: lighting, skin tone, and the model's expression and pose all have to match the original frame and stay consistent across every angle of a product page, and across the entire catalog. Get that wrong and a product set reads like a collage of different shoots. Holding all of it steady, at that fidelity and at catalog scale, is exactly what made this a hard problem, and why the team wanted a specialist.

Purpose-built models, not a generic filter

Off-the-shelf image generators redraw the whole frame, garment included. That is the opposite of what a PDP needs. On-Model runs on models PiktID trained for this job specifically:

• Onda, our proprietary model-swap engine, replaces the person while keeping the garment pixel-faithful, so the product on screen is still the product that ships.
• Orbita, our model-generation system, produces authentic, unbiased human models who do not exist as real people. (You can read the Orbita model card for how it is built and evaluated.)

Because the models are generated rather than cast, the marketplace sidesteps licensing renewals and likeness clearances entirely, the anonymization advantage that first brought the team to us.

A talent catalog, categorized by segment

Personalization needs range. So the first deliverable was not a batch of images but a talent catalog: a library of AI models built for the marketplace and organized the way its buyers think, by gender and by fashion segment, from designer and premium to modern mainstream, sport, streetwear and young fashion.

With that catalog in place, any NOS article can be matched to a model that fits its audience, and re-matched to another one whenever the team wants a different read.

Model swap across the catalog, via API

From there the workflow is programmatic. Through the On-Model API, the marketplace swaps the model on its NOS articles at scale: same garment, same pose, a new model, feeding personalized on-model imagery straight into the storefront where it can be tested against real conversion.

The examples below are AI-swapped product pages, one womenswear, one menswear. In each, the garment and every product detail are preserved exactly; only the model has changed.

"For a marketplace the whole game is consistency: the model can change, but the garment cannot move a pixel. Once the swap is that faithful, personalization stops being a photoshoot problem and becomes a catalog operation you can run and measure at scale."

— Nunzio Alexandro Letizia, Co-founder at PiktID and creator of On-Model

Why it matters

For a catalog this size, the value is not a single nicer image. It is turning model imagery into something the team can refresh, vary and optimize on demand: no expiring licenses, no reshoots, and the freedom to serve different AI models to different audiences and let conversion decide. Photograph the core range once; keep it current forever.

If you run a large catalog and want on-model imagery you can personalize at scale, start free or explore model swap.