Why Fashion's New AI Model Doesn't Exist

In the summer of 2024, Mango's Sunset Dream campaign ran across 95 markets without anyone flying anywhere to shoot it. A photographer took real pictures of the real garments, those images were fed into a generative model trained on Mango's own product data, and the campaign that resulted, a young woman in a black jumpsuit, another n sunset-lit sportswear, never involved a person standing in front of a camera. The retoucher's studio was the only physical location involved in the entire process.

Zalando has gone even further with the same approach. By the final quarter of 2024, over 70% of its editorial campaign images were AI generated, and the company has since built digital twins of its real models, three dimensional replicas that can be dressed, posed, and lit entirely inside software. A campaign cycle that used to take six to eight weeks now takes three to four days, and Zalando says the shift has cut its campaign production costs by around 90%. H&M followed a similar path in March 2025, creating digital twins of 30 of its own models in partnership with the Swedish studio Uncut, a version of the technology that was built specifically to keep a real person's likeness in the loop rather than generating a synthetic model from nothing.

None of these three companies has described what it is doing as a sustainability initiative. If you read their own language, the word that keeps coming up is efficiency, faster turnaround, lower cost, more content produced from fewer resources. Mango calls it part of its Earn lever, and Zalando frames it as responding to trends that can flip overnight. That distinction matters, because it means the emissions question, when it gets raised at all, is being raised by everyone except the brands themselves.

The emissions case is real

A traditional campaign shoot is a fairly large logistics operation, involving models, photographers, stylists, hair and makeup artists, producers, and often an entire crew flown to a location, housed, and fed for several days, on top of the equipment that gets shipped in separately. None of that happens when a campaign is generated on a laptop instead. The travel is the largest and most obvious cost that gets removed, and travel, across almost every consumption story SBT has covered, tends to be the single most carbon intensive line item in any production process.

Nobody at Mango, Zalando, or H&M has published a comparison of the emissions saved against the emissions created by the AI systems that are now replacing those flights. The number that would actually settle the environmental question is one that nobody involved has calculated in public so far.

But it's not the whole footprint

It would be a mistake to file this as a straightforward environmental win, because the alternative has its own footprint, and that footprint is mostly invisible in exactly the way a flight is not. Generative AI images are produced by data centres that consume electricity and, in many cases, significant volumes of water for cooling. A single image generation is a small fraction of what a long haul flight costs the atmosphere, but a fashion retailer generating tens of thousands of campaign images a year, across every product line and market, is not making a single request. It is running a data centre workload continuously, the same structural pattern already visible across the wider AI industry, where the sector's 2025 emissions were comparable to those of an entire city the size of New York.

The question nobody's answering as loudly

The other question sitting underneath the emissions one is not environmental at all, and it is the one that has been getting louder since H&M's announcement. Trade unions responded almost immediately. Equity in the UK warned that few legal protections currently exist for models whose likeness is used this way, and Bectu's Philippa Childs pointed out that even a well compensated model is only one part of a much larger crew, since the photographers, stylists, and makeup artists whose work simply doesn't exist anymore when a campaign is generated rather than shot. A casting agent told the Business of Fashion that for models who aren't already established, doing lower profile e-commerce work, AI represents a fairly straightforward threat to the jobs that used to exist.

H&M's own framing, that models retain rights to their digital twin and are paid per use, is a genuine attempt at an answer. It is also, notably, an answer to the labour question rather than the emissions one. The two conversations are happening in parallel, run by different people, and they rarely actually meet.

What "model" means when nobody needs choosing

There is a third layer to this that is easy to miss underneath the jobs and the emissions, which is that what a "model" actually is starts to change once the industry no longer strictly needs one. A digital twin, built from a real person's likeness with their consent, is a fairly different proposition from a fully synthetic model with no real counterpart at all, a difference Guess's AI-generated Vogue campaign made visible when it drew criticism for using two entirely fictional women rather than a consenting twin. An industry that has always organised itself around who gets chosen to represent a brand, a look, a body type, is now able to skip the choosing altogether. Mango's fully AI-generated campaigns have, so far, defaulted to the same narrow body type the brand was already using, which suggests that removing the human model doesn't automatically remove the narrowness. It just removes the person who might have pushed back on it.

Three brands have offered three separate justifications for this shift, cost, speed, and trend responsiveness, and underneath all of them sits one comparison that nobody has actually done the maths on. The emissions saved by not flying a crew around the world are real. So is the cost of the systems that are now doing the work instead. Nobody has published the number that would tell you which one is bigger, and for now, the fashion industry seems considerably more comfortable answering the labour question in public than the environmental one.