When Levi's announced in 2023 that it would test AI-generated models to increase the diversity of bodies shown on its website, the reaction was immediate and largely hostile.
The criticism was not that the images looked bad. It was the framing. A brand had presented the simulation of representation as though it were representation — offering customers the appearance of inclusion while removing the paid work that real inclusion would have created. Levi's clarified its position and scaled back the messaging. The lesson travelled further than the campaign did.
Since then, Mango has run generated campaign imagery, H&M has publicly explored digital twins of real models, and synthetic figures have appeared in mainstream advertising to recurring controversy. AI generated fashion models are no longer hypothetical. The unresolved question is not whether brands will use them, but on what terms.
The Business Case, Stated Honestly
Vendors describe this as creative liberation. The actual appeal is simpler and worth naming plainly.
- Cost. A photoshoot carries model fees, photographer, stylist, hair and makeup, studio hire, location costs, and post-production. Generated imagery removes most of it.
- Speed. A campaign that took six weeks to schedule, shoot, and retouch can be produced in days.
- Volume. For a retailer listing thousands of products, photographing each on a body is economically impossible. Generation makes the long tail of the catalogue viewable on a figure at all.
- Variation. The same garment shown on several body types, skin tones, ages, and heights without four separate shoots.
- Localisation. Regional campaigns adapted without flying a crew to each market.
- Iteration. Changing a background, a pose, or a colourway without rebooking anyone.
These are real advantages and pretending otherwise helps nobody. The disagreement is about who absorbs the cost of capturing them.
Why the Diversity Argument Backfired
The most instructive part of the early controversies was the collapse of a specific justification.
The argument went: fashion has historically excluded most body types, and generating a wider range of figures shows customers people who look like them. That is a genuine problem and a plausible-sounding solution.
It failed for a reason that is obvious in hindsight. Exclusion in fashion was never a rendering problem. It was a hiring problem — a matter of who got booked, paid, and put on a set. Solving the visual symptom while eliminating the economic opportunity inverted the entire point. Models from underrepresented groups, who had spent years fighting for those bookings, watched the industry propose to synthesise their presence instead of hiring it.
The general principle applies well beyond fashion: using a technology to simulate the outcome of a fairness effort, while removing the mechanism that would have produced it, is not progress. It is the appearance of progress at a discount.
Who Actually Loses Work
Discussion tends to focus on models, but the affected group is broader:
- Models, particularly those at the start of their careers doing e-commerce work — the entry-level jobs that build a portfolio.
- Photographers and their assistants.
- Stylists, hair and makeup artists, and set designers.
- Studio operators and equipment rental businesses.
- Retouchers, whose work is being absorbed into the generation process itself.
The e-commerce shoot mattered more than its glamour suggested. It was the reliable, unremarkable work that funded people while they built toward something else. Removing the bottom rung of a ladder does not only affect whoever is standing on it.
The Consent Model That Might Actually Work
The most promising development is not synthetic figures invented from nothing. It is licensed digital replicas of real people.
The structure works roughly like this: a model is photographed and scanned under a specific agreement, a digital version is created, and the brand licenses that version for defined uses over a defined period. The model is paid for the scan and again for each usage, retaining approval rights over how the likeness appears.
H&M's publicly discussed approach moved in this direction, with the model retaining ownership of their digital twin and agency representation continuing to negotiate on their behalf.
For this to be genuinely fair rather than cosmetically so, several conditions have to hold:
- Scope limits. The licence covers named uses, not "any purpose in perpetuity."
- Ongoing compensation. Payment per usage rather than a single buyout that transfers the asset forever.
- Approval rights. The person can refuse specific placements, particularly anything political, sensitive, or off-brand for them.
- Expiry and deletion. The replica has an end date, and the underlying data is destroyed.
- Non-transferability. The licence cannot be sold to another company without fresh consent.
- Independent representation. Agencies and unions at the table when the terms are set.
The single-payment buyout is where this goes wrong. A one-off fee for a permanently reusable likeness is not a licence. It is the sale of a career.
Disclosure Is Becoming Non-Optional
Regulation has moved faster here than most brands expected. The European Union's AI Act includes transparency obligations for synthetic content, advertising regulators in several countries have taken positions on undisclosed generated imagery, and consumer protection law already prohibits materially misleading product representation.
The practical position for a brand is straightforward:
- Label generated imagery clearly, near the image rather than buried in a footer.
- Never generate a product's appearance in a way that misrepresents the actual garment — fabric behaviour, colour, or fit.
- Keep records of what was generated and how, because you will eventually be asked.
- Assume disclosure will be required in your largest market within the planning horizon of any campaign you are commissioning now.
Brands that treat labelling as a competitive disadvantage are misreading the audience. Undisclosed synthetic imagery discovered by customers is far more damaging than disclosed synthetic imagery accepted upfront.
The Body Image Question
There is a harm here that is easy to overlook because it is diffuse.
Fashion imagery already shaped body expectations through casting and retouching. Generated figures remove the last physical constraint — the requirement that the body in the image belong to someone who exists. A generated model can be proportioned in ways no human is, and can be produced at a scale and consistency that previous techniques could not match.
The audience most exposed to this is young and already navigating a media environment that measurably affects self-image. A brand deploying this technology without thinking about that is making a decision, whether or not it acknowledges making one.
Where This Is Genuinely Uncontroversial
Not every application raises these questions. Several are close to universally accepted:
- Ghost mannequin and flat-lay imagery — no human was going to be photographed anyway.
- Background replacement and set extension on shoots that did use real people.
- Colourway variants of a product already photographed on a model.
- Rendering from 3D design files, where the garment itself is digital.
- Fit visualisation shown to an individual shopper on their own body.
These represent most of the actual efficiency available, without touching the ethical core of the debate. A brand can capture the majority of the commercial upside while leaving the casting of human beings alone — a trade many are quietly making.
The Commercial Risk Nobody Budgets For
Beyond the ethics, there is a plain business risk that gets underestimated: generated imagery can quietly damage conversion.
Product photography does a job beyond looking attractive. It communicates scale, texture, weight, and how a fabric moves. A generated image can get all of that subtly wrong — a knit that reads heavier than it is, a colour rendered a shade too saturated, a drape the material cannot produce. The shopper does not consciously notice. They simply receive a parcel that does not match the expectation the image set, and they return it.
The failure is invisible in the marketing dashboard, because the campaign performed well. It appears three weeks later in the returns report, and few teams connect the two.
Retailers testing this properly are running it as an experiment rather than a rollout:
- Split-test generated imagery against photography on matched products.
- Measure return rate and return reasons, not only click-through and conversion.
- Watch for a rise in "not as pictured" and "colour different" reason codes.
- Track repeat purchase rate, which reveals whether trust survived the first order.
A Policy Checklist for Brands
If you are going to do this, do it deliberately:
- Decide and publish where you will and will not use generated figures.
- Never use a real person's likeness without a specific, compensated, time-limited agreement.
- Disclose clearly and consistently, across every market.
- Keep the product representation accurate — this is a legal line, not a preference.
- Maintain a real photography budget, and say what it is.
- Consult the people whose work is affected before announcing, not after the backlash.
- Review annually, because both the law and the audience are moving.
The Underlying Question
The technology is not the interesting part any more. Generation quality crossed the threshold of usefulness some time ago, and it will keep improving whether or not anyone is comfortable with it.
The interesting question is one the industry keeps deferring: when a technology makes labour optional, who captures the saving?
Right now, the answer in most implementations is the brand, entirely. The models, photographers, and crews who built the visual language being imitated receive nothing. That is a choice, not an inevitability, and the licensed-replica model shows that a different arrangement is technically possible.
AI generated fashion models will be judged less on how convincing they look than on how the businesses using them chose to distribute the benefit. That decision is being made now, mostly quietly, and mostly without the affected parties in the room.
0 Comments