AI Models vs Real Models: How Fashion Brands Decide

No published test answers the AI-or-human question, and the numbers usually quoted for it measure something else. The most-cited study compared AI-generated garment designs against human-designed garments, not AI photography against studio photography. So decide on what you can observe: which gallery slot the shot fills, whether it matches the garment you ship, and who carries the disclosure risk.

That reframing matters because the question is usually asked at the wrong level. Brands ask “should we switch to AI models,” decide once, and apply the answer to a whole catalogue. The gallery does not work that way. A packshot, a fit shot, and a campaign hero each have to prove a different thing, and only one of them is really about a person. The order they move in is set by how much your existing images constrain each slot.

This covers what the research measured, why the published lifts cannot settle the argument, the slot-by-slot rule that replaces them, and what EU law has required since 2 August 2026.

Key Takeaways

• The 13% figure is not about photography. It comes from an Alibaba deployment that compared AI-generated garment designs against human-designed garments (arxiv.org/abs/2503.22182, KDD 2026 ADS track, accessed September 2026). The AI arm contained different products.

• Triple-digit case-study lifts are a sample-size artefact. At a 2.5% baseline, detecting a 10% relative lift needs roughly 64,000 sessions per variant. Tests that report a 128% win are usually reading noise.

• Decide per gallery slot, not per catalogue. The packshot, the scale shot, the fit shot, and the campaign hero each answer a different question. Only two of them depend on a person at all.

• Platforms regulate accuracy first. Zalando accepts AI-generated partner imagery that meets its quality criteria, rejects assets that fail to represent the product accurately, and asks for invisible AI marking (partner.zalando.com, updated 31 August 2026). Walmart requires AI images to be “truthful, accurate and not misleading” (marketplacelearn.walmart.com, accessed September 2026).

• The disclosure risk is concentrated, not spread. Article 50 of the EU AI Act has applied since 2 August 2026. The marking duty falls on the provider of the generation system. The deep fake disclosure duty falls on you, and it turns mostly on realistic people. A product still counts only if it misleads about the product.

• Fit is where AI on-model imagery costs you money. Fit problems are a leading reported cause of apparel returns, and a flattering drape that the garment does not have converts once and comes back.

• Produce both arms from the same source photo. If the control is last year’s studio shot and the variant is a fresh render, you compared two eras of your brand, not two treatments.

Do AI models convert better than real models?

There is no reliable general answer, and the effect on your catalogue is unmeasured until you measure it. The studies and case studies circulating in this category either compared a different variable or ran at sample sizes too small to detect the effect sizes that image treatments produce. Treat AI on-model imagery as a cost, coverage, and speed decision with an unknown conversion effect, rather than as a conversion tactic with a known return.

That is a less satisfying answer than the one most vendor pages give. It is also the only one the evidence supports.

What the most-cited study compared

One statistic anchors nearly every article on this topic: AI visuals delivering roughly 13% relative improvements to click-through and conversion, alongside a 7.9% drop in returns. The research is real and peer reviewed. It is routinely described as proof that AI fashion photography beats studio photography.

It is not that.

The paper is “Sell It Before You Make It: Revolutionizing E-Commerce with Personalized AI-Generated Items”, deployed at Alibaba and accepted to the KDD 2026 ADS track (arxiv.org/abs/2503.22182, accessed September 2026). Merchants described a garment in text. The system generated the design and a photorealistic image of it on a digital model, and the listing went live before the garment existed. Manufacturing started once orders arrived.

The comparison was AI-generated garment designs against human-designed garments. Different products, selected by a preference model trained on shopper behaviour. The photography was downstream of the variable that moved.

Quote that number to justify replacing your packshots and you have imported an effect size from an experiment that did not run your test. Your expected lift is unknown, and sizing a test around a borrowed 13% will size it wrong.

Why the published lifts cannot settle it

The other evidence on offer is vendor case studies, and they tend to report lifts of 128%, 150%, or 30%. Those numbers are not dishonest. They are unreadable, and the reason is arithmetic rather than motive.

Sample size comes from your baseline conversion rate and the smallest lift worth detecting. At a 2.5% baseline, detecting a 10% relative lift needs roughly 64,000 sessions per variant. A 30% lift needs around 7,800. Most single-brand case studies run on a fraction of that.

An underpowered test cannot return a small wrong answer. It returns a large one. A test with a few thousand sessions can only ever surface enormous swings, so the wins that get written up are precisely the ones least likely to survive a rerun. The full session counts each level of lift requires sit in the traffic math, and they are the first thing to check before you believe any published number in this category, including your own.

Image treatment effects on an established catalogue usually land in the single digits to low teens. That is the range hardest to measure and the range where the decision matters.

Decide per gallery slot, not per catalogue

Once the conversion argument is set aside, the decision becomes tractable, because the gallery is not one thing. Each position answers a different shopper question, and the questions have different tolerance for a synthetic person.

Read the “depends on a person” column and the argument resolves itself. Four of the six slots barely involve a model. Those are the slots where AI generation is uncontroversial, cheap, and fast, and where most catalogues have gaps today. The two that do involve a person are where the real questions live, and they pull in opposite directions.

The practical consequence: most brands should not be choosing between AI and real models. They should be filling four slots with AI, shooting one with a camera, and thinking hard about the fifth.

That decision gets made once per slot, and then it is mechanical. Settle the framing, the model and the brand rules for the packshot slot on one garment, save the job, and run the same job on the next garment, and the four hundred after it. Deciding what a scale shot should look like takes an afternoon. Producing the four hundredth one to the same standard as the first is the part that eats a season, and it is the part worth building as a workflow rather than a habit.

Where AI on-model imagery is the stronger choice

Coverage you do not have at all. A catalogue missing the in-scale frame, the detail crop, and half its colourways loses more to those gaps than it will ever gain from optimising a treatment it already ships. Coverage beats optimisation when the gaps are this obvious.

Long-tail SKUs that will never justify a shoot. The economics of a studio day do not stretch to a 400-SKU tail. What a product photoshoot costs sets the baseline that decides where the line falls for your catalogue.

Colourway and variant permutations. The garment is identical and only the fabric colour changes. Reshooting that is waste.

Demand testing before manufacture. This is the one place the Alibaba result genuinely applies, because it is the thing that study measured.

Representation range. Showing a garment across more body types than a casting budget covers is a real gain, provided the drape stays honest for each one.

Where a real shoot still wins

Anything that carries brand meaning. A campaign hero is an argument about who the brand is. That is the slot with the least tolerance for approximation and the one where an audience is most primed to notice.

Fit fidelity on complex garments. Tailoring, drape, knitwear, and anything with structure are where generated on-model imagery is most likely to flatter. Fit is not a cosmetic problem. Fit is a leading reported cause of apparel returns, so an image that improves the click and misrepresents the hang converts once and comes back. Run the accuracy checks before you ship on your hardest garment, not your easiest.

Texture that has to be legible. Where the purchase decision is the material itself, photograph the material.

Categories where the body is the product. Swimwear, lingerie, and activewear are judged on fit against a real body more than on styling.

The honest version of the tradeoff: AI on-model imagery is strongest where the garment is simple and the shot is functional, and weakest where the garment is structured and the shot is emotional. The consistency locks that hold a model steady from drop to drop matter more as you push further into the second category.

What platforms and regulators require

The rules in force do not ask whether pixels came from a camera. They ask whether the image matches what ships.

Zalando accepts AI-generated partner imagery and holds it to its quality criteria. Its image guidelines, updated 31 August 2026, say “Synthetic and AI-generated content must meet our high-quality editorial and technical criteria”. Zalando reserves the right to reject AI assets that “fail to represent the product accurately”, and it rejects images of poor quality from “low quality AI generation (e.g. distorted logos, mispositioned buttons, unnatural teeth or shifting facial features)” (partner.zalando.com, accessed September 2026). It does not allow visible on-image labels, tells partners not to provide content that requires visible labelling under applicable laws, and strongly recommends invisible AI marking in the metadata or an invisible watermark. Its video guidelines add that “low-quality AI video footage characterised by pixelation, motion blur, unnatural textures or anatomical inconsistencies is not permitted” (partner.zalando.com, accessed September 2026).

Walmart requires that “images generated by artificial intelligence must be truthful, accurate and not misleading”, and sellers must check AI content before they publish it (marketplacelearn.walmart.com, accessed September 2026). Etsy requires disclosure when the item itself is made with AI (etsy.com/legal/sellers, accessed September 2026). Separately, Etsy requires original photos of the actual product, and does not accept renderings or stock photos in their place, with two exceptions for mockups (etsy.com, accessed September 2026).

The pattern is consistent. Fidelity to the product is the binding rule, and provenance comes second. Full per-marketplace detail sits in what each marketplace requires for AI product images.

Article 50 of the EU AI Act has applied since 2 August 2026. The Commission’s guidelines, published 20 July 2026, treat “realistic AI-generated human avatars or personas” as persons, so a photorealistic AI-generated person is likely in scope even when no real individual is depicted (Commission guidelines, C(2026) 5054, accessed September 2026). The guidelines are not binding. The machine-readable marking duty under Article 50(2) sits with the provider of the generation system. The deep fake disclosure duty sits with you as the deployer.

So the exposure is uneven, and it concentrates in one slot: the synthetic human. A packshot or a flat lay of your product is not a deep fake as long as it does not mislead about the product. A photorealistic person wearing your garment likely needs a visible disclosure in the EU. In New York, an ad that you know contains a synthetic performer must say so, since 9 June 2026 (nysenate.gov, accessed September 2026). On Amazon, images of photorealistic people made fully by AI take the contains-synthetic-performer tag (Amazon product image guide, accessed September 2026). Which assets in a drop trigger a label, and what a label costs, is covered in labeling AI-generated fashion images.

That is a real cost line on the on-model slot specifically, and it belongs in the comparison alongside the cost of the run.

Verify your own position against the regulation and your marketplace’s current policy before scaling a treatment across a catalogue. This is general information, not legal advice, and a compliance question should not rest on a blog’s reading of it, this one included.

Produce both arms from the same product photo

Whatever you conclude, the test is only as clean as the assets. If your control is a two-year-old studio shot and your variant is a fresh render with a different crop and different lighting, you compared two eras of your brand.

Generate both arms from the same source photo of the real garment and change exactly one attribute. That is the practical argument for producing variants from your own product image rather than commissioning two separate shoots: the source is held constant by construction.

In DesignerBox, one workflow starts from one product photo and makes the packshot, the styled scene and the on-model shot from the same input, with virtual try-on as the on-model step. The cost depends on the model you pick, and the cost is shown before the run, which is what turns a 200-SKU plan into a budget instead of a guess. Run one SKU’s gallery, read the cost, and plan from that number. Virtual try-on starts on the Premium plan at $75 a month billed monthly, or $30 a month billed annually, which is the honest gate to know about before planning around it. The commercial licence starts on Pro. Plan detail is on the pricing page.

Two further things worth knowing. A model saved once from a model creator template keeps the same face and body across a drop instead of a new stranger per SKU, and where an AI fashion model comes from decides what you owe on disclosure. Model choice is not cosmetic either: image models differ in how they handle fabric on a body, so test your hardest garment on more than one.

For the fit question specifically, what the research shows on virtual try-on accuracy covers what fit-aware generation changes and what it does not.

FAQ

Do AI fashion models convert better than real models?

No published evidence settles it. The most-cited study compared AI-generated garment designs against human-designed garments rather than AI photography against studio photography, and single-brand case studies typically run below the sample size needed to detect realistic image effects. Treat it as a cost and coverage decision with an unmeasured conversion effect.

Can I use AI models for my whole catalogue?

You can, but the gallery-slot table above suggests you should not want to. Packshots, scale shots, detail crops, and variant coverage are the slots where generation is strongest. Campaign heroes and fit-critical categories are where a real shoot still earns its cost.

Do I have to disclose AI-generated model images?

In the EU, a photorealistic AI-generated person wearing your garment likely needs disclosure. Article 50 has applied since 2 August 2026, and the Commission’s guidelines, published 20 July 2026 and not binding, treat realistic AI-generated people as persons even where no real individual is depicted. Packshots and flat lays are outside that trigger as long as they do not mislead about the product. In New York, an ad with a synthetic performer must say so. Check the regulation and your marketplace’s policy directly for your own market. This is general information, not legal advice.

Will marketplaces reject AI-generated product images?

Not for being AI-generated, in the cases verified here. Zalando accepts AI imagery that meets its quality criteria and represents the product accurately, and asks for invisible AI marking. Walmart requires AI images to be truthful and accurate. Etsy requires original photos of the actual product, not renderings or stock photos, which is a fidelity rule rather than a provenance one.

Do AI model images increase returns?

They can, if the render flatters the garment’s drape or fit. Fit is a leading reported cause of apparel returns, so an image that lifts click-through while misrepresenting the hang can raise returns enough to erase the gain. Track return rate on tested SKUs for a full return window before calling any winner. Generous returns policies also mask the effect, which is why generating a size range has to be read against returns rather than conversion alone.

How many sessions do I need to test AI against real model images?

At a 2.5% baseline conversion rate, detecting a 10% relative lift needs roughly 64,000 sessions per variant and a 30% lift needs about 7,800. Below those counts you cannot separate a result from noise. Run the arithmetic on your own baseline before you design the test.

What is the cheapest way to compare the two treatments?

Generate both arms from the same source photo so only one attribute differs, then screen them at the ad layer where impressions are purchasable rather than on a product page where traffic is not. Take the surviving two candidates to the PDP only if your traffic supports it.

Sources

All accessed September 2026.

• What the 13% click-through and conversion figures and the 7.9% return reduction compared: “Sell It Before You Make It: Revolutionizing E-Commerce with Personalized AI-Generated Items”, KDD 2026 ADS track (arxiv.org/abs/2503.22182)
• Zalando partner image guidelines, including AI quality criteria, the ban on visible on-image labels and the invisible marking recommendation, updated 31 August 2026 (partner.zalando.com)
• Zalando partner video guidelines and the AI video footage policy, updated 31 August 2026 (partner.zalando.com)
• Walmart Marketplace requirement that AI-generated images be truthful, accurate and not misleading (marketplacelearn.walmart.com)
• Etsy seller policy on AI disclosure for items (etsy.com/legal/sellers) and the listing image requirements (etsy.com)
• EU AI Act Article 50 and the Commission’s guidelines published 20 July 2026, C(2026) 5054, on realistic AI-generated people (ec.europa.eu)
• New York General Business Law 396-b on synthetic performers in ads (nysenate.gov)
• Amazon’s contains-synthetic-performer tag (sellercentral.amazon.com)
• DesignerBox plans and feature gating: DesignerBox pricing page (designerbox.ai/pricing), September 2026

Research claims verified against arXiv:2503.22182, Zalando partner guidelines, Walmart Marketplace policy, Etsy seller policy, New York law, Amazon’s image guide and European Commission AI Act guidance as of September 2026. Sample sizes computed with the standard two-proportion formula at 95% confidence and 80% power. Regulatory position varies by market. This is general information, not legal advice. Individual results vary.