Should your clothing brand use AI models? Costs, results and the legal catch

By Andrej Ruckij · November 1, 2026

TL;DR: For product-page and catalogue images, probably yes, if you test first and check every image against the real garment. Tools cost from a few cents to about a dollar per image and need a clean flat lay or ghost-mannequin photo. The cost case is real: Zalando said in 2025 that generative AI cut the time to produce its imagery from 6-8 weeks to 3-4 days, at about 90% lower cost. The risks are less obvious. An image that shows the wrong texture or an idealised fit lands in the second most common reason people return clothes, and I found no published sales or returns data comparing AI models with real ones. The rules are also tightening: Google Shopping requires AI images to be marked, Zalando requires invisible marking by December 2026, the EU’s disclosure rule has applied since August 2, 2026, and New York requires written consent for any digital replica of a real model. I’d use AI where the image’s job is to show the product, keep real people where the brand relies on a face, and never reuse a real model’s likeness without a separate written consent.

Primores produces AI product photography for ecommerce: see how it works.

What changed

Reuters reviewed more than a dozen draft contracts that fashion brands send models before shoots. Some let the brand swap a model’s face onto another body, create digital replicas or train AI on the photos: a Diesel casting document asked models to accept face swaps, and a Calvin Klein draft allowed training on “reasonably unrecognisable images of Model” without extra pay. Calvin Klein says it is “testing and learning”; Diesel’s owner says that contract was for one campaign and isn’t typical.

It’s already in court. In Pujols v. Rainbow USA, a model says the retailer used AI to make new images of her from earlier commercial shoots, in poses and settings she never did; Rainbow denies wrongdoing. A New York judge refused her request to have the images taken down before trial, so the case is open. The question it raises for every brand is whether the right to use a photo covers AI-made new versions of the person in it. A Day Pitney analysis puts it plainly: “technological capability does not create a legal right to use the resulting content.”

At least one big retailer has gone much further: Zalando said in March 2026 that more than 90% of its content is now generated with AI; the example it gave was product photography, which went from about six weeks to a few days.

What it costs, and what the tools need from you

Every tool I looked at wants the same input: a clean photo of the garment, as a flat lay, on a ghost mannequin or on a hanger. FASHN’s documentation asks for photos “taken from a clean, top-down angle (flat lay style)” with “even lighting and minimal background clutter”, and says: “Avoid shadows, wrinkles, or cropped edges in product images.” That input photo is the part you control, and it decides most of the quality.

Option	What it needs	Price per image (as listed, November 2026)	Who it suits
General image models through an API (Google’s Nano Banana 2 and Nano Banana Pro)	Your garment photo plus a prompt; you build the workflow	about $0.07-0.24 at standard rates, about half in batch	A team with someone technical who will do its own checking
FASHN	Flat lay, ghost mannequin or hanger photo, up to 3 per image	roughly $0.07-0.50 on its plans, depending on resolution and whether you add a face reference	Small brands doing it themselves
Botika	Smartphone, supplier or customer photos are accepted	roughly $0.50-0.90 on its annual plans, with 2-3 retouch rounds included; its pricing page was inconsistent when read, so check	Brands that want fixes done for them
Photoroom Virtual Model	A photo of someone wearing the product, or a flat lay	not listed	Teams already using Photoroom
Lalaland (lalaland.ai now leads to Browzwear’s site), Veesual	Sales-led	no prices on their pages	Larger retailers

Two notes on these numbers. The vendor pages also carry results claims, such as Photoroom’s “96% average boost in sales”, with no sample, baseline or brand named; I’d ignore those. And the retouch rounds built into Botika’s price tell you something on their own: first outputs often need fixing, so budget time for checking, not only the per-image fee.

Where AI model images go wrong

The most useful list of failures comes from a retailer that rejects them. Zalando’s partner image guidelines (updated October 1, 2026) say it may reject AI images that “break believability” or fail to represent the product accurately, and group what it looks for in three kinds:

The product is wrong: textures, patterns, shapes, how the material behaves or finishes that don’t match reality, details that change across the image set, and garbled logos or text. Its examples include misplaced buttons and a brand logo merged with another brand’s.
The setting is implausible: unrealistic backgrounds and lighting, unnatural reflections, shadows at the wrong angle, or a model or product at the wrong scale for the room.
The model isn’t consistent: airbrushed skin, unnatural teeth, and faces, body proportions or skin tones that change across one model’s images.

Zalando is building the same check into its own pipeline: a system meant to compare each generated image with the original product photo for colour, construction and brand marks and flag details for correction, with people still reviewing.

Shoppers notice the same things. In a survey of 411 people by Stylitics (which sells AI imagery, so read it as vendor research), the examples of what broke trust were a wrong button colour, fabric with no wrinkles, a face that looked too airbrushed and images that looked “too perfect”, and the study adds that AI can idealise fit. 71% said the real and AI images they were shown looked the same or only slightly different; the study’s conclusion is that shoppers lose confidence quickly when product details are inaccurate.

A checking routine that follows from this: put each AI image next to the physical sample and check colour, buttons and hardware, logo and print placement, fabric texture and creases, how the fabric hangs, and whether the fit looks like the real garment on a real body. Then look at the model: hands, teeth, skin, and whether the same model looks the same across your catalogue. Reject anything that fails one item; don’t fix it in your head.

The returns risk

This is where a cheap image can get expensive. In PowerReviews’ 2023 survey of 25,452 US shoppers, 53% of recent apparel buyers had returned something. The reasons: sizing 39%, “item didn’t look like I thought it would” 28%, and “didn’t match the description” 13%. AI image errors sit exactly in that second reason: a fabric that looks heavier, a colour that’s off, a fit that’s flattered.

I found no published study of returns for AI-model images against real-model images. Zalando reports size-related returns down 8%, but nothing ties that to AI images. So treat returns as the number to watch in your own test, because no published data answers it for you.

What the results show so far

Cost and speed, self-reported: Zalando’s 6-8 weeks to 3-4 days and about 90% lower cost for imagery (2025); in 2026 it said it makes 70% more content at no extra cost.
Engagement: Scayle’s write-up of a masterclass on Zalando’s AI content lists 10% higher customer engagement and 17% higher click-through as results, with no source, baseline or method given.
Sales or conversion: I found no brand that has published a figure for AI models specifically. Veesual’s CEO claims, as Sourcing Journal reports it, that its model-switching feature doubles conversion rates, without published data.
How labels land: in three studies with more than 700 US women, published by SMU researchers in the Journal of Advertising, a plain “AI-generated” label on AI models in body-positivity ads lowered brand attitudes and purchase intent. Adding one sentence, that the AI was trained only on licensed photos of real people with diverse body types, brought the ads back to par with ads that had no AI. How you disclose seems to matter as much as whether you do.

Where the backlash came from

The public blow-ups so far share a cause, and it isn’t catalogue shots:

Levi’s and Lalaland, 2023: AI models announced as a way to add diversity. Critics read it as cheap representation that takes jobs from real models; Levi’s said AI would add to its live shoots, not replace them.
Guess in Vogue, 2025: an AI model in a print ad credited only in small print. A TikTok about it passed 2 million views in three days, with complaints about beauty standards, jobs and hidden labelling.
Rainbow, 2026: the lawsuit above, over reusing a real model’s likeness.

My read of the pattern: trouble came from claiming a virtue, hiding the disclosure or using a real person’s face without consent. A plain product image on a plain background hasn’t caused a backlash in any case I found.

The rules you have to follow

Google Shopping: Merchant Center requires AI-generated product images to carry the IPTC metadata tag trainedAlgorithmicMedia, and says not to remove it. It doesn’t ban AI images. Many image pipelines strip metadata during compression or upload, so check the file that actually reaches your feed.
Zalando partners: no visible AI labels or watermarks on images. Zalando requires an invisible marker by December 2026: “You must embed identifying data within the metadata or as an invisible watermark in the file code,” and the marker has to survive upload. It also won’t accept content that legally needs a visible label, including deepfakes, which it defines as AI content resembling real people, objects or places that falsely appears authentic.
Meta ads: images “created or significantly edited” with AI tools get an “AI info” label in the ad’s details; resizing or colour correction doesn’t count.
Amazon: I couldn’t read Amazon’s own rules, which sit behind a seller login. A third-party guide says AI-generated main images aren’t compliant and AI is allowed in secondary images; check Seller Central before you rely on that.
EU: Article 50 of the AI Act has applied since August 2, 2026. Anyone publishing a deepfake must disclose it, and a deepfake is defined as content that resembles existing people, objects or places and would falsely appear authentic. Whether a real garment on an invented model counts is still open; I’d assume it might and ask a lawyer. Fines go up to €15 million or 3% of worldwide turnover. Separately, AI tools already on the market before August 2 have until December 2, 2026 to add machine-readable marking to what they generate.
New York: since June 9, 2026, ads must clearly disclose a synthetic person who isn’t recognisable as a real performer, with fines of $1,000 and then $5,000. The law speaks about ads; it’s not clear whether product pages count. Digital twins of real models fall under a different law, the Fashion Workers Act.

If you use a real model’s likeness

New York’s Fashion Workers Act has required, since June 19, 2025, “clear and conspicuous prior written consent” before a brand creates or uses a model’s digital replica, stating scope, purpose, rate of pay and duration, with the agency’s consent obtained separately. Outside New York the protection is weaker: in the UK, as John Horner of the BFMA, the UK model agents’ association, puts it, “image rights do not exist as a standalone legal concept”, which is exactly why the contract matters.

What lawyers advising fashion brands recommend putting in it, from AFS Law’s analysis of the Rainbow case:

A separate AI consent, not a line in the general image release.
Tiers of consent: routine retouching, AI-assisted editing, material changes, and fully AI-generated replicas, each agreed separately.
Scope, purpose, pay and duration for each use, and whether the AI images can still be used after the campaign ends.
A check before publishing: was this image edited or generated, and is consent on file for that?
Clauses with your AI vendor on who verifies consent and who carries the risk.

On pay, the two reports I found point the same way: H&M says it pays its digital twins as it would for a shoot, and a model told City AM that for Zara’s AI edits she would “still be paid the same amount as if I’d travelled for another shoot”. The UK agents’ association also advises models to refuse open AI-training clauses, so expect agencies to push back on those.

How to decide

Who this is for: clothing and accessories brands selling online, with catalogue or product-page images to produce every season. Who it’s not for: brands built on a known face or ambassador, and luxury lines where the shoot itself is part of the brand story.

Use AI models if most of these are true:

Your images show the product, not a person. Product pages, catalogue and colour variants are the strongest case; brand campaigns with a recognisable face are the weakest.
Volume is high. If you shoot dozens of products or colourways a season, the cost difference matters; for five hero products it may not.
You can check every image against the sample. If nobody has time to do the routine above, the savings turn into returns.
You can mark and disclose where required. That means keeping metadata intact for Google Shopping and Zalando, and labelling synthetic people in ads where the law asks.
Any real person’s likeness is covered in writing. Otherwise use fully invented models, or real product photos with AI backgrounds only.

How to run a test

Pick 10-20 products with steady sales and a normal return rate.
Shoot clean inputs: flat lays or ghost-mannequin photos with even light, no wrinkles, nothing cropped.
Generate and check: make the AI-model images, check each against the physical sample, and record how many you rejected. That rejection rate is your real cost per usable image.
Split the products, not the shoppers, if your platform can’t A/B test images: half get AI images, half keep real ones, matched by category and price.
Run it for four to six weeks, long enough for returns to come in, and compare conversion, add-to-cart and returns, especially returns marked “not as expected” or “looks different”.
Decide on returns as well as conversion. A higher conversion rate that brings more “didn’t look like the picture” returns isn’t a win.

What I’d do this week

Take three of your products, shoot a clean flat lay of each, and run them through one of the cheaper tools above. Put the results next to the real garments and go through Zalando’s list. You’ll know within an hour whether the quality is there for your fabrics, and how much checking it will take.

More on AI imagery and disclosure: glossary/content-provenance and Do AI-generated ads work?

Sources

More of this every day on the channel: https://t.me/aidaily_business