What is a digital twin of a model?

An AI version of a real, named model, built from photos of that person so a brand can make new images of them without a shoot. In H&M’s 2025 program, the 30 models kept the rights to their twins. In New York, a brand needs the model’s written approval before making or using one.

Can an AI model photo be the main image on Amazon?

Amazon’s main image for adult clothing must show the garment on a standing model on pure white, and every image must show the product accurately, including its color. Its rules do not ban AI-made people, but images of fully AI-made, photorealistic people must carry the contains-synthetic-performer tag.

Is a flat lay or a ghost mannequin photo the better input?

FASHN’s API and Photoroom’s Virtual Model both accept either one. A flat lay is easier to shoot well at home. A ghost mannequin shot shows the garment’s volume, which leaves the image model less to guess on structured pieces such as blazers and coats.

AI fashion photography: what it can do and what it gets wrong

Checked September 24, 20268 min read

Short answer

AI fashion photography turns photos of a real garment into on-model, flat lay and campaign images without booking a shoot. In 2026 it handles clean studio store photos and simple scenes well, as long as you give it good references and approve one model first. It still gets fit, drape, print placement, logos and hardware wrong often enough that a person has to check every photo against the real garment before it goes live.

What AI can make for a clothing brand today

Most AI fashion work starts from a photo of the real garment, either a flat lay (the garment laid flat and shot from above) or a ghost mannequin shot (the garment shot on a mannequin that is then edited out, so it looks worn by an invisible body). An image model then draws the garment again, on a generated person or in a new scene. It does not paste your photo in, so every seam and button in the result is the model’s own drawing of your product.

Job	Works well when	Still weak at
On-model store photos from a flat lay or ghost mannequin	The garment is pressed, you supply the front and the back, and the background is a plain studio sweep	Exact length on the body, how tight a size sits, stretch
Back and side views	You give the model a photo of the back and, ideally, the side	Anything it never saw, such as back pockets, yokes and labels, which it invents
Editorial and campaign scenes	The garment is simple and the pose is relaxed	Busy prints across seams, hands gripping fabric, layered outfits
Colorways	You photograph each color, or at least give the model a real swatch of it	Matching a color from a name or hex code alone, though Amazon requires the image color to match the product sold
Fabric close-ups	You name the weave and light it from a low side angle	It can draw a different weave or knit that still looks plausible
Try-on for shoppers	This is a different product that runs on the shopper’s own photo	See virtual try-on vs AI model photos

From our own apparel jobs and the limits tool makers publish. In every row, a photo of the real garment is the input.

What AI still gets wrong with clothes

The model makers say so too. OpenAI’s documentation for its image models warns that they can struggle to keep recurring characters and brand elements consistent from one image to the next. Google’s help page for its shopping try-on calls the result an approximation that does not guarantee how a garment fits.

What goes wrong	What it looks like	How to catch it
Fit	The garment looks tailored to the model. A boxy tee turns slim, a cropped jacket reaches the hip	Lay the photo next to the flat lay and compare length and width at chest, waist and hem
Drape	Heavy denim flows like rayon, or a chunky knit hangs without weight	Check that the folds match the weight of the fabric you sell
Prints and stripes	The pattern changes scale, or stripes stop lining up at the side seams	Count pattern repeats across the chest and compare with the garment
Logos and text	Letters change, or a printed graphic is redrawn	Zoom to 100 percent and read every word
Hardware	Buttons change count or color, rivets appear on trousers that have none, a zip changes metal	Count buttons, rivets, eyelets and zip pulls
Added details	Pleats, belt loops, pockets or a back yoke the garment does not have	List every construction detail before you render and tick each one off after
Size	The model looks like a size you do not sell, or not the size you list	Say which size the model wears, and see showing clothes on other sizes

On our own men’s trousers job, the store-photo prompt named every construction detail, including what the trousers do not have ("no rivets"), and closed the list with "nothing added and nothing removed." That prompt is printed in full in how to turn a flat lay into on-model photos.

How big fashion brands have used it

Brand	When	What they did	What followed
Levi’s	March 2023	Announced a pilot with Lalaland.ai to show products on AI-generated models of more body types, ages, sizes and skin tones	Critics said Levi’s should hire diverse models instead. On March 28, 2023 it added a note saying the pilot was not a way to advance diversity and that live shoots would continue
Zalando	Q4 2024, reported May 2025	Made about 70 percent of its editorial campaign images with generative AI. Its vice president of content solutions told Reuters this cut production time from 6 to 8 weeks to 3 to 4 days, and costs by 90 percent	It is also piloting 3D digital twins of real models, who take part voluntarily
H&M	Announced March 2025, first images July 2025	Began making digital twins of 30 real models with the Swedish tech company Uncut	H&M said the models keep the rights to their twins, are paid in line with rates their agencies negotiate, and AI images carry watermarks
Guess, in Vogue	August 2025 issue	Ran a print ad with two AI-generated models made by the agency Seraphinne Vallora	Readers objected to how real the models looked and to what it meant for models’ jobs. The AI credit ran in small print
Zara	December 2025	Started using AI to edit photos of real models so they appear in other outfits	Inditex said it agrees terms with the models and pays in line with industry practice. Separately, its shopper try-on reached 43 markets by March 2026

The criticism in these cases followed two choices: presenting AI-made people as progress on diversity (Levi’s), and running AI-made people in a fashion ad that readers took for a real shoot (Guess). H&M, Zalando and Zara each described consent and pay for real models when they announced their programs. Read how to cast AI models responsibly before you cast.

A workflow that keeps the garment right

Photograph every garment yourself: flat, front and back, pressed, on a plain background, plus a close-up of the fabric and one of each label, print and piece of hardware.
Write down the garment facts before any prompt: fiber and weight, fit (slim, straight, relaxed), where it ends on the body, and every construction detail, including the ones it does not have.
Cast one model and approve a character sheet first. That is one image of the same person from the front, side and back with face close-ups, wearing your garment on a plain background. See how to keep the same model in every AI photo.
Plan the shots: front, back, side, a detail, then one or two scenes. Keep backdrop, light and framing the same across the store photos so the carousel lines up.
Give each shot only the references it needs: the sheet and the front flat lay for a front view, the sheet and the back flat lay for a back view, the fabric close-up for a detail.
Render two to four versions of each shot, keep the best, and check it against the real garment on the table, using the AI image QA checklist.
Tag and label AI-made people where platforms or laws ask for it (see the last section).

If you already shoot flat lays or ghost mannequin photos, you have most of the inputs. Ghost mannequin, flat lay or on-model compares the three styles and what each costs.

How many tries it takes, and what it costs

Our agency made two apparel campaigns with AI image models in 2026, built by hand. A men’s trousers job, store photos on a model plus editorial shots, took 40 renders for 11 photos we kept, about 3.6 tries per keeper. A surf apparel job, one model in a studio, on a cliff and in the sea, took 163 renders for 12 keepers, about 13.6 each. Its character sheet alone ran 20 times before one was kept.

The model bill is small. At Google’s published prices in September 2026, one 1K image costs about 3 cents on Nano Banana 2 Lite, 7 cents on Nano Banana 2 and 13 cents on Nano Banana Pro, so even 14 tries stay under two dollars per keeper. The bigger cost is the time someone spends reviewing. For full budgets, see how much AI product photos cost.

Questions people also ask

What is a digital twin of a model?: An AI version of a real, named model, built from photos of that person so a brand can make new images of them without a shoot. In H&M’s 2025 program, the 30 models kept the rights to their twins. In New York, a brand needs the model’s written approval before making or using one.
Can an AI model photo be the main image on Amazon?: Amazon’s main image for adult clothing must show the garment on a standing model on pure white, and every image must show the product accurately, including its color. Its rules do not ban AI-made people, but images of fully AI-made, photorealistic people must carry the contains-synthetic-performer tag.
Is a flat lay or a ghost mannequin photo the better input?: FASHN’s API and Photoroom’s Virtual Model both accept either one. A flat lay is easier to shoot well at home. A ghost mannequin shot shows the garment’s volume, which leaves the image model less to guess on structured pieces such as blazers and coats.

Where Overs fits

Overs makes on-model photos from one photo of your garment. For worn products it first makes a character sheet, which you approve before any other photo, and each later photo gets only the reference pictures it needs, such as the sheet and then the back of the garment. The plan is free for 40 photos a month; the AI costs a few cents a photo on your own OpenRouter key (OpenRouter is one account and bill for many AI models).

Free for 40 photos a month. The AI that makes the photos is billed separately, on your own key, with no markup from Overs.

Start free How it works

Sources

Checked on September 24, 2026. Prices, specs and rules change; follow the links for the current versions.