AI Fashion Photoshoot: Keeping the Garment Real

On-model photography is the most expensive thing a clothing brand does and the most obvious thing to point AI at. A model, a studio, a stylist and a photographer for one day costs more than most small labels spend on everything else that season.

So the question is not whether the images look good. They do. The question is whether the garment in the picture is the garment in the box — and that is a narrower question than it sounds.

Everything below comes from one test run: a flat-lay generated as a stand-in for a real product photo, then put on a model twice. Both results are shown unretouched.

Can AI put your actual garment on a model?

Better than expected, for the things a shopper notices first. Given a flat-lay as a reference, the on-model result kept the stripe pattern exactly — seven teal stripes across the body in both — along with the colourway, the camp collar, the wooden buttons and the embroidered fox on the chest pocket.

That is a genuinely strong transfer, and it is the part most guides get right. A reference-guided edit is not inventing a shirt; it is dressing a person in one it has been shown.

Which makes the failures more interesting than the successes, because they are all in the details a customer only discovers after the parcel arrives.

What does the model get wrong?

The material and the pattern pieces. Magnified side by side, the reference fabric is a fine linen slub and the generated version is a coarse, open basketweave — a visibly different cloth. The chest pocket also changed shape, from an angled hem in the reference to a curved one on the model.

Neither is visible at listing-thumbnail size. Both are visible in the hand, and both are the kind of thing that generates a return.

• Weave and drape. Models render a plausible textile rather than your textile. Linen becomes cotton, jersey becomes ribbed, a fine gauge becomes a chunky one.
• Pattern pieces. Pocket shapes, hem curves, placket widths, collar points. These are your pattern, and the model treats them as suggestions.
• Hardware. Button size, hole count and finish drift, which matters when the button is a design feature.
• Fit. The model renders a flattering fit rather than your grading, which is a sizing complaint waiting to happen.

What happens to the parts the pose hides?

The model invents them, and it invents them wrongly. Asked for a lifestyle shot with the subject turned away, it produced a convincing street scene — with a full patch pocket and the fox embroidery stitched onto the back of the shirt, where no pocket exists.

This is the most useful failure in the whole test, because it explains the mechanism. The model is not reasoning about garment construction. It is satisfying a description, and if the described feature cannot be seen from the requested angle, the most direct way to satisfy the description is to move the feature.

The consequence is a rule rather than a fix: only generate angles where the referenced surface is actually facing the camera. Back views, side views, detail crops of unseen panels and anything three-quarters away are where invention creeps in, and they are exactly the shots a listing needs most.

Does the same person come back?

Not automatically. The two shots in this test came from the same garment reference but returned two different men, because the reference described a shirt and said nothing whatsoever about the person wearing it. The garment was anchored and the model was left free to change.

For a lookbook or a product range this matters as much as the garment does. A collection shot on four different faces reads as four unrelated listings, and shoppers use the model's continuity to understand that items belong together.

The fix is the same technique applied twice: anchor the person and the garment separately.

• Approve one model image first, before any garment work, and treat it as a fixed asset.
• Supply both references — the person and the flat-lay — and state that neither may change.
• Re-anchor to the originals every time rather than chaining from the last output, which compounds drift. The same rule as in the character consistency guide.

Will it follow an exact specification?

No, and this failed before the model was even involved. The flat-lay prompt asked for exactly nine teal stripes and the generated shirt came back with seven. Counted on the finished file, not estimated. If you generate the garment itself, you are not producing your product — you are producing a product.

That is the whole argument for photographing real clothing rather than generating it. The stripe count is trivially checkable, which is why it makes a good test; the things that are not trivially checkable fail the same way and you will not notice.

It also means the reference image has to be a real photograph. A generated flat-lay used as a reference simply propagates its own errors into every shot that follows, which is what happened here and is worth being explicit about.

So what is it actually good for?

Context rather than record. Lifestyle backdrops, seasonal mood, campaign concepts, location scouting and testing whether an idea is worth shooting properly are all cases where nobody will hold the image up against a delivered parcel, which is exactly where the drift documented above stops mattering.

Based on the test run described in this article. Marketplace rules on AI-generated product imagery differ and change — check the policy where you list.

The economics still work in that narrower lane. A location lifestyle shot that would need a travel day costs about nine cents to try, and you can decide whether the concept is worth a real shoot before booking one.

What does it cost against a real shoot?

Around nine cents a frame on your own API key against a day rate that runs into the hundreds or thousands once model, photographer, studio and styling are counted. The saving is real but it buys exploration rather than replacement, because the shots you can trust are the ones you still have to take.

A more useful way to frame it: generation is cheap enough that the sensible workflow is to generate first and shoot second. Try twenty settings, twenty crops and twenty moods for a couple of dollars, pick the three that work, then book a shoot that produces those three properly.

That is the same argument as testing ad creatives before commissioning them, applied to a category where the product itself has to be accurate.

When should you just book the shoot?

Whenever the image is the thing a customer buys from. Product listings, size guides, fabric detail shots, anything a return would be argued over, and any brand whose proposition is the material itself. Generated imagery cannot document a garment it has only been shown a picture of.

Stated plainly, because this guide is published by a company that sells image generation:

• The main listing image should be a photograph. Every drift documented above lands in the customer's hands eventually.
• Fabric-led brands have nothing to gain. If the cloth is the selling point, the one thing that reliably changes is the cloth.
• Fit and sizing imagery must be real. A rendered fit is a returns problem, not a photography problem.
• Disclosure rules apply and change. Several marketplaces now ask whether imagery is AI-generated; check the current policy where you list rather than an article.
• Real models are people. If you would have hired someone, replacing them with a synthetic face is a decision worth making deliberately rather than by default — and it carries its own disclosure questions.

Frequently asked questions

The bottom line

Photograph the garment, then let AI place it. Reference a real flat-lay rather than a generated one, only request angles that show the referenced surface, anchor the model separately from the clothing, and keep generated frames out of the listing itself.

The encouraging part of this test is how much did transfer. Seven stripes stayed seven, the fox stayed on the pocket, the collar stayed a camp collar. Given a real photograph to work from, these models are good at dressing people.

The limit is that they are dressing people in something that resembles your garment. For a campaign, resemblance is enough. For the picture someone clicks buy on, it is not.