How to put your clothes on an AI model from one photo

Any photo that shows your garment clearly can be the starting point. Each kind has a trade-off:

Use a photo of your own product. Don’t start from another brand’s photos, an influencer’s post or any picture of someone else’s face or body: borrowing their work or likeness is exactly what buyers and platforms object to.

You don’t need a studio. You need a photo where nothing about the garment is hidden or distorted.

1. Use soft, even light. Daylight from a window, without direct sun, keeps colours true and avoids hard shadows that the AI may read as a pattern or a seam.
2. Get the whole garment in the frame: collar, both sleeves and the hem. Anything cut off has to be invented.
3. Steam or iron it first. Creases and fold lines can come back as part of the fabric.
4. Use a plain background that contrasts with the garment: a white top on a white sheet loses its edges.
5. Shoot straight on, from roughly the middle of the garment, so the proportions aren’t stretched.
6. Make sure the print, embroidery and any text are sharp. If they are small, move closer rather than zooming in later.
7. Compare the photo with the garment in the same light before you upload it. If the colour on screen is off, fix the photo, because the AI will copy the wrong colour faithfully.

Before making anything, write down what the garment is: the category, cut and fit, length, neckline, sleeves, fabric, colour and print. Then decide which of these stay fixed in every image and which one you want to test.

• Lock the things that define the product: the category, the cut, the fabric, the length. If these drift, you are no longer looking at your bestseller.
• Open only what you want to learn about, such as a new colour, a new print, a different occasion or one changed detail.
• Change one thing per version. If version B has a new colour and a shorter hem, you won’t know which one buyers responded to.
• Treat detail changes (a new neckline, a new length, different buttons) as concepts. They don’t exist yet, so they need a sample before you can promise them.

If version A is on one model in a studio and version B is on another model outdoors, a better result for B could be the model, the setting or the garment. To learn something about the garment, everything else has to stay the same: the same AI model, the same pose, the same light and the same framing in every version.

• Choose a model who looks like the people who buy from you, in skin tone, age and body type.
• Be specific about size. Describe the model’s size and build; left vague, image models tend to draw slimmer bodies.
• Use the same model for the whole test. Try a different model in a separate test, if you want to know how the garment reads on another body.

AI images of clothing are good at the overall look and weak at small, exact things. Put the real garment next to your screen and go through every image before it goes into an ad.

If an image is wrong, redo it rather than fixing it by hand in an editor. An image that shows a detail your product doesn’t have will mislead buyers, even in a test.

No AI image can show how the garment actually fits, stretches or feels. Treat it as a picture of the idea; get a sample made before you promise fit or fabric details to a buyer.

Ads and stores want different shapes. Make each one natively, framed for that shape, rather than cropping one image into the others: cropping a square into a tall image either cuts off the garment or leaves you with too few pixels.

Two more things from the same pages. Shopify recommends a consistent aspect ratio for product images, so they line up on collection pages ^[3]. Google asks for the product to fill between 75% and 90% of the image ^[1], which matters more for a flat lay than for an on-model shot. For TikTok, check TikTok’s own ad specifications; we haven’t summarised them here.

An on-model image of a version you haven’t made yet is a concept image. Say so wherever a buyer can see it, and swap it for real photos as soon as the product exists.

• Keep the metadata. Google says images made with generative AI must carry metadata saying so, such as the IPTC DigitalSourceType property, and that you shouldn’t strip it ^[1]. The IPTC value for an image created with generative AI is trainedAlgorithmicMedia ^[2]. Re-saving through some editors or compressors can remove it, so check after exporting.
• Be careful with shopping feeds. Google wants the main image to show the exact item and variant being sold, and doesn’t accept mockups or illustrations that don’t depict the physical product ^[1]. Our suggestion: keep concept images out of your Google Shopping feed until you have real photos, and read Google’s page for your own case.
• Label ads and product pages. Meta, TikTok and Google each have their own AI labelling rules; our guide to AI labels on ads and product images goes through them.
• If you take orders before the product is made, give a real ship date. See pre-orders for clothing.

Can I start from a photo of my garment being worn by a real person?

Yes, if it is your own product photo. Only the garment is read from it. The person shouldn’t be edited, copied or turned into the model; the new images should use an AI model instead.

Will the AI image show how my garment fits?

No. Drape, stretch and fit in the image are the AI’s guess. Use the images to see whether people are interested in a version, and use a sample to check fit.

Can I use these images on my product page?

For testing interest or taking pre-orders, with a clear label that they are concept images and a real ship date. Replace them with real photos once the product is made. For Google Shopping, see the caution above ^[1].

Why not crop one image into all three ad sizes?

A tall 9:16 crop of a square image either cuts off the hem and shoulders or has to be enlarged. An image framed for 9:16 from the start keeps the whole garment inside the area that the app’s buttons and captions don’t cover ^[5].

How many versions should I test at once?

Few enough that each one gets a fair share of your ad budget. Each version should differ from the original in one thing only, so a result tells you something about that one change.