The right product photo for AI try-on: what the model looks at

Cabina does not ask you to upload your photos a second time: it reads the garment image from the product page the shopper is on. It looks first for the og:image tag, the one social networks use for the preview when someone shares the link; if it is missing, it tries the main image of the theme’s gallery.

Shopify exposes og:image on every product page, so there is nothing to do there. WooCommerce, on the other hand, does not emit it without an SEO plugin: in that case Cabina falls back to the gallery, and if it cannot find that either the try-on stops partway. The fix is adding the og:image tag to your product template — the same thing that makes your links look good on WhatsApp and Instagram.

A simple rule follows: the photo that counts is the first one, the one social networks would show. If your product page has ten images and the first is a close-up of a button, the model will see a button.

The generative model has to understand one thing only: which garment to put on the shopper, and what it looks like. Anything that makes this reading ambiguous makes the result worse.

• One garment per image. A complete outfit — jacket, top and trousers in the same photo — forces the model to choose, and it does not always choose what you are selling.
• The whole garment. A long dress cropped at the waist leaves the model to make up the rest.
• From the front. The front view carries the most information: neckline, sleeves, length, how the fabric falls. A three-quarter or back view can work as a second image, not as the first.
• Flat lay, mannequin or worn: all fine. A packshot on a neutral background is the simplest case, but a photo worn from the front works too: the garment reads just the same.

A uniform background helps the model separate the garment from everything else; a busy one — a street, a shop, another garment hanging behind — confuses it. Even light matters for the same reason: a hard shadow can be read as a crease or a different colour.

And watch out for text: to the model, everything in the image is part of the image. A “-30%” badge printed over the photo, a watermarked logo, an arrow pointing at a detail — these are things it may try to put on the garment. Promotions belong in the theme, not in the photo.

You do not have to tell Cabina whether the garment is a top, trousers, a dress or outerwear: a vision model reads the image and works it out, and that reading decides how the garment is worn — a coat, for instance, stays over what the person already has on. The earlier rule applies here too: a clear photo classifies well, an ambiguous one does not.

If you have uploaded size charts, the same category is used to pick the right chart for the size recommendation. Charts are uploaded from CSV in the dashboard, once.

It is the other half of the result, and it does not depend on you: it depends on how they posed. For that Cabina has a dedicated page, which shoppers see before uploading and which you can link too: the photo guide. In short: standing, straight, facing the camera, hands visible. Anyone who prefers not to use their own photo picks one of the preset models and sees the garment on a body shape similar to theirs.

• The first image on the product page is the one I want tried on.
• There is a single garment, whole, seen from the front.
• Uniform background, light without hard shadows.
• No badge, text or watermark over the photo.
• The page exposes og:image (automatic on Shopify).
• I tried the garment on the free plan before enabling it for everyone.

The last point is the most important one: the Freemium plan gives 10 try-ons a month, and ten are enough to see how your garments come out before switching the button on across the catalogue. How to install it is in Shopify virtual try-on in 5 minutes.