Experience the look before the customer orders
At engbers, customers can upload a photo of themselves on the product detail page and see the garment on their own body. Live since July 2026, built with Google Vertex AI. How the service works, what it costs and where it reaches its limits.
engbers has been one of the online pioneers in fashion since 2000, and we have supported the menswear retailer's digital development since 2009. In July 2026, a virtual try-on was added to the product detail page. Anyone who wants to see how a colour works with the rest, or whether the material suits their own style, uploads a photo of themselves and sees it, instead of imagining it. The function is available on mobile and opens as a panel right on the product page. We are still in the hypercare phase, measuring usage and adding what shows up in real operation.
What it looks like in the shop
The journey of the image
Photo in the frontend
The customer picks a photo right on the product page. The frontend checks format and size before the upload.
Our service on AWS
Assembled from AWS components. It protects the endpoint against misuse, manages the credentials and fetches the product image.
Google Vertex AI computes
Customer photo and product image go in, one result image comes out.
Back to the customer
The same way back into the frontend, labelled in line with Article 50 of the EU AI Act.
The real work is not in the model, it is in the layer in front of it. We are happy to show in detail which AWS components we put together for this, and why.
What happens to the customer's photo
It stays on the device
Permanently, the photo sits only in the customer's browser. It is transmitted once for the computation, but not stored.
Nothing happens without consent
The customer actively agrees before choosing a photo. The consent is recorded together with the consent text in the wording shown to them.
The customer stays in control
They can delete the images at any time by clearing the website data in their browser. A new photo overwrites the old one.
What a try-on costs and how long it takes
Figures from live operation, as of July 2026.
The questions data protection asks
Consent is a separate step in the process, before the upload. The customer is informed and has to actively agree. Only then is the button unlocked and the next interaction possible.
Yes. We store the consent and the version of the consent text that was valid at that moment. If the wording changes later, it stays traceable who was told what and when.
In the customer's browser. The photo is transmitted once for the computation and is not stored in the process.
The images disappear when the consent session changes, when the customer clears the website data in their browser or when they upload a new photo for the same item.
Yes, exactly as Article 50 of the EU AI Act requires for synthetic image content.
Where the approach reaches its limits
Six points from our tests and from live operation. Some come from the model, some from the technology around it. Anyone planning a virtual try-on should know them beforehand.
It does not check the fit
The model renders the garment onto the existing photo and makes the person slightly slimmer in the process. It shows colour, material and combination, not size and fit. That expectation must not arise in the first place.
Accessories get invented
With bow ties, ties and shoes the model hallucinates. We did not roll the function out for those categories.
Staged product shots interfere
Elaborately staged product images produce artefacts in the result. The cleaner the source image, the more usable the try-on.
The upload was the hurdle
The 6 MB Lambda limit was the real technical boundary. High-resolution photos and Apple formats could not be uploaded at first, and we added that.
No local storage, no try-on
Anyone browsing in private mode cannot use the function, because the images are kept locally. The customer needs to be told this in plain language.
Product data stays as it is
That was the pleasant surprise. Neither special preparation nor new product images were necessary, the model works with what is already there.
Hypercare means we are not done yet
A function like this is not finished when it goes live. Since July we have been measuring where users drop off and working through what stands out in real operation.
The first feedback on the result images is positive. The data basis is not yet sufficient for solid statements on conversion or returns, we are building it up right now.
Is this worth it for your shop?
A virtual try-on is not a plugin you switch on. For some assortments it pays off, for others it does not. The effort sits in consent, the image pipeline and the error cases. We look at your assortment and tell you whether it fits.