What we shipped, twice
A plus size shopper uploaded her photograph and came back slim. A sixty year old came back twenty five. A male shopper came back on a woman's frame, in women's clothing.
In each case the face was recognisably hers. Everything below the neck belonged to the model in the product photograph.
What did not fix it
Saying so in the prompt, emphatically and repeatedly. You are not putting their face on the model's body. It made no difference.
Reversing the order of the two reference images did not fix it. Reframing the whole task as an edit of the shopper's own photograph, which is a shape we knew worked elsewhere, did not fix it either.
The actual cause
There were two bodies in the input. One belonged to the shopper and one belonged to the catalogue model wearing the garment, and the model's was the one lit properly, posed properly and photographed professionally.
No sentence outranks a photograph. The instruction was competing with an image, and images win.
The fix
Remove the competing body. The garment reference has to be a flat lay with nobody in it. Once the input contained exactly one person, all three shoppers came back as themselves on the first attempt.
The extraction prompt has to be blunt about it: remove the person entirely. No model, no body, no body parts, no head, no hands, no feet, no mannequin, no torso form, no implied wearer.
The practical problem this creates
Merchants have on model photographs, not flat lays. So the garment is derived once per product, cached, and reused for every try-on of that product afterwards. One extra generation, amortised across every shopper who ever tries that piece on.
That is the whole trick. Everything else, the ten second turnaround, the daily cap, the domain check, is ordinary engineering around one finding about what belongs in the frame.