The Challenge
Helping customers see the full fit
A styling service that sends clothes to people who have never seen those clothes on a body like theirs has always been asking for a small act of faith, because a photo of a garment on a stranger tells a shopper almost nothing about how it will sit on them, and faith is not something a subscription business can keep asking for indefinitely without losing the people who run out of it. The company wanted its clients to see themselves in the outfit being recommended, not a model standing in for them, since confidence in a recommendation and willingness to keep the recommended item were the same problem wearing two names.
The Approach
Generating photorealistic looks
Solvd built a system that generates a photorealistic image of the client wearing the proposed outfit, built from the client’s own uploaded photos and constructed to preserve their actual fit and proportions rather than approximating them, because an image that gets the proportions wrong does more damage to trust than no image at all. Every generated image passes through an automated realism and fidelity check before a client ever sees it. That meant the threshold for “good enough” couldn’t be subjective; it had to be a consistent pipeline decision, applied the same way every time before an image was shown to any customer.
The Outcome
Rollout at scale
The feature reached 100,000 enrolled users generating roughly three personalized images per person every week, at a scale the company was confident enough in to announce publicly on its own. GPT models were used in building the automated quality-assurance pipeline that gates every image before release, which is the detail that tends to get lost in a headline about photorealistic try-on: the interesting engineering problem was never generating a plausible image, it was deciding, automatically and at volume, which images were good enough to show a real customer.
