Look Builder

The idea

Fashion sites show you one product at a time. People buy outfits. Look Builder is an attempt to close that gap: pick a top, trousers, shoes and a bag from a store's actual catalogue, and see them together on a person.

The hard requirement is honesty. If the shirt in the picture has a pocket the real shirt does not, the picture is worse than no picture.

What exists

• A browser wardrobe. Prepared avatars, slot-by-slot selection, a randomize button and a running total. Choosing an outfit is instant, because nothing is generated at that moment.
• An asset studio. Takes catalogue items, including from a Shopify development store, and works out what preparation each one needs.
• A try-on pipeline. Takes real product photos, renders them on the avatars with image models, and holds every result for human review before it can be used.

What I have found so far

• One garment is the easy case. Models do a fair job with a single item, some of the time.
• Outfits are the hard case. Combine two or three products and the models start inventing: an extra pocket, missing buttons, a different fabric. So far, no model has produced an exact multi-item look reliably.
• Paying more did not fix it. The hosted models fail on the same requirement as the local ones.
• Confidence scores cannot be trusted. In one batch, the outputs I rejected by eye were the ones the model was most sure about. A confidence threshold would have approved known failures.
• The simple route fails too. Pasting a cut-out product photo onto a portrait, with no AI at all, looked wrong without hand-tuning.

Where it stands

It is unfinished, and I am still working on it. My current guess is that the answer is not one model. It is several routes, each good at one kind of garment, behind a single review process.

This one may yet fail. That is allowed in here.