ZabSoft

I build software for curtain salons. A salon's client picks a fabric from a swatch and has to imagine it at home. CurtainBot, our existing product, lets them send a photo of their room and a photo of the fabric and get a try-on back. Its weak part was the engine: a general image-editing model, asked to change the curtains, sometimes changed the room too, and 6 of the first 12 production jobs failed. A salon can't sell a try-on of a room that isn't the client's room. During the hackathon I built a new try-on engine on Meta's models, in a fresh repository. The generator never gets the last word. Every picture is measured against the client's photo (the photo's edges, the window frame, the tulle) and judged by three votes, and one that fails is drawn again with a note on what was wrong. After three tries the client is told it didn't work instead of being shown a bad picture. The first version pasted only the generated curtains into the original photo, so no other pixel could change. On real rooms the pasted curtains came out skewed, with cut hems and halos. I threw that out, switched to showing the generator's whole picture and spent the effort on measuring it. On a fixed set of 20 real interiors, 17 try-ons passed every check, 14 on the first attempt. In the accepted pictures a median 99% of the photo's edges outside the curtains are still in place; raw Muse Image output changes a median 25% of the room. One accepted try-on costs about $0.11, failed attempts included, and takes one to two minutes. What existed before: the CurtainBot bot, accounts and payments. What was built during the hackathon: everything in the repository, meaning the engine, the demo page and the evaluation.