World Studio

World Studio is a shared 3D workspace where people build and test spaces with an embodied AI collaborator. Small shop owners can start with a photograph, ask Muse to reconstruct the store, and edit fixtures and merchandise through natural language or direct manipulation. The problem: a photo captures appearance, but trying different layouts in a real shop takes time and disruption. Muse Spark interprets the image, describes visible evidence and uncertainties, creates custom fixture geometry and a product catalog, and proposes scene edits. Deterministic code validates scene data, detects footprint overlaps, and offers placement repair before changes are applied. Simulated shoppers then navigate the layout, browse products and reach checkout. A fixed seed supports repeatable comparisons; results can be sent back to Muse for another proposal. Prices, stock and dimensions inferred from photos are estimates, and shopper outcomes are programmed simulations rather than sales forecasts. In a live blocked-entrance test, Muse moved one object and the same 12 simulated shoppers went from zero purchases to 11, with no give-ups or geometry conflicts. A separate photo reconstruction produced a 13-object mini-mart; code caught two overlaps, placement repair resolved them, and all 12 simulated shoppers purchased. The workspace also includes an experimental EV concept scene with editable body dimensions, generated components and analytical packaging checks. Its bounded design-check-repair loop feeds calculated failures back to Muse for up to three attempts. These are concept estimates, not manufacturing validation. Built with Next.js, Three.js and a Python Meta API bridge. Scenes can be exported and reopened. Public UI: https://unwritten-demo.onrender.com/studio . Live AI and photo processing currently require running locally with a server-side Meta API key; the hosted demo keeps those endpoints restricted. Latest source and setup:https://github.com/blazefluz/unwritten