Abstract
Recent developments in 3D generative AI enable users to create bespoke 3D models from text or image prompts. However, language is often insufficient to express spatial design intent, and customization of generative output remains limited to surface appearance, making it difficult to personalize designs to the geometry and constraints of a user’s environment. Bridging this gap typically requires returning to conventional CAD tools to measure, edit, and adapt geometry, tasks that remain time-consuming and error-prone. We present MiXR, an augmented reality system that enables users to create new 3D designs by remixing parts harvested from real-world objects. Users paint to segment real-world geometry into reusable parts, recompose the parts directly in AR through spatial manipulation, and refine the assembly into a unified, fabrication-ready 3D model via a generative AI pipeline. By treating spatial assembly itself as the input modality, MiXR replaces text prompting with rough physical arrangement as a more reliable and expressive way to specify design intent. In this demonstration, attendees will construct novel 3D models by harvesting parts from objects in their environment, recomposing them in AR, and refining them with generative AI into coherent unified designs.