Turning floor plans into 3D rooms with AI-generated code

Generating realistic 3D rooms from images has been tricky—text descriptions lose spatial details, and AI agents tend to loop infinitely or crash when building entire rooms. Code-as-Room solves this by having a multimodal AI analyze a floor plan, then systematically write Blender code to create geometry, materials, and lighting in stages. A memory module tracks decisions across stages to prevent the agent from forgetting context. The authors released a benchmark for this task and show their approach outperforms existing agent-based methods.