Text2Villa: Hierarchical Generation of 3D Indoor Environments with Physics-Aware Analysis-by-Synthesis

Text2Villa is a new hierarchical system for generating villa-scale 3D interiors from natural language, with a focus on geometry that actually holds up in a game engine. The big shift is that it does not stop at room layout: it also tries to keep objects physically believable during placement, which is where many text-to-3D approaches fall apart.

At the macro level, the system uses a multi-story dataset to fine-tune an autoregressive layout generator. That lets text drive building foundations with polygonal boundaries, multiple floors, and vertical connectivity instead of the usual single-room or boxy results. For environment artists and tools teams, that matters because it moves text generation closer to something that can seed real production blockouts.

At the micro level, Text2Villa introduces an Affordance-driven Physical-Semantic Scene Graph, or A-PSSG, to encode support surfaces, containment cavities, and other placement constraints. Scene instantiation is then treated as a closed-loop analysis-by-synthesis problem, where collision detection and multimodal LLM reasoning work together to observe, evaluate, and modify the scene until penetrations, floating assets, and containment failures are resolved.

The practical takeaway is that AI scene generation is starting to look less like a one-shot prompt and more like a constrained editing system. If this approach holds up outside the paper, it could be useful for rapid prototyping, procedural interior generation, and interactive content tools where physical plausibility is non-negotiable.