ProxyBuild: Text-Guided Structured 3D Building Generation with Mesh-Anchored Procedural Proxies

arXiv cs.GR details ProxyBuild, a text-guided 3D building generation framework from Xiang Tang, Ruotong Li, and Xiaopeng Fan. The core idea is to move away from monolithic meshes and toward a structured pipeline that keeps building parts editable and spatially grounded.

ProxyBuild introduces Mesh-Anchored Procedural Proxies, or MAPP, as an intermediate representation. In practice, that means the system first predicts proxy structure on top of a geometric shell, then instantiates assets with placement logic and hard constraints. The authors also train a face-edge bigraph encoder on a building dataset annotated for MAPP, so the model can reason about the semantic roles of faces and edges rather than treating the shell as a flat surface.

For developers, the interesting part is the workflow shift: text prompts and parsed attributes are used to drive high-precision retrieval and assembly, while the structure stays post-editable. The paper says this reduces common generative failures such as over-smoothing, component collisions, and structural corruption, which are exactly the kinds of issues that make procedural or AI-assisted environment generation hard to ship.

The broader implication is that building generation may be moving toward hybrid systems that combine learned inference, procedural rules, and hard spatial constraints. That matters for teams building cityscapes, modular architecture kits, VR scenes, or digital twin content where editability and structural correctness matter as much as visual plausibility.