Reproduce “BuildingBlock: A Hybrid Approach for Structured Building Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Junming Huang, Chi Wang, Letian Li, Changxin Huang, Qiang Dai, Weiwei Xu
Researchers from Zhejiang University developed BuildingBlock, a hybrid framework that generates diverse, structured, and editable 3D buildings directly from text descriptions. It integrates a Transformer-based diffusion model, Large Language Models, and procedural content generation to achieve state-of-the-art results in layout generation and high-quality 3D models, while also enabling localized editing.
Key insights
- A hybrid approach effectively combines the global synthesis capabilities of generative models, the semantic reasoning of LLMs, and the precise control of PCG to overcome individual limitations in complex 3D building generation.
- Representing building layouts as unordered sets of component boxes and processing them with a Transformer-based diffusion model proves effective for generating coherent layouts without sequence dependency.
- LLMs can successfully translate high-level text prompts into component-specific style attributes and enrich hierarchical information within rule-based layouts, bridging the semantic and geometric gaps.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.