Building-CVGAE: Constrained Variational Graph Autoencoder For Architectural Volumetric Design Generation
Year: 2025
Type: conference-paper
Abstract: Multiple-floorplan generation is a challenging and underexplored research area. While single-floorplan graph generation has been widely studied, extending these methods to multiple floors remains complex due to the hierarchical and structured nature of multi-floor voxel representations and the need for continuity between floors, where elevators and stairs must be aligned. In this work, we propose Building-CVGAE, a Constrained Variational Graph Autoencoder (CVGAE) for voxel-based multiple-floorplan generation, where the voxel graph structure remains fixed during generation, and only the voxel node labels are predicted. Our approach follows a two-stage training strategy: first, a Self-Supervised Learning (SSL) model generates embeddings of program graphs, capturing high-level structural information. These embeddings are then used to guide the CVGAE, which predicts voxel labels while preserving spatial and architectural coherence. We investigate the impact of conditioning our CVGAE using embeddings with different sizes and evaluate various architectures for voxel graph generation. The model is evaluated on the Building-GAN dataset using two metrics: Floor Area Ratio (FAR) distance and Target Program Ratio (TPR) similarity. Our model achieves strong results in key metrics and generates high-quality visuals, showing clear improvements over existing methods. The code will be made publicly available at https://github.com/Sehaba95/Building-CVGAE.
Language: English
Subfield: Architecture
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416213608
FWCI:
Cites:
Cited by:
Related to: