Building-Diffusion: Graph Discrete Diffusion Model For Architectural Volumetric Design Generation
Résumé
Diffusion models have recently achieved remarkable suc- cess in continuous domains such as images and videos, yet applying them to discrete and structured data like graphs remains challenging. We propose Building-Diffusion, a conditional discrete diffusion model for node-attribute gen- eration in architectural building layout graphs. The method follows a two-stage pipeline: first, a self-supervised model encodes the input program graph (Gp) into a compact em- bedding; then, a discrete denoising network learns to it- eratively recover clean voxel labels from categorical noise while keeping the graph structure fixed. This design allows the model to capture categorical node distributions and pre- serve topological consistency. Experiments on the Building- GAN dataset demonstrate that Building-Diffusion produces layouts that are both realistic and structurally coherent. We further validate our approach through a user study and quantitative evaluations using three metrics. These results highlight discrete diffusion as a promising framework for graph-based generative modeling.