Sectional Sampling

Abstract

This paper tackles the challenge of algorithmically generating 3D architectural designs. Traditional rule-based methods and GAN-based 3D generative models struggle with extrapolation, control, and resolution. We propose a novel approach using fine-tuned latent diffusion models, specifically Stable Diffusion, to generate high-resolution architectural sections. Our pipeline synthesizes 3D models through latent interpolation, employing Low-Rank Adaptation (LoRA) for model finetuning and ControlNet for consistent outlines. The process includes creating sectional poché datasets for various building types, fine-tuning models with LoRA, generating new poché images conditioned on section boundaries via ControlNet, and using latent interpolation for smooth transitions between sections. The resulting sections are reconstructed into 3D models using the Marching Cubes algorithm in the Grasshopper environment. Our method generates 3D designs with resolution up to 1024x768x300 voxels and consistent continuity on sectional axis, achieving state-of-the-art results in 3D architectural design generation.