A Framework for Architectural Geometry Synthesis Using Cellular Automata and Conditioned Diffusion Models

We present a framework for generating 3D architectural geometry and corresponding high-fidelity renderings from procedural rules. Our pipeline integrates a 3D Cellular Automata (CA) engine for voxel massing, a geometry-conditioned diffusion model (SDXL with ControlNet) for structure-aware rendering, and a Large Language Model (LLM) for automated metadata generation. This system links procedural logic with visual design, enabling rapid exploration and the creation of structured datasets. We demonstrate the framework by generating over 15,000 unique configurations at 70 seconds per model.