HypergraphFormer: LLM-Based Floor Plan Generation

HypergraphFormer is a novel approach to floor plan generation that uses a large language model (LLM) to learn hypergraph representations. The model is trained via supervised fine-tuning on the RPLAN dataset and generates a hypergraph-based textual representation encoding spatial relationships and connectivity. It outperforms state-of-the-art techniques using rasterized or vectorized representations across diverse metrics. The method also shows improved data efficiency, especially under distribution shift. By decoupling apartment footprints from functional and geometric subdivisions, it enables generation for arbitrary, irregular user-specified boundaries. The paper releases a separate out-of-distribution dataset to demonstrate generalizability.