This study explores the use of generative graph models to create architectural layouts, focusing on building footprints as a foundational layer within a layered graph framework. We evaluate two prominent methods: the Deep Generative Model of Graphs (DGMG) with the Deep Graph Library (DGL) and GraphGPT, a transformer-based model adapted for graph generation. DGMG employs a graph-native, stepwise approach to add nodes and edges probabilistically, allowing it to generate architectural graphs that maintain structural integrity and spatial coherence. The model’s incremental design process makes it well-suited for generating closed-loop architectural footprints, though scalability and node feature representation pose challenges when applied to complex layouts. Conversely, GraphGPT transforms architectural graphs into token sequences, allowing it to leverage the transformer architecture for graph generation. While promising in its ability to encode both sequential and relational information, adapting GraphGPT to architectural data proved challenging due to tokenization and computational constraints. Despite these limitations, GraphGPT offers the potential for multimodal applications, where both text and graph data can inform generative processes. Our preliminary findings suggest that DGMG demonstrates feasibility in generating simple footprint graphs, while GraphGPT’s transformer approach may hold promises for complex, context-rich tasks if adapted effectively. This study contributes insights into the relative strengths and limitations of graph-native and transformer-based models for architectural design, setting a foundation for future research in generative graph applications in architecture.