From floorplans to user interfaces: generative AI for layout design

Automated layout design aims to generate spatial arrangements of elements within a given space that are both visually appealing and functionally coherent. Traditionally reliant on expert knowledge and manual effort, layout design is now increasingly augmented by generative models capable of producing high-quality designs across domains such as architectural floorplans, user interfaces, and graphic compositions. Despite recent advances, existing generative approaches face three fundamental challenges: enforcing fine-grained geometric constraints, integrating heterogeneous functional and modality-specific requirements, and maintaining computational efficiency at scale.

This thesis addresses these challenges through a series of novel methods grounded in three core principles: vector-first generation, vector-raster co-generation, and end-to-end synthesis. Departing from raster-dominated workflows, we develop generative models that operate directly in vector space, enabling precise geometric reasoning and direct constraint enforcement without reliance on post-processing. To bridge structural rigor with visual richness, we propose dual-domain models that co-generate vector and raster outputs, accommodating the diverse needs of content and design tasks. Furthermore, by designing fully end-to-end architectures, we streamline the generation pipeline, eliminating intermediate stages and improving scalability.

The proposed models demonstrate strong performance across various layout tasks, including architectural reconstruction, UI generation, and visual content composition. By integrating domain-specific insights with scalable generative techniques, this thesis advances the field of automatic layout design and lays the groundwork for practical, high-quality design automation in real-world applications.