A deep learning framework for generating layout and style in interior design

The field of spatial design style generation faces challenges including lengthy generation and modification cycles and ambiguous design style terminology. This study combines a multimodal encoder with a U-Net-based dual-branch decoder to construct a framework for joint generation of style and layout. Compared to studies that consider only style or only layout design, model demonstrates superiority in optimizing style and layout consistency and rationality. Multimodal inputs are weighted using an attention mechanism to integrate user information, user constraints, and information from reference layout and style maps, constraining the model to generate results that meet user satisfaction. This study also incorporates an automated modification scheme based on user feedback, which can mitigate the inefficient interaction between designers and users. Experimental results shows, the SL-JGM achieves a high constraint satisfaction rate of 92% on the Structured3D dataset, compared to five baseline models, with a user satisfaction score of 4.6. It also demonstrates satisfactory stability in generating models for common styles such as Nordic and New Chinese. Furthermore, SL-JGM demonstrates significant advantages over other models in layout rationality and style consistency. This research contributes to the simplification of spatial layout design workflows.