FloorPlan-LLaMa: Aligning Architects' Feedback and Domain Knowledge in Architectural Floor Plan Generation

Floor plans serve as a graphical language through which architects sketch and communicate their design ideas. Actually, in the Architecture, Engineering, and Construction (AEC) design stages, generating floor plans is a complex task requiring domain expertise and alignment with user requirements. However, existing evaluation methods for floor plan generation rely mainly on statistical metrics like FID, GED, and PSNR, which often fail to evaluate using domain knowledge. As a result, even high-performing models on these metrics struggle to generate viable floor plans in practice. To address this, (1) we propose ArchiMet-ricsNet, the first floor plan dataset that includes functionality, flow, and overall evaluation scores, along with detailed textual analyses. We train FloorPlan-MPS (Multi-dimensional Preference Score) on it. (2) We develop FP-LLaMa, a floor plan generation model based on an autoregressive framework. To integrate architects' professional expertise and preferences, FloorPlan-MPS serves as the reward model during the RLHF (Reinforcement Learning from Human Feedback) process, thereby aligning FP-LLaMa with the needs of community. (3) Comparative experiments demonstrate that our method outperforms baseline models in both text-conditional and class-conditional tasks. Validation by professional architects confirms that our approach yields more rational plans and aligns better with their preferences.