ChatHouseDiffusion:提示词引导的建筑平面图生成与编辑方法

ChatHouseDiffusion: Prompt-guided generation and editing of floor plans

关键词:  建筑平面图生成  建筑平面图编辑  交互设计  扩散模型  大语言模型  Graphormer

Key words:  floor plan generation  floor plan editing  interaction design  diffusion model  large language model  Graphormer

摘要:

建筑平面图的生成与编辑是建筑智能设计中的关键环节,需要兼具灵活性与设计效率。针对现有方法过度依赖信息输入、缺乏交互性以及难以实现局部精确编辑等问题,文中提出ChatHouseDiffusion模型。该方法首先通过大语言模型将用户自然语言输入解析为结构化JSON提示信息,继而利用Graphormer捕捉房间之间的拓扑关系,最后通过扩散模型在房间轮廓约束下生成满足条件的建筑平面图。研究结果表明:在编辑阶段,借助交叉注意力图的替换机制可以实现对局部区域的精确修改,避免整体重构。 基于RPLAN数据集的测试表明,ChatHouseDiffusion在微平均交并比（Micro-IoU）和宏平均交并比（Macro-IoU）两项评价指标上均优于现有模型,尤其在使用准确输入条件时,其生成结果与真实标注高度一致,展现出良好的实用性与泛化能力。该模型不仅严格遵循用户要求,更能通过交互功能实现更直观的设计流程,为智能化建筑平面图设计提供了新的技术路径。同时基于该方法开发了一个支持绘图、文本输入与交互式编辑的可视化设计平台,提升了建筑平面图设计的实用性与可操作性。

Abstract:

The generation and editing of floor plans are critical in intelligent architectural planning, requiring a high degree of flexibility and efficiency. In response to the limitations of existing methods—such as excessive reliance on input information, lack of interactivity, and insufficient support for precise local modifications—this paper presents ChatHouseDiffusion. The proposed approach first utilizes a large language model to parse natural language instructions into structured JSON prompts, then applies Graphormer to capture the topological relationships between rooms, and finally adopts a diffusion model to generate floor plans under room boundary constraints. The results indicate that during the editing phase, the cross-attention maps replacement mechanism enables localized modifications without reconstructing the entire layout. Experiments conducted on the RPLAN dataset demonstrate that ChatHouseDiffusion outperforms existing models in both Micro-IoU and Macro-IoU, especially when accurate input prompts are used, achieving results highly consistent with ground truth, which exhibits strong practical utility and generalization performance. The model not only strictly follows user requirements but also enables intuitive and iterative design through interactive operations, providing a novel pathway for intelligent floor plan design. Based on this method, we further developed a visual platform enabling the drawing of outlines, inputting text prompts, and generating and editing floor plans, enhancing the practicality and usability of floor plan design.