【目的】针对风景园林平面设计中多要素耦合约束复杂、方案迭代效率较低的问题,构建能够感知场地环境多维数据(如场地实景、交通流线、空间要素类型及分布),并支持风景园林平面设计三阶段流程的自动化生成模型。【方法】构建“初步设计—方案修改—总平面设计”三阶段风景园林平面设计框架,研发与之匹配的多维约束扩散生成模型(multi-dimensional constrained diffusion model,MCDM);通过扩展模型的输入输出通道,对红绿蓝三基色(RGB)场地实景、灰度交通流线和类别栅格等多维数据进行联合学习,实现多维空间特征融合与可控生成。【结果】在“初步设计”阶段,模型可依据输入的场地现状实景、交通流线、空间要素类型及分布等数据,自动生成具有视觉真实感的场地实景方案及与之对应的空间布局方案;在“方案修改”阶段,模型以人工优化后的空间要素布局为附加约束,对上一阶段设计方案进行定向优化生成,实现方案的高效迭代;在“总平面设计”阶段,模型以上一阶段优化方案及相关空间要素数据为多维约束条件,批量生成具有专业手绘风格的总平面方案图。【结论】MCDM能够提升风景园林方案的生成效率与方案多样性,为人机协同设计提供可行的技术范式。

【Objective】In order to address the complex coupled constraints among multiple design elements and the relatively low efficiency of scheme iteration in landscape architectural plan design, this study developed an automated generative model capable of perceiving multidimensional site-environment data, such as existing-site imagery, circulation routes, and the types and spatial distributions of site elements, while supporting a three-stage workflow for landscape architectural plan design. 【Method】A three-stage framework comprising “preliminary design-scheme modification-master plan design” was established, and a corresponding multidimensional constrained diffusion model (MCDM) was developed. The input and output channels of the model were expanded to enable joint learning from multidimensional data, including RGB site imagery, grayscale circulation maps, and categorical raster data, thereby achieving multidimensional spatial feature fusion and controllable generation. 【Result】In the “preliminary design” stage, the model automatically generated visually realistic site design renderings and corresponding spatial layout schemes based on input data, including existing-site imagery, circulation routes, and the types and spatial distributions of site elements. In the “scheme modification” stage, manually refined spatial-element layouts were introduced as additional constraints to guide the targeted optimization and regeneration of schemes produced in the previous stage, thereby enabling efficient scheme iteration. In the “master plan design” stage, the optimized scheme from the previous stage and the associated spatial-element data were used as multidimensional constraints to generate batches of master plan drawings in a professional hand-drawn style. 【Conclusion】The MCDM can improve the generation efficiency and diversity of landscape architecture schemes, providing a viable technical paradigm for human-machine collaborative design.