Objective This research addresses the persistent challenges of controllability and interpretability in generative algorithms for landscape architecture, aiming to bridge the gap between theoretical model development and practical design application. Current end-to-end generative frameworks often lack semantic transparency and controllable intermediate mechanisms, limiting their adaptability to complex, real-world design contexts. To overcome these limitations, this research proposes a multi-model, multi-path generative framework that integrates semantic and process-level interpretability with designer interaction. The framework is designed to enhance both the rationality and professional logic of generative outcomes, providing a controllable and transparent computational pathway for intelligent park and open space design.

Methods The proposed framework is built upon a “functional − semantic − spatial” integrated node classification system, in which each landscape node is abstracted into a multi-attribute characteristic vector. These node representations function as the semantic mediation layer throughout the generative process, supporting the implementation of the following three core tasks. Further, a nodal functional relationship refers to the structured spatial and functional connections among nodes within a landscape system, describing how individual functional units (e.g., plazas, paths, vegetated zones, and waterfronts) interact, overlap, or depend on each other within a coherent spatial organization, thereby capturing both the topological (positional and connectivity) and semantic (functional purpose and hierarchy) dimensions of landscape composition, and serving as the mediating mechanism that bridges abstract design intent and concrete spatial generation. 1) Inference of nodal functional relationship: Image-to-structure translation is adopted to predict spatial and functional linkages among landscape elements. 2) Generation of complete layout scheme: Nodal functional relationships are translated into coherent site layouts under varying spatial and functional constraints. 3) Terrain-aware adaptive generation: Elevation and slope data are integrated to enable topographically responsive design outcomes. To ensure flexibility across different design needs, the framework supports two application paths, namely the rapid generation path and the directed generation path, which are detailed as follows. 1) The rapid generation path, automatically infers nodal functional relationships via Model A and generates complete layouts through Models B-1 and B-2, requiring only minimal designer adjustments. This path is suitable for conceptual and early-stage design, where efficiency and iterative exploration are prioritized. 2) The directed generation path entails designers to manually define or modify nodal relationships based on site conditions, design intent, or functional strategies. This path allows targeted intervention and stronger alignment with specific planning objectives, supporting semi-automatic, user-guided generation. Depending on the inclusion of topographic data, two complementary sub-models are employed: Model B-1 handles general sites without terrain information using a CycleGAN-based end-to-end architecture, while Model B-2 integrates topographic characteristics into a multi-channel input (B for spatial structure, G for semantic function, and R for terrain information), enhancing adaptive learning for complex terrains. A multi-dimensional evaluation system is incorporated to assess and optimize generative outcomes. The evaluation framework encompasses indicators of structural coherence, spatial configuration, land use ratio, and road attributes, enabling both quantitative comparison with existing reference plans and independent evaluation under unconstrained conditions. In the latter case, a norm-constrained multi-objective optimization process is applied to ensure design practicality, guiding the generative system beyond visual similarity toward functional and implementable outcomes.

Results Empirical validation is conducted using two representative cases — Beixiaohe Park in Beijing and Fanchuan Park in Xi’an — demonstrating the framework’s robustness and adaptability across site typologies and environmental constraints. In the Beixiaohe Park case, where terrain variation is minimal, the system effectively produces spatially coherent and functionally reasonable layouts through the rapid generation path, achieving fast scheme generation and iterative refinement with minimal manual intervention. The integration of the evaluation framework enables automatic selection of optimized results based on multi-dimensional indicators, verifying the system’s operational flexibility and practicality in early-stage design contexts. In contrast, in the Fanchuan Park case characterized by complex terrain and larger spatial heterogeneity, the advantages of the directed generation path and the terrain-aware Model B-2 are highlighted. By incorporating multi-channel inputs that encoded spatial, functional, and topographic data, the system significantly improves road network continuity, spatial organization, and topographic adaptability, producing results closely aligned with expert-designed schemes. Quantitative evaluation confirms that the system boasts higher structural coherence and spatial rationality compared with baseline generative approaches, while maintaining a balance between diversity and functionality. The integrated multi-dimensional evaluation system proves applicable under both comparative (compared with real reference plans) and non-comparative conditions, offering an objective, transparent mechanism for solution screening and optimization. This capability effectively compensates for the prevailing “similarity-oriented but less practical” tendency in current generative design research, demonstrating the framework’s value as a decision-support tool for real-world landscape planning.

Conclusion This research establishes a multi-model, multi-path generative framework incorporating a semantic mediation mechanism providing a systematic and process-level interpretable technical route for intelligent spatial layout in complex landscape sites. By integrating node semantics, terrain constraints, and quantitative evaluation into a unified workflow, the framework advances both the theoretical and applied dimensions of generative design in landscape architecture. The research contributes three major insights: 1) Introducing nodal functional relationships as mediating mechanism effectively decomposes the opaque end-to-end generative process, enhancing structural clarity and interpretability; 2) the dual-path strategy allows seamless transition between rapid automated generation and designer-directed customization, achieving a practical balance between efficiency and control; 3) the multi-dimensional evaluation system establishes a standardized, data-driven basis for assessing design rationality and spatial performance, promoting the shift of generative design from similarity-based learning to utility-oriented application. Future work will focus on developing goal & function−driven dynamic tuning mechanisms and multi-source data integration to strengthen model generalization and applicability. Efforts will also be made to explore the coupling of behavioral, functional, and spatial dimensions, as well as real-time human − AI co-design interfaces, so as to further enhance collaboration, adaptability, and practical impact in intelligent landscape and urban design.