A Cross-Stage Human-GenAI Collaborative Workflow for Landscape Architecture Design: A Pocket Park Feasibility Study

Abstract

Research on generative artificial intelligence (GenAI) in landscape architecture has largely focused on individual design stages. Few studies have examined how selected design information can transfer across design stages and how landscape architecture expertise can be incorporated into a workflow-level process. This study proposes a human–GenAI collaborative workflow that integrates design knowledge at each stage and facilitates information transfer across 2D plan generation, 3D spatial construction, and scene rendering. The workflow is implemented with a Stable Diffusion (v1.5) model fine-tuned with LoRA and conditioned by ControlNet, and comprises three stages. First, professional knowledge is encoded as semantic constraints, and the designer validates the results. Next, the designer translates the selected 2D plan into a 3D spatial model through designer-led spatial interpretation and modeling. Finally, the designer combines multi-condition rendering with professional knowledge and selects the final alternative. A pocket park case study demonstrates the feasibility of the proposed workflow. The generated alternatives were assessed against the stated design criteria in this case, and the results show the designer’s decisions at each stage and how those decisions informed the subsequent stage.