Generating Landscape Layouts with GANs and Diffusion Models

Artificial intelligence (AI), particularly machine learning (ML), is dramatically expanding in design, in combination with computer vision, image recognition, segmentation methods, and satellite imagery, for the use of AI in two-dimensional orthogonal imagery. This study utilizes Generative Adversarial Networks (GANs) and AI Diffusion Models (DM) to generate novel garden plans. We utilized a large dataset of two-dimensional garden plans obtained from an online design repository to train a deep learning (DL) system on a) qualitative aspects, such as aesthetic scores, and b) quantitative attributes, such as functional scores. We used FastGAN, an existing DL architecture within a PyCharm environment, to generate rough landscape layouts, and then utilized the Stable Diffusion Model (SDM) to provide them with higher definition and resolution. The system is trained on a custom dataset consisting of projects evaluated by a large number of people. It learns and generates patterns, ratios, and relationships related to land use. Hence, the outputs are high-quality landscape layouts suitable for smaller-scale home garden projects. Our goal is to demonstrate an AI-based workflow that can assist landscape architects in their explorations for smaller-scale landscape design projects. Finally, the outcomes were assessed across 100 AI-generated plans based on eight criteria: graphic language, plan readability, building mass, land-use patterns, circulation, softscape pattern, diversity, and readability. The successful and negative aspects of the study, which scored above 72 percent in all criteria, were identified and discussed.