Designing restorative outdoor environments using generative adversarial networks

Designing restorative outdoor environment (ROE) layouts is time-consuming, labor-intensive, and tends to lack variety, yet Generative Adversarial Networks (GANs) offer a promising shortcut. However, controlled GAN-based ROE generation based on design semantics and early-stage systematic evaluation has been missing. Unlike previous studies that focused on visual entity generation, this study introduced semantically-labeled datasets to enable functional control over design elements such as walkability and therapeutic spaces. It aims to apply GANs in ROE layout design, determine the applicable characteristics, and evaluate the generated results through qualitative and quantitative assessments. Three stylized datasets—“natural,” “semi-artificial,” and “artificial”— were used to train 12 models. The results show that Pix2Pix produces the best results and CycleGAN the worst. In the application stage, GauGAN performs the lowest practical creativity, though it produces the highest quality output in the test stage. Models trained on the “artificial” dataset show the most obvious main features in the test stage, and those trained on the “natural” dataset display the most prominent style features in both the test and application stages. The “semi-artificial” images are mislabeled but inspire the models to produce better outputs. The quantitative, qualitative, and application assessment progress progressively cross-validate GAN-generated ROE layouts and works. The study equips urban designers with valuable insights and practical experience in implementing GANs in the ROE layout design. • Applied generative adversarial networks (GANs) to the design of restorative outdoor environment (ROE) layouts. • Created three stylized datasets of ROE layouts with design intentions and interpretability. • Established a systematic evaluation framework for ROE design using GANs. • Ascertained the applicable characteristics of different GANs in ROE layout design.