Jiyu Wang
Designing residential landscape morphology means negotiating simultaneous trade-offs among microclimate comfort, visual quality, functional accessibility, and ecological service provision—objectives that experience-driven workflows rarely balance in any systematic way. We propose a hybrid generative framework that couples a variational autoencoder (VAE) with a generative adversarial network (GAN) to produce diverse yet structurally coherent morphologies conditioned on site boundary, building layout, and density constraints. Each generated scheme is assessed in roughly 6.5 s by a four-dimensional performance pipeline that pairs a microclimate surrogate model, a CNN-based aesthetic scorer, graph-theoretic accessibility metrics, and ecological service estimators. NSGA-III then searches the 128-dimensional latent space for Pareto-optimal configurations, and TOPSIS-based decision support singles out balanced recommended schemes. Across three residential parcels in Changsha, China—spanning high-, mid-, and low-density typologies—the VAE-GAN hybrid attained lower Fréchet Inception Distance (24.6–31.4) and higher morphological diversity (LPIPS 0.412–0.461) than standalone GAN or VAE baselines. On average the optimized schemes scored 6.2% above professional human-designed benchmarks on the composite performance index, with the widest margin on the most spatially constrained site; because all three parcels share one subtropical humid climate, we frame this as a model-based advantage rather than validated cross-climatic performance. Ablation and sensitivity analyses confirm that each architectural component is necessary and that composite rankings stay robust under weight perturbation. Rather than replacing professional judgment, the framework positions generative intelligence as a divergent-thinking accelerator for early-stage landscape design.