Text‑to‑Garden is a multimodal generative design system that reconstructs traditional Chinese garden landscapes from simple text prompts. Instead of manual surveys or expert‑driven scans, it can interpret instructions like “a moon gate beside a lotus pond” and synthesize realistic layout proposals with deep learning models. This method streamlines restoration workflows and opens up intuitive, language‑based heritage design.

Leveraging a discrete variational autoencoder, an autoregressive transformer, and CLIP for evaluation, the project marries cultural preservation with state‑of‑the‑art generative AI. It not only lowers technical barriers but also empowers human‑centered, project‑specific restoration. By enabling participatory resilience planning through computational methods, Text‑to‑Garden charts a new path for spatial regeneration and adaptive heritage conservation.