GardenDesigner: Encoding Aesthetic Principles into Jiangnan Garden Construction via a Chain of Agents

Abstract

Jiangnan gardens, a prominent style of Chinese classical gardens, hold great potential as digital assets for film and game production and digital tourism. However, manual modeling of Jiangnan gardens heavily relies on expert experience for layout design and asset creation, making the process time-consuming. To address this gap, we propose GardenDesigner, a novel framework that encodes aesthetic principles for Jiangnan garden construction and integrates a chain of agents based on procedural modeling. The water-centric terrain and explorative pathway rules are applied by terrain distribution and road generation agents. Selection and spatial layout of garden assets follow the aesthetic and cultural constraints. Consequently, we propose asset selection and layout optimization agents to select and arrange objects for each area in the garden. Additionally, we introduce GardenVerse for Jiangnan garden construction, including expert-annotated garden knowledge to enhance the asset arrangement process. To enable interaction and editing, we develop an interactive interface and tools in Unity, in which non-expert users can construct Jiangnan gardens via text input within one minute. Experiments and human evaluations demonstrate that GardenDesigner can generate diverse and aesthetically pleasing Jiangnan gardens. Project page is available at https://monad-cube.github.io/GardenDesigner.

Figures (17)

• Figure 1: The motivation of GardenDesigner. Traditional manual modeling of Jiangnan gardens requires document searching, data modeling, and expert design, making it time-consuming and expertise-dependent. GardenDesigner automates Jiangnan garden construction via analyzing the user text and acquiring the assets, and then optimizes the garden layout. For applications, users can construct a Jiangnan garden through text input, which can be used for creating VR/AR experiences, film and game development, and real garden construction.
• Figure 2: Overview of the GardenDesigner pipeline. GardenDesigner transforms the user input into a Jiangnan garden through Hierarchical Garden Composition and Knowledge-Embedded Asset Arrangement. First, Hierarchical Garden Composition transfers the user input into parameters for terrain and road generation with aesthetic principles. Subsequently, Knowledge-Embedded Asset Arrangement chooses the objects based on the garden knowledge and area information, and then optimization loss is used to get the feasible solution for layout.
• Figure 3: The five constraints categories: (a) Global, edge, and middle; (b) Position, around, and backed up; (c) Distance, near and far; (d) Alignment, aligned; And (e) Rotation, face to.
• Figure 4: GardenVerse construction from Internet repositories and manual modeling. We invite experts to modify the architectures and construct object combinations. Finally, garden experts annotate the basic information and garden knowledge for assets.
• Figure 5: Qualitative analysis. In (a), we input the same prompt to GardenDesigner and the baseline conlan to evaluate the generated garden quality with three different views for each garden. In (b), we compare four methods: (1) Baseline conlan, (2) Baseline with GardenVerse assets, (3) GardenDesigner, and (4) GardenDesigner without Knowledge-Embedded Asset Arrangement to conduct the ablation experiment.