3DBonsai: Structure-Aware Bonsai Modeling Using Conditioned 3D Gaussian Splatting

Hao Wu, Hao Wang, Ruochong Li, Xuran Ma, Hui XiongThe Hong Kong University of Science and Technology (Guangzhou), China · 2025

Read 3DBonsai: Structure-Aware Bonsai Modeling Using Conditioned 3D Gaussian Splatting by Hao Wu, Hao Wang, Ruochong Li, Xuran Ma, Hui XiongThe Hong Kong University of Science and Technology (Guangzhou), China (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-01619.txt.

Introduction

Hao Wu and co-authors present 3DBonsai in 2025, a system that grows detailed 3D bonsai trees from text prompts. Text-to-3D tools have improved fast by mixing learned 3D priors with image diffusion models. Thin, branching structures like bonsai defeat those priors, which carry only simple shape detail.

The authors first built a trainable growth algorithm that colonizes space the way real branches do, giving the bonsai a believable skeleton. A conditioned splatting method then dresses that skeleton with foliage and bark detail. The split between structure and surface keeps trees both intricate and coherent. Graphics researchers get a template for other fine-structured objects that current generators blur or break.

Catalog record

CREATORHao Wu, Hao Wang, Ruochong Li, Xuran Ma, Hui XiongThe Hong Kong University of Science and Technology (Guangzhou), China
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE28 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-01619

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

3DBonsai: Structure-Aware Bonsai Modeling Using Conditioned 3D Gaussian Splatting Hao Wu, Hao Wang, Ruochong Li, Xuran Ma, Hui XiongThe Hong Kong University of Science and Technology (Guangzhou), China Abstract Recent advancements in text-to-3D generation have shown remarkable results by leveraging 3D priors in combination with 2D diffusion. However, previous methods utilize 3D priors that lack detailed and complex structural information, limiting them to generating simple objects and presenting challenges for creating intricate structures such as bonsai. In this paper, we propose 3DBonsai, a novel text-to-3D framework for generating 3D bonsai with complex structures. Technically, we first design a trainable 3D space colonization algorithm to produce bonsai structures, which are then enhanced through random sampling and point cloud augmentation to serve as the 3D Gaussian priors. We introduce two bonsai generation pipelines with distinct structural levels: fine structure conditioned generation, which initializes 3D Gaussians using a 3D structure prior to produce detailed and complex bonsai, and coarse structure conditioned generation, which employs a multi-view structure consistency module to align 2D and 3D structures. Moreover, we have compiled a unified 2D and 3D Chinese-style bonsai dataset. Our experimental results demonstrate that 3DBonsai significantly outperforms existing methods, providing a new benchmark for structure-aware 3D bonsai generation. Demos and more details are available at https://3dbonsai.github.io/ . 0 0 footnotetext: This work was supported in part by the National Natural Science Foundation of China(Grant No.92370204), in part by the National Key R&D Program of China(Grant No.2023YFF0725001), in part by Guangzhou-HKUST(GZ) Joint Funding Program(Grant No.2023A03J0008), in part by the Education Bureau of Guangzhou Municipality. I Introduction 3D assets are significant in various industrial sectors, such as game development, film production, metaverse, etc. Creating 3D assets is challenging since it demands highly professional and extensive design experience. Recently, text-to-3D object generation methods [ 1 , 2 , 3 , 4 , 5 ] have shown notable capabilities in producing 3D assets, including the common objects in Objaverse dataset [ 6 ] . However, relevant studies on generating 3D objects with complex and irregular structures are still rare. This paper investigates an open research problem of generating 3D bonsai. Bonsai refers to the traditional Asian art of growing and shaping miniature trees in containers. This is a challenging yet valuable task, due to the structure complexity and aesthetic standard of target 3D bonsai. Previous methods for generating 3D complex objects can be divided into two principal categories. The first involves the optimization-based 2D lifting methods [ 7 , 4 ] . They introduce Score Distillation Sampling (SDS) [ 7 ] to guide the synthesis of 3D objects through 2D diffusion. This approach mitigates the lack of 3D or multi-view data for training 3D models. However, SDS loss can lead to oversaturated and oversmoothed results. This limitation makes it difficult for the model to effectively accommodate objects with complex and irregular structures, such as bonsai. The second category of method adopts procedural generation [ 8 , 9 ] . These methods leverage programming capabilities in 3D modeling software, such as Blender, to generate the 3D assets automatically. To be specific, Infinigen [ 8 ] attempted to use procedural code to generate trees but was limited by the defined tree species from blender and generation algorithms, resulting in somewhat monotonous outcomes. 3DGPT [ 9 ] attempted to combine Large Language Models (LLMs) with Blender for the design of 3D assets, yet still encountered issues with fidelity and artistic style transfer in the generation of complex objects. To solve these problems in complex 3D object generation, we propose the 3DBonsai framework, designed to generate complex 3D bonsai effectively and efficiently. Technically, our approach combines 3D structure priors with structure-aware 3D Gaussian splatting. Firstly, we adapt the space colonization algorithm (SCA) [ 10 ] to 3D space and generate initial 3D structure priors. To this end, we propose a novel 3D Space Colonization Algorithm (3D SCA), which is implemented through the generation of leaf nodes in 3D space and a trainable branch growth algorithm. Then, these structure priors generated by 3D SCA are further processed through two different structure-aware 3DGS. The first directly utilizes the 3D structure prior as an initial state for 3DGS, followed by a strong constraint applied to the bonsai structure generation. Another approach employs a weaker constraint strategy, utilizing the 3D diffusion [ 11 , 12 ] to produce coarse 3D prior. Then the multi-view structure consistency learning is utilized to extract depth map information from the 3D structures across multiple views [ 13 , 14 ] . This process guides the splatting process by enhancing the complexity of 3D structures and improving the structure consistency across different viewpoints. The above two pipelines transform the 3D structure prior to an initialized 3D Gaussian distribution, which is then refined using the 2D diffusion model. We have compiled extensive datasets to support and validate our proposed method. We collect and release a dataset of Chinese-style 2D bonsai images and a smaller collection of high-quality 3D bonsai assets, which serve as the benchmark datasets and help with the bonsai generation quality assessment. Our proposed 3D bonsai generation framework demonstrates significant effectiveness in generating bonsai from complex structures. Our contributions are summarized as follows: • We propose a trainable 3D space colonization algorithm that serves as the structure guidance for structure-aware 3D Gaussian splatting, ensuring high fidelity of the generated bonsai while maintaining 3D consistency. • Our framework introduces two pipelines for structure constraints within 3DGS: fine structure and coarse structure. We employ a 3D structure prior to detailed fine structure 3DGS and multi-view structure consistency learning module to achieve coarse structure 3DGS. • We present a comprehensive bonsai dataset featuring 20 high-quality Chinese-style bonsai 3D models and over a thousand high-quality images. II Related Work II-A 3D Gaussian Splatting Recently, the 3D Gaussian splatting (3DGS) [ 15 ] method has optimized most of the issues associated with the methods above through its efficient rendering approach and superior rendering quality. It stands out as an impressive and practical 3D representation method. Not only does it facilitate real-time rendering, but it can also be conveniently integrated into 3D asset creation pipelines. Several current methods utilize 3D Gaussian splatting to generate high-quality 3D assets. DreamGaussian [ 4 ] employs a 3D Gaussian splatting framework using a single image to produce satisfactory outcomes. However, it faces common limitations like multi-face anomalies. Expanding on 3DGS, GaussianDreamer [ 5 ] and GSGEN [ 16 ] address the text-to-3D task. They both leverage 3D priors for 3D consistency and use a 2D diffusion model for high-resolution detail refinement. Their approaches significantly boost the optimization speed and generation quality of the text-to-3D models. II-B 3D Trees Models Generation The tree generation algorithms typically employ fractals and repetitive patterns, environmentally sensitive automata, and particle systems for the generation pipeline. L-systems [ 17 ] , grounded in fractal principles, utilize rule-based procedural generation to address the creation of various plant types. However, the tree models generated by L-systems still lack realism and artistic quality. Traditional sketch-based methods and data-driven methods are capable of accurately generating images or point clouds of plants based on user input. Currently, neural network-based methods for generating detailed tree images have achieved impressive results [ 18 ]

Read the complete text