Introduction
Generating realistic 3D trees is a notorious challenge in computer graphics. Traditional procedural methods, such as L-systems, require expert-level parameter tuning and often struggle to adapt to diverse environmental constraints. Conversely, modern deep learning models typically produce voxels or meshes, which are computationally expensive; a detailed tree mesh can contain hundreds of thousands of faces, making it nearly impossible for standard transformer architectures to process. Furthermore, maintaining "structural coherence"—ensuring that branches connect to their parents in a biologically plausible manner—remains a significant hurdle for most generative models.
To address these issues, researchers have developed HourglassTree, a transformer-based model that treats 3D trees as sequences of branches rather than complex meshes. This approach allows for the generation of static trees, the completion of partial structures, and the simulation of 4D growth processes with high structural integrity.
Methods
The HourglassTree framework relies on three core innovations to make tree generation both tractable and accurate.
Branch-Based Representation
Instead of generating a dense mesh, the model generates a set of branches. Each branch is defined by two endpoints $(x, y, z)$ and their associated radii $(r)$. This representation provides massive compression; a complex tree with 200 branches can be described by just 1,600 parameters, compared to the millions required for a raw mesh.
Hourglass Transformer Architecture
Inspired by the U-Net structures common in computer vision, the "Hourglass" architecture utilizes downsampling layers to reduce the number of tokens in the middle of the network. This allows the model to capture global tree structure at the "bottleneck" while maintaining local detail at the input and output stages. This design significantly reduces GPU memory usage—from 15.9G in a standard baseline to 6.1G in the full Hourglass model—while increasing processing speed.
Botanical Token Ordering
While most 3D models sort points by spatial coordinates (e.g., Z-Y-X or Hilbert curves), this paper demonstrates that Depth-First Search (DFS) and Breadth-First Search (BFS) are superior for tree generation. By following the tree’s natural branching graph, the model learns hierarchical parent-child relationships more effectively. This "botanical" ordering provides a strong inductive bias, resulting in nearly 100% connectivity in generated structures.
Results and Applications
The model was trained on a synthetic dataset of six distinct species (Elm, Vitellaria, Hickory, Shadbush, Spruce, and Tulip), scaling up to 50,000 samples for conditional tasks.
Key results include:
- Fidelity: On Elm trees, the DFS-ordered model achieved an FID score of 5.64, a dramatic improvement over the 36.30 score achieved with standard spatial ordering.
- Connectivity: The DFS model reached a connectivity score of 0.9866, ensuring that almost every generated branch is properly attached to the skeleton.
- Versatility: The model successfully performs conditional tasks such as "point-cloud-to-tree" and "sketch-to-tree" conversion, maintaining user intent while filling in plausible branching details.
- 4D Growth: By modeling growth as a sequence of developmental stages, the transformer learns the "delta" between ages, effectively acting as a 4D growth simulator.
Conclusion
HourglassTree represents a significant step forward in automated botanical modeling. By prioritizing topological structure over raw spatial coordinates and utilizing an efficient multi-resolution architecture, it achieves high-fidelity results across a variety of species and tasks. Future work will focus on scaling the model to handle larger structures like entire forests and integrating leaf placement directly into the generative process.
Summary updated