Diff-Tree: A Diffusion Model for Diversified Tree Point Cloud Generation with High Realism

Three-dimensional (3D) virtual trees play a vital role in modern forestry research, enabling the visualization of forest structures and supporting diverse simulations, including radiation transfer, climate change impacts, and dynamic forest management. Current virtual tree modeling primarily relies on 3D point cloud reconstruction from field survey data, and this approach faces significant challenges in scalability and structural diversity representation, limiting its broader applications in ecological modeling of forests. To address these limitations, we propose Diff-Tree, a novel diffusion model-based framework for generating diverse and realistic tree point cloud with reduced dependence on real-world data. The framework incorporates an innovative tree realism-aware filtering mechanism to ensure the authenticity of generated data while maintaining structural diversity. We validated Diff-Tree using two distinct datasets: one comprising five tree species from different families and genera, and another containing five Eucalyptus species from the same genus, demonstrating the method’s versatility across varying taxonomic levels. Quantitative evaluation shows that Diff-Tree successfully generates realistic tree point cloud while effectively enhancing structural diversity, achieving average MMDCD and COVCD values of (0.41, 65.78) and (0.56, 47.09) for the two datasets, respectively. The proposed method not only significantly reduces data acquisition costs but also provides a flexible, data-driven approach for virtual forest generation that adapts to diverse research requirements, offering a more efficient and practical solution for forestry research and ecological modeling.