TreeFlow: Conditional Flow Matching for 3D Tree Point Cloud Generation from Inventory Attributes

Model checkpoints, validation visualizations, and generated samples for the TreeFlow paper.

This archive contains all trained model weights, training-time validation visualizations, and generated point cloud samples used in:

Marcozzi, A.; Tenny, J.; Martin, D.; Castorena, J.; Crennen, Z.; Wells, L.; Hillman, S. TreeFlow: Conditional Flow Matching for 3D Tree Point Cloud Generation from Inventory Attributes. Remote Sensing, 2026. DOI: 10.3390/rs1010000

What is TreeFlow?

TreeFlow is a conditional flow matching model that generates realistic 3D individual-tree point clouds from three scalar inventory attributes: species, acquisition platform (TLS / MLS / ULS), and tree height. It is trained directly on real laser scanning data from the FOR-species20K benchmark (17,707 manually segmented trees spanning 33 species and 19 genera, captured across temperate, boreal, and Mediterranean biogeographic regions). The model uses a transformer backbone with U-ViT long skip connections, NeRF-style sinusoidal positional encoding, and token-prepend conditioning, and is trained with the Conditional Optimal Transport flow matching objective plus classifier-free guidance.

Contents.

Three experiment directories corresponding to a three-stage training curriculum at progressively higher point densities:- pretrain-8-512-4096/ — Stage 1. Trained from scratch on 4,096-point clouds for 2,000 epochs.- pretrain-8-512-8192/ — Stage 2. Initialized from Stage 1 and trained on 8,192-point clouds for 1,000 epochs.- finetune-8-512-16384/ — Stage 3. Final production model used for all results in the paper. Trained on 16,384-point clouds for 5,000 epochs.

Each directory contains the full training config (config.json), all saved checkpoints (checkpoints/epoch_*.pt), per-epoch validation visualizations, and training logs. The 4,096-point and 16,384-point directories additionally contain a samples/ subdirectory with the generated point clouds (zarr arrays with metadata), comparison images, a flat metadata index (samples_metadata.csv), and the complete morphological evaluation outputs (per-pair metrics, Wasserstein-1 baselines, and summary tables by genus and height bin) reported in the paper.

Reproducibility.

All training, sampling, post-processing, and evaluation code, together with the SLURM submission scripts used to produce this archive, is available at https://github.com/amarcozzi/TreeFlow. A top-level EXPERIMENTS_README.md in this archive documents the directory layout, hyperparameters, and the commands to regenerate every artifact.

Training data.

The FOR-species20K benchmark is distributed separately by Puliti et al. (2025), Methods in Ecology and Evolution 16, 801–818, https://doi.org/10.1111/2041-210X.14503. This archive does not redistribute it.

License. CC BY 4.0 — please cite the paper above if you use these materials.