NeuralFur

Fur reconstruction from multi-view images using 3D Gaussian Splatting.

Prerequisites

• CUDA 11.8 (download)
• Conda
• Blender 3.6 (optional, for strand visualization) (download)

Make sure PATH includes <CUDA_DIR>/bin and LD_LIBRARY_PATH includes <CUDA_DIR>/lib64.

Installation

cd submodules/GaussianHaircut
bash install.sh

This will clone external dependencies (pytorch3d, simple-knn, glm), create a conda environment, and install all required packages. NeuralHaircut is already included in ext/.

Data

Download the preprocessed data from Google Drive (link) and place it into your desired location, e.g.:

/path/to/data/panda_processed_GH2/walk/

pip install gdown
gdown https://drive.google.com/uc?id=1rIxvQKXVaMZ6Xzx7MAmHfucnbezLSbZt
unzip *.zip

The data directory should contain:

• sdf_grid.npy, min_bound.npy, max_bound.npy -- SDF volume for penetration loss
• neus_lr.obj -- reconstructed mesh (used for chamfer loss)
• furless.obj -- furless body mesh (used for strand initialization)
• tan_furless.npy -- precomputed tangent directions
• annotations_furless_reshaped2.json -- body part annotations (controls per-region fur length and gravity)
• eyes.ply -- eye landmarks (used for metric scale estimation)
• images/ -- input images
• masks/ -- segmentation masks
• orientations/ -- orientation maps

Configuration

Edit submodules/GaussianHaircut/simple_run_panda.sh and set your paths:

CUDA_HOME=/path/to/cuda-11.8       # your CUDA installation
DATA_PATH=/path/to/data/panda_processed_GH2/walk/
ENV_PATH=/path/to/conda/env        # conda environment name or path

DATA_PATH is passed to the training and export scripts via --data_root, which automatically replaces the DATA_ROOT placeholder in the YAML config. No need to edit the YAML config separately.

The YAML config (src/arguments/metrical_panda_furless_15k_small.yaml) contains animal-specific parameters like per-region fur length, gravity directions, and loss settings. These don't need to change between runs of the same animal.

Running

cd submodules/GaussianHaircut
bash simple_run_panda.sh

The pipeline runs two stages:

1. Fur strand reconstruction -- optimizes strand geometry and appearance using orientation, mask, chamfer, SDF, shape consistency, and gravity losses
2. Strand export -- exports the reconstructed strands as a .ply point cloud

Results are saved to the SAVE_EXP_PATH directory. Use Tensorboard to monitor training progress.

Project Structure

NeuralFur/
  submodules/
    GaussianHaircut/
      src/
        train_latent_fur.py          # main training script
        preprocessing/export_fur.py  # strand export
        scene/                       # scene, cameras, gaussian models
        gaussian_renderer/           # rendering
        utils/                       # losses, camera utils, etc.
        arguments/                   # config and CLI argument parsing
      ext/
        NeuralHaircut/               # strand prior, texture networks
        diff_gaussian_rasterization_hair/  # custom CUDA rasterizer
      simple_run_panda.sh            # example run script
      install.sh                     # installation script

TODO

• Code for processed scenes
• Reconstruction results
• Preprocessing pipeline for new data (deadline: April 12, 2026)
• Detailed guide for using obtained ply inside unreal engine

License

The code is distributed for research purposes under CC BY-NC-SA 4.0.

Citation

If you find this work useful, please consider citing:

@inproceedings{NeuralFur26,
  title = {{NeuralFur}: Animal Fur Reconstruction from Multi-view Images},
  aword_paper = {Best Paper Runner Up},
  booktitle = {Int.~Conf.~on 3D Vision (3DV)},
  month = mar,
  year = {2026},
  author = {Sklyarova, Vanessa and Kabadayi, Berna and Yiannakidis, Anastasios and Becherini, Giorgio and Black, Michael J. and Thies, Justus},
  month_numeric = {3}
}