GaussianFluent 3DGS Assets

This repository contains pretrained 3D Gaussian Splatting assets and simulation configuration files for:

GaussianFluent: Gaussian Simulation for Dynamic Scenes with Mixed Materials

• Project page: https://hb-pencil-zero.github.io/GaussianFluent/
• Paper: https://arxiv.org/abs/2601.09265
• Code: https://github.com/HB-pencil-zero/GaussianFluent

Contents

The model/ folder contains standard 3DGS checkpoints with:

model/<scene>/
  cameras.json
  cfg_args
  input.ply
  point_cloud/iteration_<latest>/point_cloud.ply

Released standard 3DGS scenes:

a752b28d-f
bowl
bullet_0_psnr36
cake
cookie
dragonfruit
garden
garden_ours
jelly
kiwi
kiwi_0.04_psnr42
lollipop
milk2
milk_0.03_psnr32
oreo
pineple
pumkin
sand_castle
toast
watermelon
watermelon_fruitninja

The model/trained_gs_fruitninja/ folder contains standalone FruitNinja assets stored as .ply + .json pairs:

apple
bread
cake
orange
pomegranate
watermelon

The config/ folder contains all 32 simulation configuration JSON files from the research codebase, including single-object and multi-object examples such as watermelon_config.json, jelly_config_nacc.json, fruits*.json, milk2_and_watermelon_config.json, and backup/variant configs.

asset_manifest.json records the local staging source, selected checkpoint iteration, and matched config for each released scene.

Usage

Clone the code repository with submodules:

git clone --recurse-submodules https://github.com/HB-pencil-zero/GaussianFluent.git
cd GaussianFluent

Download this Hugging Face repository and place the model/ and config/ folders at the root of the code repository. The resulting layout should look like:

GaussianFluent/
  model/a752b28d-f/...
  model/watermelon/...
  model/jelly/...
  model/garden/...
  model/trained_gs_fruitninja/apple.ply
  config/watermelon_config.json
  config/jelly_config_nacc.json
  ...

Example simulation commands:

python gs_simulation/watermelon/gs_simulation_watermelon.py \
  --model_path model/watermelon \
  --output_path output/watermelon \
  --config config/watermelon_config.json \
  --render_img \
  --compile_video

python gs_simulation/jelly/gs_simulation_jellynacc.py \
  --model_path model/jelly \
  --output_path output/jelly \
  --config config/jelly_config_nacc.json \
  --render_img \
  --compile_video

Interior Filling

The cleaned interior filling and back-projection code is in the GitHub repository under interior_filling/. The inpainting stage can use the external third-party MVInpainter project:

https://github.com/ewrfcas/MVInpainter

MVInpainter is not part of GaussianFluent and is not redistributed in this asset repository.

Notes

• Standard 3DGS scene folders keep only the selected latest checkpoint point_cloud.ply to reduce download size.
• trained_gs_fruitninja/ is included separately because its local layout is standalone PLY assets instead of a standard 3DGS training directory.
• This asset repository should not contain code submodules, build outputs, TensorBoard logs, or temporary PLY files.

License and Use

These assets are released for academic research use with the GaussianFluent codebase. Please follow the license and citation requirements of GaussianFluent and any upstream projects used in your experiments.

Citation

@article{huang2026gaussianfluent,
  title={GaussianFluent: Gaussian Simulation for Dynamic Scenes with Mixed Materials},
  author={Huang, Bei and Chen, Yixin and Lu, Ruijie and Zeng, Gang and Zha, Hongbin and Pei, Yuru and Huang, Siyuan},
  journal={arXiv preprint arXiv:2601.09265},
  year={2026}
}