Home Original page

GitHub - jiawei-ren/dreamgaussian4d: [arXiv 2023] DreamGaussian4D: Generative 4D Gaussian Splatting

DreamGaussian4D:
Generative 4D Gaussian Splatting

Jiawei Ren* Liang Pan* Jiaxiang Tang Chi Zhang Ang Cao Gang Zeng Ziwei Liu†

S-Lab, Nanyang Technological University  Shanghai AI Laboratory 
Peking University   University of Michigan  
*equal contribution
†corresponding author

Arxiv 2023

arXiv Paper

examples.mp4

News

  • 2024.6.10: add Gradio demo.
  • 2024.6.9:
    • support LGM for static 3D generation.
    • support video-to-4d generation. Add evaluation scripts for the Consistent4D benchmark. Results are in our updated project page and report.
    • improve the implementation for better speed and quality. Add a gradio demo for image-to-4d.

Install

# python 3.10 cuda 11.8 
conda create -n dg4d python=3.10 -y && conda activate dg4d
conda install -c "nvidia/label/cuda-11.8.0" cuda-toolkit

pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu118
pip install xformers==0.0.23 --no-deps --index-url https://download.pytorch.org/whl/cu118

# other dependencies
pip install -r requirements.txt

# a modified gaussian splatting (+ depth, alpha rendering)
git clone --recursive https://github.com/ashawkey/diff-gaussian-rasterization
pip install ./diff-gaussian-rasterization

# simple-knn
pip install ./simple-knn

# for mesh extraction
pip install git+https://github.com/NVlabs/nvdiffrast/

To use pretrained LGM:

# for LGM
mkdir pretrained && cd pretrained
wget https://huggingface.co/ashawkey/LGM/resolve/main/model_fp16_fixrot.safetensors
cd ..

Image-to-4D

(Optional) Preprocess input image

python scripts/process.py data/anya.png

Step 1: Generate driving videos

python scripts/gen_vid.py --path data/anya_rgba.png --seed 42 --bg white

Step 2: static generation

Static generation with LGM:

python lgm/infer.py big --test_path data/anya_rgba.png

Optionally, we support static generation with DreamGaussian:

python dg.py --config configs/dg.yaml input=data/anya_rgba.png

See configs/dghd.yaml for high-quality DreamGaussian training configurations.

Step 3: dynamic generation

# load static 3D from LGM
python main_4d.py --config configs/4d.yaml input=data/anya_rgba.png

# (Optional) to load static 3D from DreamGaussian, add `radius=2`
python main_4d.py --config configs/4d.yaml input=data/anya_rgba.png radius=2

# (Optional) to turn on viser GUI, add `gui=True`, e.g.:
python main_4d.py --config configs/4d.yaml input=data/anya_rgba.png gui=True

See configs/4d_low.yaml and configs/4d_demo.yaml for more memory-friendly and faster optimization configurations.

(Optional) Step 4: mesh refinment

# export mesh after temporal optimization by adding `mesh_format=obj`
python main_4d.py --config configs/4d.yaml input=data/anya_rgba.png mesh_format=obj

# mesh refinement
python main2_4d.py --config configs/refine.yaml input=data/anya_rgba.png

# (Optional) to load static 3D from DreamGaussian, add `radius=2`
python main2_4d.py --config configs/refine.yaml input=data/anya_rgba.png radius=2

Video-to-4D

Prepare Data

Download Consistent4D data to data/CONSISTENT4D_DATA. python scripts/add_bg_to_gt.py will add white background to ground-truth novel views.

Step 1: static generation

python lgm/infer.py big --test_path data/CONSISTENT4D_DATA/in-the-wild/blooming_rose/0.png

# (Optional) static 3D generation with DG
python dg.py --config configs/dg.yaml input=data/CONSISTENT4D_DATA/in-the-wild/blooming_rose/0.png

Step 2: dynamic generation

python main_4d.py --config configs/4d_c4d.yaml input=data/CONSISTENT4D_DATA/in-the-wild/blooming_rose

# (Optional) to load static 3D from DG, add `radius=2`
python main_4d.py --config configs/4d_c4d.yaml input=data/CONSISTENT4D_DATA/in-the-wild/blooming_rose radius=2

Run demo locally

Load exported meshes in Blender

  • Install the Stop-motion-OBJ add-on
  • File -> Import -> Mesh Sequence
  • Go to logs directory, type in the file name (e.g., 'anya'), and tick Material per Frame.
blender_lowfps.mov

Tips

  • Black video after running gen_vid.py.
    • Make sure pytorch version is >=2.0

Acknowledgement

This work is built on many amazing research works and open-source projects, thanks a lot to all the authors for sharing!

Citation

@article{ren2023dreamgaussian4d,
  title={DreamGaussian4D: Generative 4D Gaussian Splatting},
  author={Ren, Jiawei and Pan, Liang and Tang, Jiaxiang and Zhang, Chi and Cao, Ang and Zeng, Gang and Liu, Ziwei},
  journal={arXiv preprint arXiv:2312.17142},
  year={2023}
}