Skip to content
/ DS4D Public

[ICCV 2025 Highlight] Not all Frame Features are Equal: Video-to-4d Generation via Decoupling Dynamic-static Features

Notifications You must be signed in to change notification settings

LiyingCV/DS4D

Repository files navigation

Not all Frame Features are Equal: Video-to-4d Generation via Decoupling Dynamic-static Features

Liying Yang1,Chen Liu2,Zhenwei Zhu1,Ajian Liu3,  Hui Ma1,  Jian Nong1,Yanyan Liang1,†
1Macau University of Science and Technology  2The University of Queensland  3Institute of Automation, Chinese Academy of Sciences
Corresponding Author

ArXiv

ICCV Highlight🔥

Example

🛠️ The codes and datasets will be released after we clean down

Installation

The environment was tested on NVIDIA 4090, A100, A10, 3090, and Tesla V100.

Clone the code repository

git clone https://github.com/LiyingCV/DS4D.git

Create a new environment from environment.yml

conda env create -f environment.yml
conda activate ds4d

Install Gaussian splatting and simple-knn

pip install ./diff-gaussian-rasterization
pip install ./simple-knn

Install pointnet2_ops

cd tgs/models/snowflake/pointnet2_ops_lib && python setup.py install && cd -

Install PyTorch3D

Prepare data

Before starting training, we first use InstantMesh to reconstruct the 3D object of the middle frame. For more details about how to reconstruct, please refer to InstantMesh. Then, we need to scale and rotate the 3D object to fit Gaussian splatting. By the way, we can also use other 3D reconstruction models or 3D generation models to produce the 3D object, while we use InstantMesh in the paper.

Training

During trainig, the code will produce the result of 360° video and front/back/left/right video.

python runner.py --config /path/to/config/ --item /the/name/of/item/

📚 Citation

If you find our work useful for your research, please consider citing our paper:

@InProceedings{Yang_2025_ICCV,
    author    = {Yang, Liying and Liu, Chen and Zhu, Zhenwei and Liu, Ajian and Ma, Hui and Nong, Jian and Liang, Yanyan},
    title     = {Not All Frame Features Are Equal: Video-to-4D Generation via Decoupling Dynamic-Static Features},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {7494-7504}
}

About

[ICCV 2025 Highlight] Not all Frame Features are Equal: Video-to-4d Generation via Decoupling Dynamic-static Features

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published