Skip to content
Projects

yikaiw/Vidu4D has 371 GitHub stars and 12 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-18 00:31 UTC. Last push 2025-01-12.

How these counts are measured

Measured star growth

in 24h
0
in 7 days
Update delayed
in 15 days
Update delayed
in 1 month
Update delayed

Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.

About

[TPAMI 2025, NeurIPS 2024] Video4DGen: Enhancing Video and 4D Generation through Mutual Optimization

Creative & Media

README.md

Original

Vidu4D

This repository is an implementation of "Vidu4D: Single Generated Video to High-Fidelity 4D Reconstruction with Dynamic Gaussian Surfels", in NeurIPS 2024.

Vidu4D: Single Generated Video to High-Fidelity 4D Reconstruction with Dynamic Gaussian Surfels
Yikai Wang*, Xinzhou Wang*, Zilong Chen, Zhengyi Wang, Fuchun Sun, Jun Zhu

   

Overview

4D Results

See accompanying videos in our project page for better visual quality.

Training

First, specify the experiment name,

seqname=cheetah

Then, create a seqname folder under database/raw and put a video (e.g., 0.mp4) in it.

Stage 1, data preprocess, including data cropping, optical flow, DINO feature, rough poses, etc,

python scripts/run_preprocess.py $seqname $seqname other "0"

Stage 2, optimize neural sdf,

python lab4d/train.py --seqname $seqname --logname base --fg_motion bob --num_rounds 21 --rgb_timefree --rgb_dirfree

Stage 3, optimize Gaussian surfels,

python lab4d/train.py --seqname ${seqname} --logname gs-frzwarp --fg_motion gs-bob --num_rounds 61 --load_path logdir/${seqname}-base/ckpt_0020.pth --gs_init_mesh logdir/${seqname}-base/021-fg-geo.obj --imgs_per_gpu 1 --pixels_per_image -1 --eval_res 256 --rgb_timefree --rgb_dirfree --rgb_loss_only --gs_optim_warp=False --data_prefix full --force_center_cam

Render after training,

python lab4d/render.py --flagfile=logdir/${seqname}-gs-frzwarp/opts.log --load_suffix latest --render_res 512

Todo

  • DGS Refinement
  • Benchmarks

Acknowledgements

Our code is built based on 2DGS and Lab4D. We thank the authors for their great repos.

BibTeX

@inproceedings{wang2024vidu4d,
  title={Vidu4D: Single Generated Video to High-Fidelity 4D Reconstruction with Dynamic Gaussian Surfels},
  author={Yikai Wang and Xinzhou Wang and Zilong Chen and Zhengyi Wang and Fuchun Sun and Jun Zhu},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2024}
}