[CVPR 2026] 🎞️MoVieS

MoVieS: Motion-Aware 4D Dynamic View Synthesis in One Second

Chenguo Lin^1*, Yuchen Lin^1,3*, Panwang Pan^2†,
Yifan Yu², Tao Hu², Honglei Yan², Katerina Fragkiadaki³, Yadong Mu¹

¹Peking University, ²ByteDance, ³Carnegie Mellon University

This repository contains the official implementation of the paper: MoVieS: Motion-Aware 4D Dynamic View Synthesis in One Second. MoVieS is a feed-forward framework that jointly reconstructs appearance, geometry and motion for 4D scene perception from monocular videos in one second.

Feel free to contact me (chenguolin@stu.pku.edu.cn) or open an issue if you have any questions or suggestions.

🔥 See Also

You may also be interested in our other works:

• [CVPR 2026] Diff4Splat: a generative model for 4D dynamic scenes from a single-view image.

📢 News

• 2026-02-26: The source codes for inference and training, and pretrained checkpoint are released.
• 2026-02-21: The paper is accepted to CVPR 2026.
• 2025-07-15: This repo is initialized and MoVieS technical report is released on arXiv.

📋 TODO

• Provide source codes for inference and training.
• Provide the pretrained checkpoint.
• Provide detailed instructions for inference and training.
• Make the codebase cleaner.

🔧 Installation

You may need to modify the specific version of torch in settings/setup.sh according to your CUDA version. There are no restrictions on the torch version, feel free to use your preferred one.

conda create -n movies python=3.10 -y
conda activate movies
git clone https://github.com/chenguolin/MoVieS.git
cd MoVieS
bash settings/setup.sh

🚀 Inference Usage

📸 Novel View Synthesis

In-the-wild DAVIS dataset

1. Download the pretrained checkpoint and preprocessed videos with poses from 🤗HuggingFace and put them in resources/. In-the-wild videos are provided by the DAVIS dataset, and their camera poses are estimated by MegaSAM.

mkdir -p resources && cd resources
hf download chenguolin/MoVieS movies_ckpt.safetensors DAVIS/ --local-dir resources

2. Run the inference script for novel view synthesis. In the inference script, we first render the dynamic scene with a fixed camera, and then fix the timestamp and render the scene with a moving camera to synthesize novel views. You can also apply other desired combinations of camera poses and timestamps for novel view rendering.

# python src/infer_davis_nvs.py --name <DAVIS_SAMPLE_NAME>
# For example:
python src/infer_davis_nvs.py --name motocross-bumps

Inference results will be saved in out/<DAVIS_SAMPLE_NAME>. You will get:

Input Video (input_video.mp4)	Predicted Motion (output_motion_camera0.mp4)	Novel View Synthesis (output_render.mp4)

📍 3D Point Tracking

• TODO

🌊 Scene Flow Estimation

• TODO

✂️ Dynamic Object Segmentation

• TODO

🏋️‍♂️ Training Usage

📊 Training Datasets

• We use three static scene datasets (RealEstate10K, TartanAir and MatrixCity) and five dynamic scene datasets (PointOdyssey, DynamicReplica, Spring, VKITTI2 and Stereo4D) to train MoVieS.
• Support combining multiple datasets for training via src/data/easy_dataset.py.
• Support dynamic numbers of input frames and aspect ratios via src/data/dynamic_dataloader.py.
• Set your dataset directory in src/options.py before training.

🧑‍🏫 Training Instructions

• TODO

😊 Acknowledgement

We would like to thank the authors of DiffSplat, VGGT, NoPoSplat, and CUT3R for their great work and generously providing source codes, which inspired our work and helped us a lot in the implementation.

📚 Citation

If you find our work helpful, please consider citing:

@inproceedings{lin2026movies,
  title={MoVieS: Motion-Aware 4D Dynamic View Synthesis in One Second},
  author={Lin, Chenguo and Lin, Yuchen and Pan, Panwang and Yu, Yifan and Hu, Tao and Yan, Honglei and Fragkiadaki, Katerina and Mu, Yadong},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
  year={2026}
}