MotionCtrl

We have released both Training and Inference code of MotionCtrl deployed on AnimateDiff

We release the codes, models and demos for MotionCtrl on Stable Video Diffusion (SVD).

https://github.com/TencentARC/MotionCtrl/assets/19488619/45d44bf5-d4bf-4e45-8628-2c8926b5954a

Official implementation of MotionCtrl: A Unified and Flexible Motion Controller for Video Generation.

MotionCtrl can Independently control complex camera motion and object motion of generated videos, with only a unified model.

Results of MotionCtrl+AnimateDiff

Results of MotionCtrl+SVD

More results are in showcase_svd and our Project Page.

Results of MotionCtrl+VideoCrafter

More results are in our Project Page.

Changelog

• 20231225: Release MotionCtrl deployed on LVDM/VideoCrafter.

• 20231225: Gradio demo available.

• 20231228: Provide local gradio demo for convenience.

• 20240115 More camera poses used for testing are provided in dataset/camera_poses

• 20240115 Release MotionCtrl deployed on SVD. Codes are in the branch svd and Gradio Demo is available in .

• ❗❗❗ Gradio demo of MotionCtrl deployed on VideoCrafter2 is available in . You can also run it locally by python -m app --share.

• ❗❗❗ Release MotionCtrl deployed on AnimateDiff are available in branch animatediff, containing both training and inference code.

• 20240920 Provide scripts for collecting object trajectories with ParticleSfM.

• 20240920 We provide a HandyTrajDrawer to customize object trajectories more conveniently.

⚙️ Environment

conda create -n motionctrl python=3.10.6
conda activate motionctrl
pip install -r requirements.txt

Inference

• Run local inference script

1. Download the weights of MotionCtrl motionctrl.pth and put it to ./checkpoints.
2. Go into configs/inference/run.sh and set condtype as 'camera_motion', 'object_motion', or 'both'.

• condtype=camera_motion means only control the camera motion in the generated video.
• condtype=object_motion means only control the object motion in the generated video.
• condtype=both means control the camera motion and object motion in the generated video simultaneously.

1. Running scripts: sh configs/inference/run.sh

• Run local gradio demo

  复制

  python -m app --share

Training Details

Preparing Dataset

• RealEstate10K

  1. Following https://github.com/cashiwamochi/RealEstate10K_Downloader to download and process the videos.
  2. Corresponding Captions and List are provided in GoogleDrive.

• WebVid with Object Trajectories

  1. Preparing ParticleSfM. Our experiments is running on CentOS 8.5 and we provide a detailed install note in dataset/object_trajectories/ParticleSfM_Install_Note.pdf.

  2. Moving dataset/object_trajectories/prepare_webvideo_len32.py and dataset/object_trajectories/run_particlesfm_obj_traj.py to ParticleSfM project.

  3. Step 1: Prepare sub-videos with lenth of 32 and size of 256 x 256.

     bash复制

     ## start_idx and end_idx is used to process a subset of the dataset in different machines parallelly
     
     python prepare_webvideo_len32.py --start_idx 0 --end_idx 1000

  4. Step 2: Get object trajectories

     bash复制

       root_dir="WebVid/train_256_32"
       start_idx=0
       end_idx=1000
     
       CUDA_VISIBLE_DEVICES=0 python run_particlesfm_obj_traj.py \
       --root_dir $root_dir \
       --start_idx $start_idx \
       --end_idx $end_idx \

• You can customize object Trajectories with our provided HandyTrajDrawer.

:books: Citation

If you make use of our work, please cite our paper.

@inproceedings{wang2024motionctrl,
  title={Motionctrl: A unified and flexible motion controller for video generation},
  author={Wang, Zhouxia and Yuan, Ziyang and Wang, Xintao and Li, Yaowei and Chen, Tianshui and Xia, Menghan and Luo, Ping and Shan, Ying},
  booktitle={ACM SIGGRAPH 2024 Conference Papers},
  pages={1--11},
  year={2024}
}

Acknowledgment

The current version of MotionCtrl is built on VideoCrafter. We appreciate the authors for sharing their awesome codebase.

❓ Contact

For any question, feel free to email [email protected] or [email protected].