| Project Page | Paper |

This is a version of AC3D built on CogVideoX. AC3D is a camera-controlled video generation pipeline that follows the plucker-conditioned ControlNet architecture originally introduced in VD3D.

Install PyTorch first (we used PyTorch 2.4.0 with CUDA 12.4).

pip install -r requirements.txt

Prepare the RealEstate10K dataset following the instructions in CameraCtrl. The dataset path will be used for video_root_dir in the train and inference scripts. This is the folder structure after pre-processing:

- RealEstate10k
  - annotations
    - test.json
    - train.json
  - pose_files
    - 0000cc6d8b108390.txt
    - 00028da87cc5a4c4.txt
    - ...
  - video_clips
    - 0000cc6d8b108390.mp4
    - 00028da87cc5a4c4.mp4
    - ...

AC3D: CogVideoX-2B: Checkpoint

AC3D: CogVideoX-5B: Checkpoint

AC3D: CogVideoX-2B

bash scripts/inference_2b.sh

AC3D: CogVideoX-5B

bash scripts/inference_5b.sh

The 2B model requires 48 GB memory and the 5B model requires 80 GB memory. Using one node with 8xA100 80 GB should take around 1-2 days for the model to converge.

These are the fine-tuning scripts to train the ControlNet models on top of a pre-trained base model.

AC3D: CogVideoX-2B

bash scripts/train_2b.sh

AC3D: CogVideoX-5B

bash scripts/train_5b.sh

• This code mainly builds upon CogVideoX-ControlNet
• This code uses the original CogVideoX model CogVideoX
• The data procesing and data loading pipeline builds upon CameraCtrl

@article{bahmani2024ac3d,
  author = {Bahmani, Sherwin and Skorokhodov, Ivan and Qian, Guocheng and Siarohin, Aliaksandr and Menapace, Willi and Tagliasacchi, Andrea and Lindell, David B. and Tulyakov, Sergey},
  title = {AC3D: Analyzing and Improving 3D Camera Control in Video Diffusion Transformers},
  journal = {arXiv preprint arXiv:2411.18673},
  year = {2024},
}