🎬 ReDirector: Creating Any-Length Video Retakes with Rotary Camera Encoding

This repository contains the official implementation of "ReDirector: Creating Any-Length Video Retakes with Rotary Camera Encoding", including both training and inference code.

🔥Note: You can adjust config parameters using hydra configs here. Feel free to modify them to match your own experimental setup, and if you encounter any issues, please open an issue or contact us.

Setup

• Step 1: Install conda envs

  conda env create -f environment.yml
  conda activate redirector
  pip install -r requirements.txt
  pip install flash_attn --no-build-isolation
  pip install --no-build-isolation git+https://github.com/mohammadasim98/met3r
  pip install --no-build-isolation -e vipe

• Step 2: Download checkpoints (Or, you can manually download from Wan Checkpoints and ReDirector Checkpoints, and place them in models folder.):

  python download_model.py

• Step 3: Prepare the training dataset MultiCamVideo dataset.

• Step 4: Extract features for the faster training:

  torchrun --nproc-per-node=8 extract_features.py dataset_path=<path/to/dataset> extract.reverse=true
  torchrun --nproc-per-node=8 extract_features.py dataset_path=<path/to/dataset> extract.reverse=false

Training

• We train our model for 20k iteration on 8 RTX Pro 6000 Blackwell GPUs. It takes around 90 hours.

  torchrun --nproc-per-node=8 train.py dataset_path=<path/to/dataset>

Inference and Evaluation

• Given test the example videos (.mp4 or images directory), we follow ReCamMaster for the preset camera types as below.

  python inference.py eval.ckpt_path=models/step20000.safetensors eval.video_path=example_data/bear eval.cam_type=1 eval.cam_speed=1

  cam_type	Trajectory
  1	Pan Right
  2	Pan Left
  3	Tilt Up
  4	Tilt Down
  5	Zoom In
  6	Zoom Out
  7	Translate Up (with rotation)
  8	Translate Down (with rotation)
  9	Arc Left (with rotation)
  10	Arc Right (with rotation)

• The reported metrics in the paper except VBench were obtained by running the code below.

  python evaluate.py --data_path results/bear_camera_1_speed_1.mp4

Citation

If you find this repository helpful for your project, please consider citing our work. :)

@article{park2025redirector,
  title={ReDirector: Creating Any-Length Video Retakes with Rotary Camera Encoding},
  author={Park, Byeongjun and Kim, Byung-Hoon and Chung, Hyungjin and Ye, Jong Chul},
  journal={arXiv preprint arXiv:2511.19827},
  year={2025}
}