UniVid (WACV2026)

UniVid: Unifying Vision Tasks with Pre-trained Video Generation Models
Lan Chen, Yuchao Gu and Qi Mao
MIPG, Communication University of China; Show Lab, National University of Singapore

Getting Started

1. Environment setup

git clone https://github.com/CUC-MIPG/UniVid.git
cd Univid

conda create -n Univid python=3.11
conda activate Univid

2. Requirements installation

pip install requirements.txt

We use the open-source DiffSynth-Studio codebase to train UniVid. Please merge the diffSynth folder from this repository into the corresponding folder in DiffSynth-Studio.

Then, run:

pip install -e .

3. Dataset Preparation

We provide an example dataset of camera movement task. Each video contains four clips arranged in the order $(A \rightarrow A' \rightarrow B \rightarrow B')$, each with the same length.
Make sure the folder structure follows:

data/
  train/
    camera_movement/
      metadata.csv
      train/

To prepare your customized data, please ensure that each clip has exactly $4n + 1$ frames, and that the total number of frames in all four clips is less than $81$.

Run the script data_process.sh to preprocess the dataset.

4. Start training

Run the script train.sh to start training.

An example UniVid model for the camera movement task can be downloaded from this link

5. Inference

Run the script inference.sh for inference.

Main Observations

Citation

@article{chen2025univid,
  title={UniVid: Unifying Vision Tasks with Pre-trained Video Generation Models},
  author={Chen, Lan and Gu, Yuchao and Mao, Qi},
  journal={arXiv preprint arXiv:2509.21760},
  year={2025}
}