Bridging Limited-Horizon Training and Open-Ended Testing in Autoregressive Video Diffusion

Haodong Li¹, Shaoteng Liu², Zhe Lin², Manmohan Chandraker^1✉

¹UC San Diego ²Adobe Research ^✉Corresponding author

Please click the figure below to watch the teaser video on YouTube.

📢 News

• 2026-03-14 Add Causal Forcing as a new base model option!
• 2026-02-22 Demo (local & online) released!
• 2026-02-11 Code released!
• 2026-02-10 Paper released on arXiv!

🛠️ Setup

This installation was tested on: Ubuntu 20.04, CUDA 12.4, NVIDIA A40.

1. Clone the repository:

git clone https://github.com/haodong2000/RollingSink.git
cd RollingSink

2. Install dependencies using conda:

conda create -n RS python=3.10 -y
conda activate RS
pip install -r requirements.txt

3. Download checkpoints:

sh shell_scripts/download_ckpt.sh

🤗 Gradio Demo

1. Run the demo online: HF Space
2. Run it locally:

python app.py

🕹️ Inference

1. Prepare prompts under prompts/

We've pre-uploaded some examples under prompts/example/

2. Run the inference command:

sh shell_scripts/cuda_i.sh

i $\in$ {0,1,2,3}
The default video length is 5-minute, which requires GPUs with $\geqslant$ 48GB memory.

🎓 Citation

If you find our work useful in your research, please consider citing our paper🌹:

@article{li2026rolling,
  title={Rolling Sink: Bridging Limited-Horizon Training and Open-Ended Testing in Autoregressive Video Diffusion},
  author={Li, Haodong and Liu, Shaoteng and Lin, Zhe and Chandraker, Manmohan},
  journal={arXiv preprint arXiv:2602.07775},
  year={2026}
}

🤝 Acknowledgement

This implementation is impossible without the awesome open-cource contributions of:

• Self Forcing
• LongLive
• CausVid
• Wan
• Gradio
• HuggingFace Hub
• PyTorch