Rolling Forcing

Autoregressive Long Video Diffusion in Real Time

Kunhao Liu¹ · Wenbo Hu² · Jiale Xu² · Ying Shan² · Shijian Lu^1
1Nanyang Technological University ²ARC Lab, Tencent PCG

💡 TL;DR: REAL-TIME streaming generation of MULTI-MINUTE videos!

https://github.com/user-attachments/assets/7b43ded2-7f29-41a1-8244-a1fc49c418e5

• 🚀 Real-Time at 16 FPS:​​ Stream high-quality video directly from text on a ​single GPU.
• 🎬 Minute-Long Videos:​​ Generate coherent, multi-minute sequences with ​dramatically reduced drift.
• ​⚙️ Rolling-Window Strategy:​​ Denoise frames together in a rolling window for mutual refinement, ​breaking the chain of error accumulation.
• ​🧠 Long-Term Memory:​​ The novel ​Attention Sink​ anchors your video, preserving global context over thousands of frames.
• ​🥇 State-of-the-Art Performance:​​ Outperforms all comparable open-source models in quality and consistency.

🛠️ Installation

Create a conda environment and install dependencies:

conda create -n rolling_forcing python=3.10 -y
conda activate rolling_forcing
pip install -r requirements.txt
pip install flash-attn --no-build-isolation

🚀 Quick Start

Download checkpoints

huggingface-cli download Wan-AI/Wan2.1-T2V-1.3B --local-dir-use-symlinks False --local-dir wan_models/Wan2.1-T2V-1.3B
huggingface-cli download TencentARC/RollingForcing checkpoints/rolling_forcing_dmd.pt --local-dir .

CLI inference

Example inference script:

python inference.py \
    --config_path configs/rolling_forcing_dmd.yaml \
    --output_folder videos/rolling_forcing_dmd \
    --checkpoint_path checkpoints/rolling_forcing_dmd.pt \
    --data_path prompts/example_prompts.txt \
    --num_output_frames 126 \
    --use_ema

Gradio demo (minimal UI)

Run a local web demo that takes a text prompt and shows the generated video.

1. Ensure the Wan base model and checkpoint above are downloaded.
2. Launch the app:

python app.py \
  --config_path configs/rolling_forcing_dmd.yaml \
  --checkpoint_path checkpoints/rolling_forcing_dmd.pt

Then open the printed local URL in your browser.

📈 Training

Download training prompts, ODE-initialized checkpoint, and teacher model

huggingface-cli download gdhe17/Self-Forcing checkpoints/ode_init.pt --local-dir .
huggingface-cli download gdhe17/Self-Forcing vidprom_filtered_extended.txt --local-dir prompts
huggingface-cli download Wan-AI/Wan2.1-T2V-14B --local-dir wan_models/Wan2.1-T2V-14B

Train Rolling Forcing on a single machine with 8 GPUs

torchrun --nproc_per_node=8 \
  --rdzv_backend=c10d \
  --rdzv_endpoint 127.0.0.1:29500 \
  train.py \
  -- \
  --config_path configs/rolling_forcing_dmd.yaml \
  --logdir logs/rolling_forcing_dmd

🔖 Citation

If you find this codebase useful for your research, please kindly cite our paper and consider giving this repo a ⭐️.

@article{liu2025rolling,
  title={Rolling Forcing: Autoregressive Long Video Diffusion in Real Time},
  author={Liu, Kunhao and Hu, Wenbo and Xu, Jiale and Shan, Ying and Lu, Shijian},
  journal={arXiv preprint arXiv:2509.25161},
  year={2025}
}

🙏 Acknowledgements

• Self Forcing: the codebase and algorithm we built upon. Thanks for their wonderful work.
• Wan: the base model we built upon. Thanks for their wonderful work.