Kunhao Liu1
·
Wenbo Hu2
·
Jiale Xu2
·
Ying Shan2
·
Shijian Lu1
1Nanyang Technological University 2ARC Lab, Tencent PCG
video.mp4
- 🚀 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.
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
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 .
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
Run a local web demo that takes a text prompt and shows the generated video.
- Ensure the Wan base model and checkpoint above are downloaded.
- 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.
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
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
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}
}- 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.