Diffusion-Transformer Extrapolation for Long Video Generation

This repository provides the official implementation of RIFLEx and UltraViCo, which achieve diffusion-transformer extrapolation for long video generation in a plug-and-play way.

This repository hosts RIFLEx and UltraViCo on separate branches, and the code is fully open source.

• RIFLEx:

  • main: HunyuanVideo-diffusers and CogVideoX-diffusers
  • multi-gpu: multi-GPU inference for HunyuanVideo

• UltraViCo:

  • ultra-wan: UltraViCo for Wan2.1
  • ultra-hunyuan:UltraViCo for HunyuanVideo

---

UltraViCo: Breaking Extrapolation Limits in Video Diffusion Transformers

 

Min Zhao*, Hongzhou Zhu*, Yingze Wang, Bokai Yan, Jintao Zhang, Guande He, Ling Yang, Chongxuan Li, Jun Zhu

This branch supports UltraViCo for HunyuanVideo. For Wan 2.1, please refer to the `ultra-wan` branch.

Installation

conda create -n ultravico_hy python=3.11 -y
conda activate ultravico_hy
pip install -r requirements.txt

Inference

export PYTHONPATH=$(pwd)/src

torchrun --nproc_per_node=8 --standalone -m parallel_examples.run_attention_patterns \
  --alpha 0.9 \
  --beta 0.6 \
  --extrapolation_ratio 3 \
  --height 544 \
  --width 960 \
  --num_inference_steps 50 \
  --prompt "Brown bear wading slowly through shallow river, splashes frozen mid-air, forest reflection steady on water surface."

• extrapolation_ratio $\in (1,4]$ : the generated video length as a multiple of the training length

• alpha < beta $\in (0,1)$: larger → stronger temporal consistency; smaller → better visual quality.

Acknowledge

• We adopt SageAttention for our UltraViCo attention kernel.

• The code of the parallel diffusers-style HunyuanVideo is build upon ParaAttention. Many tanks to its developers!

• Thank Tecent HuyuanVideo and Wan2.1 for their great open-source models!

---

References

If you find the code useful, please cite

@article{zhao2025ultravico,
  title={UltraViCo: Breaking Extrapolation Limits in Video Diffusion Transformers},
  author={Zhao, Min and Zhu, Hongzhou and Wang, Yingze and Yan, Bokai and Zhang, Jintao and He, Guande and Yang, Ling and Li, Chongxuan and Zhu, Jun},
  journal={arXiv preprint arXiv:2511.20123},
  year={2025}
}

@article{zhao2025riflex,
  title={Riflex: A free lunch for length extrapolation in video diffusion transformers},
  author={Zhao, Min and He, Guande and Chen, Yixiao and Zhu, Hongzhou and Li, Chongxuan and Zhu, Jun},
  journal={arXiv preprint arXiv:2502.15894},
  year={2025}
}