For the best experience, it's better to use a desktop computer to view this website.

Bowtie-flow icon Bowtie-flow: Efficient High-resolution Video Generation with Prior Preservation

Kaixin Ding1, Xi Chen1, Sihui Ji1, Yuan Gao2, Liang Hou2, Xin Tao2, Pengfei Wan2, Hengshuang Zhao1

1The University of Hong Kong
2Kling Team, KuaishouTechnology

Please stay tuned for all videos loading...

TL;DR: High-resolution video generation is slow: for example, Wan 2.1 takes over 50 minutes to generate a 720p video. Existing acceleration methods often compromise model priors (layout, semantics, motion). We propose Bowtie-flow, a two-stage framework: first, a fast low-resolution preview using a pretrained model; second, a Refiner to upscale while preserving priors. Key techniques include noise reshifting to reduce prior loss and shifting windows with careful training design. Bowtie-flow is simple, efficient, and compatible with various base models, achieving 12.5× speedup for generating 5-second, 16fps, 720p Wan 2.1 videos and 8.7× speedup for generating 5-second, 24fps, 720p HunyuanVideo.

Contact Us

Feel free to contact Kaixin Ding at kxding@connect.hku.hk for any question,cooperation, and communication.

If you find this work useful, please consider citing:

@article{ding2025bowtieflow,
            title={Bowtie-flow: Efficient High-resolution Video Generation with Prior Preservation},
            author={Kaixin Ding and Xi Chen and Sihui Ji and Yuan Gao and Liang Hou and Xin Tao and Pengfei Wan and Hengshuang Zhao},
            journal={arXiv preprint arXiv:},
            year={2025}
        }

Thank Jenga, UltraPixel, ControlNeXt to provide us the project page's template!