---
datasets:
- quanhaol/MagicData
base_model:
- Wan-AI/Wan2.2-TI2V-5B
---
Wan2.2-TI2V-5B-Turbo
Wan2.2-TI2V-5B-Turbo is designed for efficient step distillation and CFG distillation based on Wan2.2-TI2V-5B.
Leveraging the Self-Forcing framework, it enables 4-step TI2V-5B model training. **Our model can generate 121-frame videos at 24 FPS with a resolution of 1280ร704 in just 4 steps, eliminating the need for the CFG trick.**
To the best of our knowledge, Wan2.2-TI2V-5B-Turbo is the **first** open-source repository of the distilled I2V version of Wan2.2-TI2V-5B.
## ๐ฅVideo Demos
The videos below can be reproduced using [examples/example.csv](examples/example.csv).
## ๐ฃ Updates
- `2025/08/06` ๐ฅWan2.2-TI2V-5B-Turbo has been released [`here`](https://huggingface.co/quanhaol/Wan2.2-TI2V-5B-Turbo).
## ๐ Installation
Create a conda environment and install dependencies:
```bash
conda create -n wanturbo python=3.10 -y
conda activate wanturbo
pip install -r requirements.txt
pip install flash-attn --no-build-isolation
python setup.py develop
```
## ๐Quick Start
### Checkpoint Download
```bash
pip install "huggingface_hub[hf_transfer]"
HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli download Wan-AI/Wan2.2-TI2V-5B --local-dir wan_models/Wan2.2-TI2V-5B
HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli download quanhaol/Wan2.2-TI2V-5B-Turbo --local-dir wan_models/Wan2.2-TI2V-5B-Turbo
```
### DMD Training
```bash
bash running_scripts/train/Wan2.2/dmd.sh
```
Our training run uses 4000 iterations and completes in under 2 days using 16 A100 GPUs.
### Fewstep Inference
```bash
bash running_scripts/inference/Wan2.2/i2v_fewstep.sh
```
## ๐ค Acknowledgements
We would like to express our gratitude to the following open-source projects that have been instrumental in the development of our project:
- [CausVid](https://github.com/tianweiy/CausVid)
- [Self-Forcing](https://github.com/guandeh17/Self-Forcing)
- [Self-Forcing-Plus](https://github.com/GoatWu/Self-Forcing-Plus)
- [Wan2.1](https://github.com/Wan-Video/Wan2.1)
- [Wan2.2](https://github.com/Wan-Video/Wan2.2)
Special thanks to the contributors of these libraries for their hard work and dedication!
## ๐ Contact
If you have any suggestions or find our work helpful, feel free to contact us
Email: liqh24@m.fudan.edu.cn or zhenxingfd@gmail.com or wangrui21@m.fudan.edu.cn
If you find our work useful, please consider giving a star to this github repository and citing it:
```bibtex
@article{li2025magicmotion,
title={MagicMotion: Controllable Video Generation with Dense-to-Sparse Trajectory Guidance},
author={Li, Quanhao and Xing, Zhen and Wang, Rui and Zhang, Hui and Dai, Qi and Wu, Zuxuan},
journal={arXiv preprint arXiv:2503.16421},
year={2025}
}
```