FlowMotion: Training-Free Flow Guidance for Video Motion Transfer

Zhen Wang¹, Youcan Xu², Jun Xiao², Long Chen¹*

¹The Hong Kong University of Science and Technology (HKUST)
²Zhejiang University

CVPR 2026

Efficient Video Motion Transfer with one single RTX 3090 🚀

🚀 Quick Start

1. Setup

Create the environment and install the dependencies by running:

conda create -n flowmotion python=3.10
conda activate flowmotion
pip install -r requirements.txt

2. Data Preparation

Put your videos in the data/ folder, we have already prepared several examples.

3. Running FlowMotion

Run the following command for video motion transfer:

python run.py --video_path ./data/49f/bus.mp4 --target_prompt "A black shark is moving through crystal-clear ocean waters."

Try more examples, such as:

python run.py --video_path ./data/49f/car-turn.mp4 --target_prompt "A dog is running in a city garden."

python run.py --video_path ./data/49f/blackswan.mp4 --target_prompt "A spaceship glides silently through the universe."

python run.py --video_path ./data/49f/hike.mp4 --target_prompt "A penguin is walking along a frozen coastline."

Try longer videos with more frames, such as:

python run.py --video_path ./data/81f/car-turn.mp4 --target_prompt "A tiger is running through a dense jungle."

python run.py --video_path ./data/81f/bus.mp4 --target_prompt "A eagle soaring over rugged mountain peaks."

4. Variant of Source Motion Representation

For videos with more complex motions, try using the varient of source motion representation with clean latent, such as:

python run.py --video_path ./data/49f/motocross-jump.mp4 --target_prompt "A monkey on a motorcycle is mid-jumping on a dirt track in the forest." --guidance_type clean_latent

besides, larger learning rate and more detailed prompt all help to get better results, such as:

python run.py --video_path ./data/49f/snowboard.mp4 --target_prompt "A monkey is surfing down a towering, sun-bleached dune, a sleek figure of focused agility against the vast emptiness. A plume of golden sand arcs from the board, shimmering in the heat haze that dances above the sand. In the distance, a cluster of resilient cacti stands witness to this burst of dynamic life under the relentless sun." --guidance_type clean_latent --lr_base 0.01

💻 VRAM Requirements

Below is the VRAM usage on a single NVIDIA RTX 3090 (24GB VRAM), you can save more cost by using enable_model_cpu_offload().

Video Length	VRAM Usage	Enable CPU Offload
49 frames	19 GB	4.9 GB
81 frames	20 GB	5.8 GB

🗓️ TODO

• Release demo
• Release benchmark

🖊️ BibTeX

If you find this project useful in your research, please consider cite:

@article{wang2026flowmotion,
  title={FlowMotion: Training-Free Flow Guidance for Video Motion Transfer},
  author={Wang, Zhen and Xu, Youcan and Xiao, Jun and Chen, Long},
  journal={arXiv preprint arXiv:2603.06289},
  year={2026}
}

Acknowledgements

We thank to Wan and diffusers.