PropFly: Learning to Propagate via On-the-Fly Supervision from Pre-trained Video Diffusion Models

Official repository for PropFly: Learning to Propagate via On-the-Fly Supervision from Pre-trained Video Diffusion Models

[Project Page]  |  [arXiv]

1. Create and activate the Conda environment

If you do not already have a propfly environment, create one with Python 3.10:

conda create -n propfly python=3.10 -y
conda activate propfly

2. Install dependencies

Install the project requirements inside the active environment:

pip install -r requirements.txt

3. Prepare the demo checkpoint

Download the pre-trained PropFly weights from our Hugging Face repository: 👉 james16/PropFly

You can download the model manually from the link above or use the huggingface-cli. Place the downloaded diffusion_pytorch_model.bin file inside the ./checkpoints/ directory.

Using huggingface-cli (Recommended):

mkdir -p checkpoints
huggingface-cli download james16/PropFly --local-dir ./checkpoints

(Note: If the file downloads into a subfolder like checkpoints/PropFly-1.3B/, move the .bin file directly to the root of checkpoints/.)

demo.sh expects a transformer checkpoint at:

./checkpoints/diffusion_pytorch_model.bin

If your checkpoint lives elsewhere, either copy it to that path or edit demo.sh and update --transformer_path.

4. Verify the demo assets

The demo script uses the sample assets already included in the repo:

• assets/demo/bear/
• assets/demo/bear.png
• assets/demo/bear_input.gif
• assets/demo/bear_output.gif

If you want to test with your own inputs, update the paths in demo.sh.

5. Preview the demo result

You can view the sample input and output directly in the README preview:

Input Video

Reference Image

↓

Output Video

6. Run the demo

Run the shell script from the project root:

bash demo.sh

The script will:

• load the model from Wan-AI/Wan2.1-VACE-1.3B-Diffusers
• use the local reference image and frame folder
• run inference with the settings defined in demo.sh
• save results under ./outputs/

Notes

• The demo is GPU-intensive and may take several minutes depending on your hardware.
• If you change the input image folder, make sure the frame files are sorted the way you expect.
• Output files are written to the directory specified by --output_dir.

Citation

If you find this repository useful, please cite:

@article{seo2026propfly,
	title={PropFly: Learning to Propagate via On-the-Fly Supervision from Pre-trained Video Diffusion Models},
	author={Seo, Wonyong and Moon, Jaeho and Lee, Jaehyup and Kim, Soo Ye and Kim, Munchurl},
	journal={arXiv preprint arXiv:2602.20583},
	year={2026},
	doi={10.48550/arXiv.2602.20583}
}

Acknowledgments

This project is based on VACE:

• VACE: https://github.com/ali-vilab/VACE