Official implementation of VideoShield: Regulating Diffusion-based Video Generation Models via Watermarking.

Video Examples

ModelScope

Watermarked	Tampered	GT Mask	Pred Mask

Stable-Video-Diffusion

Watermarked	Tampered	GT Mask	Pred Mask

Environment Setup

pip install -r requirements.txt

Model Download

Download the video model to your preferred directory.

The text-to-video (T2V) model ModelScope can be downloaded from: https://huggingface.co/ali-vilab/text-to-video-ms-1.7b.
The image-to-video (I2V) model Stable-Video-Diffusion can be download from: https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt.

Running the Scripts

1. Watermark Embedding and Extraction

For ModelScope:

python3 watermark_embedding_and_extraction.py \
	--device 'cuda:0' \
	--model_name modelscope \
	--model_path <your_model_path> \
	--num_frames 16 \
	--height 256 \
	--width 256 \
	--frames_copy 8 \
	--hw_copy 4 \
	--channel_copy 1 \
	--num_inference_steps 25

For Stable-Video-Diffusion:

python3 watermark_embedding_and_extraction.py \
	--device 'cuda:0' \
	--model_name stable-video-diffusion \
	--model_path <your_model_path> \
	--num_frames 16 \
	--height 512 \
	--width 512 \
	--frames_copy 8 \
	--hw_copy 8 \
	--channel_copy 1 \
	--num_inference_steps 25

Note:

You can also skip specifying --model_path (skip Model Download). The script will automatically download the model to the default cache directory.
The generated watermarked video and watermark information will be saved in the ./results directory by default.

2. Temporal Tamper Localization

For ModelScope:

python3 temporal_tamper_localization.py \
	--device 'cuda:0' \
	--model_name modelscope \
	--model_path <your_model_path> \
	--num_inversion_steps 25 \
	--video_frames_dir './results/modelscope/a_red_panda_eating_leaves/wm/frames'

For Stable-Video-Diffusion:

python3 temporal_tamper_localization.py \
	--device 'cuda:0' \
	--model_name stable-video-diffusion \
	--model_path <your_model_path> \
	--num_inversion_steps 25 \
	--video_frames_dir './results/modelscope/a_red_panda_eating_leaves/wm/frames'

Note:

Default video frames directory: './results/stable-video-diffusion/a_red_panda_eating_leaves/wm/frames' (can be modified as needed)

3. Spatial Tamper Localization

For ModelScope:

python3 spatial_tamper_localization.py \
	--device 'cuda:0' \
	--model_name modelscope \
	--model_path <your_model_path> \
	--num_inversion_steps 25 \
	--video_frames_dir './results/modelscope/a_red_panda_eating_leaves/wm/frames'

For Stable-Video-Diffusion:

python3 spatial_tamper_localization.py \
	--device 'cuda:0' \
	--model_name stable-video-diffusion \
	--model_path <your_model_path> \
	--num_inversion_steps 25 \
	--video_frames_dir './results/modelscope/a_red_panda_eating_leaves/wm/frames'

Note:

Default video frames directory: './results/stable-video-diffusion/a_red_panda_eating_leaves/wm/frames' (can be modified as needed)
The tampered watermarked video, gt mask and pred mask will be saved in the ./results directory by default.

Acknowledgements

This code builds on the code from the GaussianShading.

Cite

If you find this repository useful, please consider giving a star ⭐ and please cite as:

@inproceedings{hu2025videoshield,
  title={VideoShield: Regulating Diffusion-based Video Generation Models via Watermarking},
  author={Runyi Hu and Jie Zhang and Yiming Li and Jiwei Li and Qing Guo and Han Qiu and Tianwei Zhang},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2025}
}