wmforger

🖼️🔓 Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models

NeurIPS 2025 (Spotlight 🏅)

[arXiv] [Colab]

🎮 Usage

See and try our example jupyter notebook in Colab or try the mimimal example of removing watermark localy using the steps below.

1. Install PyTorch and other requirements.

2. Clone the repository.

   git clone https://github.com/facebookresearch/videoseal.git
   cd videoseal/wmforger/

3. Download the pretrained model weights.

   wget https://dl.fbaipublicfiles.com/wmforger/convnext_pref_model.pth

4. Extract watermark.

   python optimize_image.py --ckpt_path convnext_pref_model.pth --image assets/tahiti_watermarked.png

🚆 Train preference model from scratch

1. Download SA-1b dataset.

2. Update path to the dataset in configs/datasets/sa-1b-full.yaml

3. Train. We trained using 8 GPUs.

   sbatch train-slurm.sh

🧾 License

Please see the LICENSE file in the root of the main repository.

✍️ Citation

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

@inproceedings{soucek2025transferable,
  title={Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models},
  author={Sou\v{c}ek, Tom\'{a}\v{s} and Rebuffi, Sylvestre-Alvise and Fernandez, Pierre and Jovanovi\'{c}, Nikola and Elsahar, Hady and Lacatusu, Valeriu and Tran, Tuan and Mourachko, Alexandre},
  booktitle={Advances in Neural Information Processing Systems},
  year={2025}
}