This code is the official implementation of Tree-Ring Watermarks.
If you have any questions, feel free to email Yuxin (ywen@umd.edu).
We propose Tree-Ring Watermarking to watermark diffusion model outputs. Tree-Ring Watermarking chooses the initial noise array so that its Fourier transform contains a carefully constructed pattern near its center. This pattern is called the key. This initial noise vector is then converted into an image using the standard diffusion pipeline with no modifications. To detect the watermark in an image, the diffusion model is inverted to retrieve the original noise array used for generation. This array is then checked to see whether the key is present.
- PyTorch == 1.13.0
- transformers == 4.23.1
- diffusers == 0.11.1
- datasets
Note: higher diffusers version may not be compatible with the DDIM inversion code.
For non-adversarial case, you can simply run:
python run_tree_ring_watermark.py --run_name no_attack --w_channel 3 --w_pattern ring --start 0 --end 1000 --with_tracking --reference_model ViT-g-14 --reference_model_pretrain laion2b_s12b_b42k
You can modify arguments to perform attack. For example, for rotation with 75 degrees:
python run_tree_ring_watermark.py --run_name rotation --w_channel 3 --w_pattern ring --r_degree 75 --start 0 --end 1000 --with_tracking
For more adversarial cases, see here.
For other watermark types mentioned in the paper, you can check scripts/.
You can download 5000 COCO examples used in the paper here. Feel free to add more data or other datasets according to the format of fid_outputs/coco/meta_data.json.
Then, to calculate FID, you may run:
python latent_watermark_fourier_fid.py --run_name fid_run --w_channel 3 --w_pattern ring --start 0 --end 5000 --with_tracking --run_no_w
You can get the pre-trained models here. For example, the link of the model used by the paper is 256x256 diffusion. Then, you may follow the script to run the experiments.
Crucial hyperparameters for Tree-Ring:
w_channel: the index of the watermarked channel. If set as -1, watermark all channels.w_pattern: watermark type: zeros, rand, ring.w_radius: watermark radius.