Detection Limits and Statistical Separability of Tree Ring Watermarks in RF-based Diffusion Transformers

Official Implementation of the paper WMARK@ICLR Detection Limits and Statistical Separability of Tree Ring Watermarks in Rectified Flow-based Text-to-Image Generation Models.

Workflow

Dataset Release

All the results in the paper can be obtained via the .h5 files given in this link

We have also released the dataset on Kaggle.

We welcome additional analysis on this dataset.

Schema of .h5 files

watermark_results_{YYYYMMDD-HHMMSS}_{config_identifier}.h5
├── initial_noise/
│   ├── "0" : Tensor<float32>[channels, height, width]
│   ├── "1" : Tensor<float32>[channels, height, width]
│   └── ...
│
├── generated_images/
│   ├── "0" : Array<uint8>[height, width, 3]
│   ├── "1" : Array<uint8>[height, width, 3]
│   └── ...
│
├── reconstructed_noise/
│   ├── "0" : Tensor<float32>[channels, height, width]
│   ├── "1" : Tensor<float32>[channels, height, width]
│   └── ...
│
├── reconstructed_images/
│   ├── "0" : Array<uint8>[height, width, 3]
│   ├── "1" : Array<uint8>[height, width, 3]
│   └── ...
│
├── watermark_verified : Array<bool>[n_samples]
└── prompts : Array<string>[n_samples]

The detailed code to reproduce the results is given in the evaluate_results.ipynb

Results