nhzjl/W-Bench

[ICLR 2025] Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances

What is it?

W-Bench is the first benchmark to evaluate watermarking robustness across four types of image editing techniques, including regeneration, global editing, local editing, and image-to-video generation.

Eleven representative watermarking methods are evaluated on the W-Bench. The W-Bench contains 10,000 instances sourced from datasets such as COCO, Flickr, ShareGPT4V, etc.

GitHub Repo: https://github.com/Shilin-LU/VINE

Dataset Structure

The evaluation set consists of six subsets, each targeting a different type of AIGC-based image editing:

• 1,000 samples for stochastic regeneration
• 1,000 samples for deterministic regeneration (aka, image inversion)
• 1,000 samples for global editing
• 5,000 samples for local editing (divided into five sets, each containing 1,000 images and 1,000 masks, with different mask sizes ranging from 10–60% of the image area)
• 1,000 samples for image-to-video generation
• 1,000 samples for testing conventional distortion (identical to the 1,000 samples used for deterministic regeneration)

How to download and use 🍷 W-Bench

Using huggingface_hub

huggingface-cli download Shilin-LU/W-Bench --repo-type=dataset --local-dir W-Bench

or

from huggingface_hub import snapshot_download
folder = snapshot_download(
  "Shilin-LU/W-Bench",
  repo_type="dataset",
  local_dir="./W-Bench/",
  allow_patterns="DET_INVERSION_1K/image/*" # to download a specific branch
)

For faster downloads, make sure to install pip install huggingface_hub[hf_transfer] and set the environment variable HF_HUB_ENABLE_HF_TRANSFER=1.

Using datasets

1. With Stream

from datasets import load_dataset
dataset = load_dataset("Shilin-LU/W-Bench", split="train", streaming=True)
next(iter(dataset))['image'].save('output_stream.png')

2. Without Stream

from datasets import load_dataset
dataset = load_dataset("Shilin-LU/W-Bench", split="train")
dataset[1]['image'].save('output.png')

Citation Information

Paper on arXiv

@article{lu2024robust,
  title={Robust watermarking using generative priors against image editing: From benchmarking to advances},
  author={Lu, Shilin and Zhou, Zihan and Lu, Jiayou and Zhu, Yuanzhi and Kong, Adams Wai-Kin},
  journal={arXiv preprint arXiv:2410.18775},
  year={2024}
}