Rel-Zero

Official code scaffold for Rel-Zero: Harnessing Patch-Pair Invariance for Robust Zero-Watermarking Against AI Editing (CVPR 2026).

Rel-Zero is a zero-watermarking method: it does not modify image pixels. Instead, it predicts a set of stable patch-pair relations from the original image and verifies whether the same relational watermark survives editing or common distortions.

Highlights

• No image embedding or pixel perturbation is required.
• The watermark is represented as top-k stable patch-pair indices.
• Inference needs only the original/suspect image and a trained pair predictor.
• The release includes example edited image pairs and robustness scripts.

Repository Layout

Rel-Zero/
  assets/                 # framework figure used in the paper
  configs/                # example pair manifest and default settings
  examples/test_pairs/    # bundled test images
  src/relzero/            # model, metrics, data, noise and visualization code
  tools/                  # command-line evaluation scripts
  checkpoints/            # put downloaded checkpoints here
  outputs/                # generated visualizations/results

The bundled example pairs intentionally exclude instruct_pix2pix_14_249632.

Installation

conda create -n relzero python=3.10 -y
conda activate relzero

pip install -e .
pip install -r requirements.txt

Install a CUDA-matched PyTorch build if your environment needs a specific CUDA version.

Pretrained Checkpoint

Download the pretrained checkpoint from Hugging Face:

• REL-ZERO checkpoint.pth

Download with wget:

mkdir -p checkpoints
wget -O checkpoints/checkpoint.pth \
  "https://huggingface.co/JERCCC/REL-ZERO/resolve/main/checkpoint.pth?download=true"

The evaluation examples below assume the checkpoint is saved as:

checkpoints/checkpoint.pth

You can also pass any local checkpoint path explicitly with --checkpoint.

The expected checkpoint format is:

{
    "model_state_dict": pipeline.state_dict(),
    "epoch": epoch,
}

Train

The release includes a VAE-label training entry point. It builds stable patch-pair targets from original images and their VAE reconstructions, then trains the original-image-only Rel-Zero predictor.

python tools/train_relzero.py \
  --coco-root path/to/coco/train2014 \
  --coco-ann path/to/coco/annotations/captions_train2014.json \
  --checkpoint-dir checkpoints \
  --run-name relzero_train \
  --device cuda:0 \
  --train-size 2000 \
  --val-size 256 \
  --epochs 8

The best checkpoint is written to:

checkpoints/relzero_train/best_stage1.pth

Evaluate Edited Pairs

python tools/evaluate_pairs.py \
  --checkpoint checkpoints/checkpoint.pth \
  --pairs configs/example_pairs.json \
  --device cuda:0 \
  --negative-match-prob 0.06 \
  --target-fpr 0.001 \
  --output-json outputs/example_pairs.json

Optional edge visualizations:

python tools/evaluate_pairs.py \
  --checkpoint checkpoints/checkpoint.pth \
  --pairs configs/example_pairs.json \
  --save-vis

Evaluate JPEG and Gaussian Noise

python tools/evaluate_noise.py \
  --checkpoint checkpoints/checkpoint.pth \
  --pairs configs/example_pairs.json \
  --device cuda:0 \
  --jpeg-quality 50 \
  --gaussian-sigma 0.05 \
  --negative-match-prob 0.06 \
  --target-fpr 0.001 \
  --output-json outputs/noise.json

Metric

For each image pair, Rel-Zero predicts two top-k patch-pair sets. We report:

precision = shared predicted pairs / top_k

For TPR@0.1%FPR, the release follows the binomial test used in the paper. With top_k=50, negative_match_prob=0.06, and target_fpr=0.001, the detection threshold is computed from the null distribution. Each observed matching probability is then converted into:

P[Binomial(top_k, observed_precision) >= threshold_matches]

Custom Data

Create a manifest with the same format as configs/example_pairs.json:

{
  "pairs": [
    {
      "name": "my_edit",
      "original": "path/to/original.png",
      "edited": "path/to/edited.png"
    }
  ]
}

Relative paths are resolved from the manifest file location.

Citation

@inproceedings{chen2026rel,
  title={Rel-Zero: Harnessing Patch-Pair Invariance for Robust Zero-Watermarking Against AI Editing},
  author={Chen, Pengzhen and Liu, Yanwei and Gu, Xiaoyan and Chen, Xiaojun and Liu, Wu and Wang, Weiping},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={3337--3346},
  year={2026}
}

License

This code scaffold is released under the MIT License. Checkpoint weights and third-party datasets may have separate licenses.