EditShield

EditShield: Protecting Unauthorized Image Editing by Instruction-guided Diffusion Models. Project Page | Paper

TL;DR: quickstart

Step 1. Set up a conda environment, and download a pre-trained model:

conda env create -f environment.yaml
conda activate Editshield
bash scripts/download_checkpoints.sh

Step 2. Create a perturbation using the EditShield with Expectation Over Transformation (EOT), we use center cropping as an example:

python EOT_Center.py

We also provide codes for Gaussian kernel smoothing and Image rotation.

Step 3. Generate descriptions of the caption using LLaVA:

python Cap_EOT_C.py

Step 4. Calculate the CLIP image similarity and CLIP text-image direction similarity:

python --number 5 --method EOT-C Caption_Metrics.py

Reference

@article{chen2023editshield,
  title={EditShield: Protecting Unauthorized Image Editing by Instruction-guided Diffusion Models},
  author={Chen, Ruoxi and Jin, Haibo and Chen, Jinyin and Sun, Lichao},
  journal={arXiv preprint arXiv:2311.12066},
  year={2023}
}