search

umd-huang-lab / WAVES

Code for our paper "Benchmarking the Robustness of Image Watermarks"
42 stars 2 forks source link

Can you tell me how the dataset should be organized in this code #6

Closed xiaoxiaostudy closed 4 months ago

xiaoxiaostudy commented 5 months ago

Can you tell me how the dataset should be organized in this code? For example, I randomly selected 5,000 images from diffusiondb, how should I organize my reference images and prompt strings?Like that? --datasets --main --diffusiondb --image1 --image2 --attacked

johnding1996 commented 4 months ago

Hi, thank you for your interest and for reaching out. Apologies for the delayed response.

For non-attacked images, which include both the original and watermarked images, we suggest the following directory structure. The dataset names diffusiondb and mscoco are used here as examples. The directory real contains non-watermarked, original images, and <watermark_method> should be replaced with your specific watermarking technique.

main
├── diffusiondb
│   ├── prompts.json
│   ├── real
│   ├── <watermark_method>
└── mscoco
    ├── prompts.json
    ├── real
    └── <watermark_method>

For attacked images, please use the same structure:

attacked
├── diffusiondb
│   ├── <attack_method>-<attack_strength>-<watermark_method>
└── mscoco
    └── <attack_method>-<attack_strength>-<watermark_method>

Please configure your .env file to specify the parent directory of both main and attacked folders as follows:

DATA_DIR=/path/to/datasets

This organization will facilitate correct file management and accessibility within your project.

xiaoxiaostudy commented 4 months ago

Thank you for your response! I really appreciate your work!