semi-truths/Semi-Truths-Evalset

Semi-Truths: The Evaluation Sample

Recent efforts have developed AI-generated image detectors claiming robustness against various augmentations, but their effectiveness remains unclear. Can these systems detect varying degrees of augmentation?

To address these questions, we introduce Semi-Truths, featuring 27,600 real images, 245,300 masks, and 850,200 AI-augmented images featuring varying degrees of targeted and localized edits, created using diverse augmentation methods, diffusion models, and data distributions. Each augmented image includes detailed metadata for standardized, targeted evaluation of detector robustness.

🚀 Leverage the Semi-Truths dataset to understand the sensitivities of the latest AI-augmented image detectors, to various sizes of edits and semantic changes!

📂 NOTE: This is a subset of the Semi-Truths dataset created for ease of evaluation of AI-Augmented image detectors. For users with memory contraints or initial exploration of Semi-Truths, we recommend using this dataset. For the full dataset, please see semi-truths/Semi-Truths.

Directions

🌅 I want to use the Semi-Truths dataset to evaluate my detector!

• The metadata.csv file organizes all image file information under columns image_id and image_path.
• Leverage this information to pass both real and fake images to the detector you're evaluating.
• Append the detector predictions to the metadata file.
• Our metadata contains data attributes and various change metrics that describe the kind of augmentation that occured.
• By grouping predictions and computing metrics on images defined by a type of augmentation, you can gauge the specific strengths and weakness of the detecor!

To leverage our evaluation and analysis protocols, please visit our Github at: [Coming Soon! ⏳]

Dataset Structure

The general structure of the Semi-Truths Dataset is as follows:

• The original, real image and mask data can be found in the folder original
• Augmented images created with Diffusion Inpainting are in inpainting
• Prompt-edited images are in the folder p2p
• Prompt-edited image masks, computed post-augmentation, are in the folder p2p_masks
• All metadata can be found in metadata.csv, including labels, datasets, entities, augmentation methods, diffusion models, change metrics, and so on.

├── metadata.csv (Image, Mask, and Change Information)
├── original (Real Images/Mask Pairs)
│   ├── images
│   │   ├── ADE20K
│   │   ├── CelebAHQ
│   │   ├── CityScapes
│   │   ├── HumanParsing
│   │   ├── OpenImages
│   │   └── SUN_RGBD
│   └── masks
│       ├── ADE20K
│       ├── CelebAHQ
│       ├── CityScapes
│       ├── HumanParsing
│       ├── OpenImages
│       └── SUN_RGBD
├── inpainting (inpainted augmented images)
│   ├── ADE20K
│   ├── CelebAHQ
│   ├── CityScapes
│   ├── HumanParsing
│   ├── OpenImages
│   └── SUN_RGBD
└── p2p (prompt-based augmented images)
    ├── ADE20K
    ├── CelebAHQ
    ├── CityScapes
    ├── HumanParsing
    ├── OpenImages
    └── SUN_RGBD

How to download Semi Truths?

You can download the whole dataset Semi Truths by cloning the dataset using the command:

  git clone https://huggingface.co/datasets/semi-truths/Semi-Truths-Evalset