NullFace: Training-Free Localized Face Anonymization

Hugging Face Paper

Test set

We curated a subset of test subjects from the CelebA-HQ and FFHQ datasets for the quantitative comparisons against baseline methods in our paper. Specifically, we selected:

• 4,852 images from CelebA-HQ
• 4,722 images from FFHQ

For each test subject, we generated a corresponding segmentation mask, which is designed to keep the eye and mouth areas visible when needed.

The test subject lists (JSONL format) and segmentation masks can be downloaded in the following structure:

nullface-test-set/
├── celeba-hq/
│   ├── mask_images/
│   │   ├── 00010.png
│   │   └── ...
│   └── metadata.jsonl
├── ffhq/
│   ├── mask_images/
│   │   ├── 00010.png
│   │   └── ...
│   └── metadata.jsonl
└── README.md