---
license: cc-by-nc-sa-4.0
viewer: false
task_categories:
- image-to-image
---

# NullFace: Training-Free Localized Face Anonymization

[HF中国镜像站 Paper](https://huggingface.co/papers/2503.08478)

## Test set

We curated a subset of test subjects from the [CelebA-HQ](https://github.com/tkarras/progressive_growing_of_gans) and [FFHQ](https://github.com/NVlabs/ffhq-dataset) datasets for the quantitative comparisons against baseline methods in our paper. Specifically, we selected:
- 4,852 images from [CelebA-HQ](https://github.com/tkarras/progressive_growing_of_gans)
- 4,722 images from [FFHQ](https://github.com/NVlabs/ffhq-dataset)

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:

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