| Domain | Dataset | Generator | Height | Width | Comment |
|---|---|---|---|---|---|
| Face | FFHQ | StyleGAN2* convert |
1024 | 1024 | |
| Face | CelebA | - | 1024 | 1024 | evaluation |
| Cat | AFHQ Cat | StyleGAN2-ADA | 512 | 512 | |
| Cat | LSUN Cat | StyleGAN2* convert |
256 | 256 | |
| Dog | AFHQ Dog | StyleGAN2-ADA | 512 | 512 | |
| Wild Animal | AFHQ Wild | StyleGAN2-ADA | 512 | 512 | |
| Horse | LSUN Horse | StyleGAN2* convert |
256 | 256 | |
| Car | Stanford Car | - | train inversion | ||
| Car | LSUN Car | StyleGAN2* convert |
384 | 512 | Crop |
| Church | LSUN Church | StyleGAN2* convert |
256 | 256 |
*: weight is from the original [StyleGAN2](NVlabs/stylegan2: StyleGAN2 - Official TensorFlow Implementation (github.com)) (TensorFlow-based), which needs to convert by script. We also provide the converted weights, which are converted by this implementation.
We use FFHQ (70,000 images) for training and CelebA-HQ test dataset (2824 images) for testing.
Download script can be found in NVlabs/ffhq-dataset.
By default, the image path follows: data/FFHQ/xxxxx.png.
CelebA-HQ is a subset of CelebA dataset, we share the test split (2824 images) on drive.
By default, the image path follows: data/CelebA-HQ/test/xxxxxx.jpg.
AFHQ consists of three categories: cat, dog, and wild. There are two versions (i.e., AFHQ and AFHQv2), and we use AFHQ by default.
TODO
TODO