Dataset Instruction

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.

Face Dataset

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