Model Zoo

Pre-trained Models

First of all, we thank the following repositories for their work on high-quality image synthesis

Please download the models you need and save them to checkpoints/.

StyleGAN Ours
Model (Dataset)	Training Samples	Training Duration (K Images)	FID
Face ("partial" means faces are not fully aligned to center)
celeba_partial-256x256	103,706	50,000	7.03
ffhq-256x256	70,000	25,000	5.70
ffhq-512x512	70,000	25,000	5.15
LSUN Indoor Scene
livingroom-256x256	1,315,802	30,000	5.16
diningroom-256x256	657,571	25,000	4.13
kitchen-256x256	1,000,000	30,000	5.06
LSUN Indoor Scene Mixture
apartment-256x256	4 * 200,000	60,000	4.18
LSUN Outdoor Scene
church-256x256	126,227	30,000	4.82
tower-256x256	708,264	30,000	5.99
bridge-256x256	818,687	25,000	6.42
LSUN Other Scene
restaurant-256x256	626,331	50,000	4.03
classroom-256x256	168,103	50,000	10.10
conferenceroom-256x256	229,069	50,000	6.20

MNIST (60,000 training samples and 10,000 test samples on 10 digital numbers)
SVHN (73,257 training samples, 26,032 testing samples, and 531,131 additional samples on 10 digital numbers)
CIFAR10 (50,000 training samples and 10,000 test samples on 10 classes)
CIFAR100 (50,000 training samples and 10,000 test samples on 100 classes)
ImageNet (1,281,167 training samples, 50,000 validation samples, and 100,100 testing samples on 1000 classes)
CelebA (202,599 samples from 10,177 identities, with 5 landmarks and 40 binary facial attributes)
CelebA-HQ (30,000 samples)
FF-HQ (70,000 samples)
LSUN (see statistical information below)
Places (around 1.8M training samples covering 365 classes)
Cityscapes (2,975 training samples, 19998 extra training samples (one broken), 500 validation samples, and 1,525 test samples)
Streetscapes

Statistical information of LSUN dataset is summarized as follows: