Module google/compare_gan/model_9_celebahq128_resnet19/1

ResNet19 trained on CelebA HQ (128x128) (FID: 29.13).

Module URL:

Open Colab notebok


ResNet19 generator and discriminator.

For the details of the setup, please refer to [1]. The code used to train these models is available on GitHub. View all available compare_gan modules in the Colab notebook.


  • Dataset: CelebA HQ (128x128)
  • Model: Non-saturating GAN
  • Architecture: ResNet19
  • Optimizer: Adam (lr=1.000e-04, beta1=0.500, beta2=0.900)
  • Discriminator iterations per generator iteration: 5
  • Discriminator normalizaton: Layer normalization
  • Discriminator regularization: DRAGAN Gradient Penalty (lambda=1.000)


  • FID: 29.13
  • Inception: 2.34
  • MS-SSIM: 0.30

Example use

# Declare the module
gan = hub.Module("")

# Use the generator signature
z_values = tf.random_uniform(minval=-1, maxval=1, shape=[64, 128])
images = gan(z_values, signature="generator")

# Use the discriminator signature
logits = gan(images, signature="discriminator")

# Drive execution with tf.Session[images, logits])


[1] Karol Kurach, Mario Lucic, Xiaohua Zhai, Marcin Michalski, Sylvain Gelly. The GAN Landscape: Losses, Architectures, Regularization, and Normalization.