Module google/progan-128/1

Progressive GAN trained on CelebA for 128x128 images.

Module URL: https://tfhub.dev/google/progan-128/1

Open Colab notebok

Overview

Image generator based on tensorflow reimplementation of Progressive GANs[1].

Maps from a 512-dimensional latent space to images. During training, the latent space vectors were sampled from a normal distribution.

Module takes <Tensor(tf.float32, shape=[?, 512])>, representing a batch of latent vectors as input, and outputs <Tensor(tf.float32, shape=[?, 128, 128, 3])> representing a batch of RGB images.

Example use

# Generate 20 random samples.
generate = hub.Module("https://tfhub.dev/google/progan-128/1")
images = generate(tf.random_normal([20, 512]))

Training

The original model has been trained on a GPU for 636,801 steps with batch size 16.

References

[1] Tero Karras, Timo Aila, Samuli Laine, Jaakko Lehtinen. Progressive Growing of GANs for Improved Quality, Stability, and Variation. arXiv preprint arXiv:1710.10196, 2017.