Colab Notebooks: StyleGAN2

What’s this about?

This network, developed by NVidia, is at the moment (6th March 2020, 4:27 pm) the most advanced generative network for images. It was trained on High-Definition-Datasets (for example, Faces from Flickr-Faces-HQ). StyleGAN2 provides automatically learned, unsupervised separation of high-level attributes, stochastic variation and control of layers with visual features.

There are various StyleGAN2-Notebooks (benefits of crowdsourced research), but my favorite is finetuned by Mikael Christensen.

Links:

• Paper: https://arxiv.org/abs/1812.04948
• Video: https://youtu.be/kSLJriaOumA
• Code: https://github.com/NVlabs/stylegan / https://github.com/NVlabs/stylegan2
• Colab Notebook

Things to try out:

• There are various default datasets by NVidia available within the notebook (mind the resolution):

• Try out new Datasets. Train your own models or use those, provided by various artists and researchers, like Michael Friesen (follow his twitter for new updates).

Bacteria StyleGAN

Art-StyleGAN:

MC Escher-StyleGAN:

• Produce videos of interpolations:

• Try out StyleGAN2 projection. With the StyleGAN2 notebook you discover (or better: re-cover) images being hidden in the Latent Space of the Network. StyleGAN Projection is a method to trace back StyleGAN2-generated images — so you can detect a photo as an AI-generated product (#DeepFake debunk). It still has some downsides: you have to know the concrete dataset of the particular image; every change in the image will make the process of projection unfeasible.

Here is a successful projection:

And this one will probably deprive you of sleep next night:

Read more:

If you want to try a bigger diversity of motives, try out ArtBreeder, a powerful image generator by Joel Simon:

Index of Series "Google Colab Notebook".

Full essay "12 Colab Notebooks that matter"