Colab Notebooks: BigGAN

BigGAN was one of the first prominent Generative Adversarial Networks and has already rather historical that qualitative relevance. Trained on ImageNet at a now-humble 128x128 resolution this Network became a standard by its manifold generative abilities.

In this notebook, you can generate samples from a long list of categories.

The selection of the category “Valley” (#979) generates a photorealistic Valley image series. 

This is how BigGAN imagines a "Clockwork":

You clearly see: the conception of something "roundish", with "arrows" and "characters" is still preserved. Still, BigGAN doesn't bear any interpretation of a clock.

Here are some AI generated comics:

And this is an interesting thing: "Envelopes":

Here you see it again: labeling is essential. BigGAN is trained on a human-labeled dataset of images. Probably most of the images with sketches and drafts were labeled as "envelopes". The Bias begins already here. And this is not an issue of AI. It's our problem.

Links:

• BigGAN paper on arXiv (Andrew Brock, Jeff Donahue, and Karen Simonyan. Large Scale GAN Training for High Fidelity Natural Image Synthesis. arxiv:1809.11096, 2018)
• Colab Notebook

Things to try out:

The same notebook allows you to create interpolation between images. This approach was — and is — mindblowing and innovative since only Surrealists and Abstract Artists were previously so courageous to combine incompatible things. Now you can do it with photorealistic results, for example, generate an interpolation between a Yorkshire terrier and a Space shuttle.

Read also:

Index of Series "Google Colab Notebook".

Full essay "12 Colab Notebooks that matter"