I’ve been experimenting with various neural network architectures for image generation, including VAEs (Variational Autoencoders) and GANs (Generative Adversarial Networks).

The video here is a timelapse of a GAN training over the course of about 2 days. Each tile is an image generated by the network, with the complete image representing 36 different generated images (6×6 grid).

Code is based on work by Animesh Karnewar, repo found here, original paper here.