This repo contains a number of approaches for generative modeling and is continuously updated as I work with new methods.
Currently, the main model is a VQVAE[0] along with a PixelCNN++ prior. I based my JAX implementation of VQVAE on this notebook and trained it on a custom dataset containing topographic map tiles of Switzerland. PixelCNN was slightly adapted from here to work with learning the latent codes from VQVAE instead of RGB images.
Todos
-
retrain PixelCNN++ for longer for better results
-
conditional generation
-
build web demo
WIP.