Vector-Quantized Variational Autoencoders

Vector-Quantized VAEs were proposed in 2017. Since its inception, it has pushed the field of high-quality image generation to a great extent. Its recipes like discrete latent space optimization, codebook sampling, etc. have gone to later become essential blocks for modern models like VQ-GAN, DALL-E, etc.

In my latest Keras example, I present an implementation VQ-VAEs including the subsequent PixelCNN part for image reconstruction and generation. I’ve included crucial pieces of visualizations as well to make it fun and interesting.

Here is the link to my example:


Well done Sayak!
always great content!

1 Like