Implementation / replication of DALL-E, OpenAI's Text to Image
...In contrast to OpenAI's VAE, it also has an extra layer of downsampling, so the image sequence length is 256 instead of 1024 (this will lead to a 16 reduction in training costs, when you do the math).