can someone give an intuition for the unified tokenizer that they mentioned? I understand the idea of using a GAN or an autoencoder to make a image tokenizer, but it's unclear how you combine that with text to have a single tokenizer over both modalities (if I'm understanding the paper correctly)