Tokenizer, serving as a translator to map the intricate visual data into a compact latent space, lies at the core of visual generative models.

Giventheir remarkable representational capacity [1], we choose deep neural networks as the function in our work.