Comment by sabareesh

It might be that our current tokenization is inefficient compared to how well image pipeline does. Language already does lot of compression but there might be even better way to represent it in latent space