Comment by naasking

Using pixels is still tokenizing. What's needed is something more like "Byte Latent Transformers", which has dynamically sized patches based on information content rather than tokens.