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

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.