Towards Models that Can See and Read