Pre-trained visual networks can develop a shared protocol for referential communication, enabling effective communication despite architectural differences.