Are multi-format embeddings — like this one, or simpler ones like text-to-image — inherently bijective? I.e. could a model like this be configured to use any of its trained sense data types as either an input or output? Or is there some bias in the way the training data is ingested that makes the model better at transforming in certain directions?