Embedding Training always makes me female

Using automatic, I am embedding training, using text file descriptors per image, but consistently I am being made female.

The text descriptions are all accurate. What am I missing here? I've tried it with multiple models.