From a machine learning standpoint, I would guess that this is achieved with a partially supervised approach. The idea would be to embed an entire face dataset

From a machine learning standpoint, I would guess that this is achieved with a partially supervised approach. The idea would be to embed an entire face dataset from a number of different poses to a latent space by some method (could be as simple as PCA, or as complex as an arbitrary neural network). An arbitrary classification method (using a partial labelling of faces by gender) yields a hyperplane which optimally separates the labelled points. Then, something like a simple reflection across this hyperplane yields another point in the latent space, and the generator yields a corresponding face of the opposite gender. Anyone know if this has been implemented for a task like this?