Celebrity Face GANs

A deep learning project demonstrating how latent space arithmetic works in GANs. Given celebrity photos of men and women, with and without glasses, we showed that you can generate a man with glasses by taking the latent vector of a woman with glasses, subtracting the vector of a woman without glasses, and adding the vector of a man without glasses. This vector arithmetic in the z-space lets you isolate and transfer specific facial attributes. We also used CycleGANs to perform unpaired style transfer, converting male faces to female and vice versa.