By plotting nonlinear trajectories through a GAN’s latent space, the method enables certain image attributes to vary while others are held fixed.