This is quite a difficult task and is still an open area of research. The idea behind the GANs is to map latent features (e.g. rotation, age) to an output image, but the problem is that the source data comes from an unknown probability distribution. GANs aim to approximate it by sampling this latent feature vector from a simple known distribution (e.g. normal).
However, such a distribution usually won't fit the target distribution and hence the latent features become entangled as they will reproduce statistics of the real data (e.g. gender and beard)
There are several attempts to solve this problem. One of them is Conditional GAN. For example, with a conditional GAN it can be possible to translate a segmentation mask into an image (Pix2Pix). Another approach has been proposed in StyleGAN. The authors introduce an intermediate feature vector that the model learns to disentangle in an unsupervised manner. Each value in this intermediate represents a meaningful attribute of the resulting image, but it is still not possible to control the latent features in a supervised manner. That is, after training you should manually test each of them to identify its relations.