This is some interesting jupyter notebook sample code dlnd_face_generation.ipynb.
Here I create a Deep Convolutional GAN (DCGAN). I train it on pictures of celebrities, then generate a totally new face from noise. The output is a bit crude, mainly due to the low resolution of the output images (in my opinion). But it's an amazing concept.
I get the celebrity images from CelebA.html. You can get pre-processed image set from here processed-celeba-small.zip. Here the images are reduced to 64x64x2 and cropped to the faces only.
This is what some of the celebrity images look like from the original training set.
And this is what some of the faces generated from "noise" look like.
The entire project is zipped project-face-generation.zip and archived archive.zip if interested.