The pair of training images in this case were the two halves of a face. The task of the neural net was to complete half of a face based on its symmetrical relationship. As in all the cases, the result is an obsessive network which, in response to any input, only generates what it knows. The faces imagined by this network are a kind of average of all the faces it has seen, modified by the characteristics of every input image.
Recursive: initial input half face
Recursive: initial input black image