You can use an image to guide the direction of a dream. Normally, deepdream is trying to maximise its recognition in the scene. For datasets trained on lots of types of picture, it can recognise and draw all sorts of things - faces, buildings, arms, legs, wings and tentacles. The guide image tells the network to maximise certain parts of the recognition - those triggered by the guide image.

I used Lucid Dream, using the original ImageNet dataset. I used layer inception_4d/output (see here for a comparison of layers) for 50 iterations.

My favourite is the bears.

Unguided Dream