Actually, he used random crops and selected the highest rated. A "deep dream selfie" would actually run the neural network in reverse so as to generate a completely different image.