Transform the Set: Memory Attentive Generation of Guided and Unguided Image Collages