Pose Guided Person Image Generation

generate an image conditioning on reference image and a specified pose

• coarse to fine: blurry result by L1 loss refined via adverserial training
  • better and faster: easy task to learn
  • predict a difference map: speed up training
  • true/fake pair: input pose / real pose
    • avoid trival solution: output input image in stage 2, maybe?
  • leverage L1 and adverserial loss
    • the latter causes artifacts