Disentangled Makeup Transfer with Generative Adversarial Network

task: disentangle identity and makeup

• an encoder for each task
  • assumption: identity and makeup independent of each other
  • perceptual loss for personal identity preservation
  • histogram matching on diff. regions, then L_2 loss for makeup transfer
• attention mask for each key part
  • face, brow, eye, lip