BeautyGlow On-Demand Makeup Transfer Framework with Reversible Generative Network

task: generate a specific makeup style

• needs a encoder to interpolate latent spaces, which doesnot exist in GAN
• BeautyGlow: decomposes latent vector into makeup features and facial identity features
  • in the same space
  • feature of image = W * (makeup) + (I - W) * (identity)
  • then switch features, gen. images
  • Glow based method
• perceptual loss
• makeup loss
  • makeup vector should be close to center of all makeup vectors
  • before/after-makeup vector should be close to center of its classes
    • kind of split the space into two halves, one for identity one for makeup