High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs

less than 1 minute read

Published:

address two problems

  • difficulty of generating images of high resolution
  • lack of details and realisim

method

  • aim: diverse result given input label map
    • object-level semantic editing
  • global generator and local enhancer
    • two scales often enough
  • 3 discriminators for different scales
    • easier to train coarse to fine
    • feature matching in each layer of distriminator
  • instance-level semantic label map -> object boundry
  • instance-wise average pooling
    • how to implement?