Example-Guided Style-Consistent Image Synthesis from Semantic Labeling

less than 1 minute read

Published:

image synthesis from semantic label map and exemplary image

  • directly using pix2pixHD will not keep style well
  • multiscale discriminator, feaure matching loss
  • style consistenct loss on pair of images
    • dataset constructed from video
  • semantic consistency loss
    • from VGG feature map
    • weighted perceptual loss: weight=1 for content consistent images, weight=1/M for style consistent images