Publications

Ji Liu*, Heshan Liu*, Mang-Tik Chiu, Yu-Wing Tai, Chi-Keung Tang. High-Resolution Appearance Transfer Network (APT) (*Equal contributions)

  • Project demo here
  • Our contribution consists of a new encoder-decoder architecture that successfully enables appearance transfer to a target pose.
    • We propose novel local descriptors (progressive local perceptual loss + local discriminators at the highest resolution to enhance local details and generation quality.
    • We apply progressive training to our autoencoder architecture to achieve outputs at unprecedented high resolution (1024 x 1024).
  • Results: We can transfer in high resolution on Human3.6M, DeepFashion and a self-collected YouTube dataset.