Sym-parameterized Dynamic Inference for Mixed-Domain Image Translation

  • 2019-04-12 15:41:01
  • Simyung Chang, SeongUk Park, John Yang, Nojun Kwak
  • 0


Recent advances in image-to-image translation have led to some ways togenerate multiple domain images through a single network. However, there isstill a limit in creating an image of a target domain without a dataset on it.We propose a method that expands the concept of `multi-domain' from data to theloss area and learns the combined characteristics of each domain to dynamicallyinfer translations of images in mixed domains. First, we introduceSym-parameter and its learning method for variously mixed losses whilesynchronizing them with input conditions. Then, we propose Sym-parameterizedGenerative Network (SGN) which is empirically confirmed of learning mixedcharacteristics of various data and losses, and translating images to anymixed-domain without ground truths, such as 30% Van Gogh and 20% Monet and 40%snowy.


Introduction (beta)



Conclusion (beta)