Monday, April 22, 2019


  • They can make entirely new image that are realistic, even they never been seen before
  • Most of the application for GANs have been images
  • Takes a textual description of the bird and than generating a high resolution of a bird matching that description.
  • These pictures have never been seen before. It is not running a image search on a database, infact GAN is drawing a probability distribution over all hypothetical images matching that description
  • We can keep running the GAN to get more images.


  • Developed in collaboration with adobe and Berkeley to help artist
  • As the user draws a very crude sketches using the mouse,iGAN searches for the nearest possible realistic image.
  • Ex scribble of green paint is turned into a lush grassland or a black triangle turned into a detailed mountain.
  • Images in one domain can be turned into image in another domain.
  • Blueprints for buildings can be turned into photos of finished buildings.
  • Drawings of cat can be changed into a realistic photos of cats.
  • Can be trained into unsupervised way.

  • Good at unsupervised image to image translation.
  • Ex : Transforms the video of a horse to a video of a zebra.
  • As the changes are completely un supervised. It changes the surroundings as zeebra live in a different environment. The background shown below looks more like a african background

  • GANS are used for imitation learning ( ex imitate the action taken by a human expert)
  • GANs are not limited to the visual domain

No comments:

Post a Comment