CycleGAN

Transferring characteristics from one image to another is an exciting proposition. How cool would it be if you could take a photo and convert it into the style of Van Gogh or Picasso! Or maybe you want to put a smile on Agent 42's face with the virally popular Faceapp These are examples of cross domain image transfer - we want to take an image from an input domain $D_i$ and then transform it into an image of target domain $D_t$ without necessarily having a one-to-one mapping between images from input to target domain in the training set. Relaxation of having one-to-one mapping makes this formulation quite powerful - the same method could be used to tackle a variety of problems by varying the input-output domain pairs - performing artistic style transfer, adding bokeh effect to phone camera photos, creating outline maps from satellite images or convert horses to zebras and vice versa!! This is achieved by a type of generative model, specifically a Generative Adversarial Network dubbed CycleGAN by the authors of this paper.