Conditional GAN architectures such as cGAN learn to generate fake images with a specific condition. For example the generator can be forced to generate images of people smiling, with or without glasses or with a specific hair color. But in order to achieve this, the GAN must be trained alongside with the conditional information. What if you are given a fixed trained GAN that was not trained with additional condition information, is it still possible to force the generator to generate images with specific features? according to this paper: Interpreting the Latent Space of GANs for Semantic Face Editing, YES you can! The paper explores how different features are encoded in the latent space and show how by moving around this space you can control the features of a generated fake image.
The full post is here
Adding a smile to an image:
Removing a smile from an image:
Changing hair color to black:
