0
$\begingroup$

In the controllable image synthesis, we are manipulating a noise vector z such that our generator ( in our GAN model ) creates images that the desired feature exists. For instance, take the feature of "smiling". In order to learn how to change the z (noise) vector , the technique is described as following :

  1. Train a classifier ( besides the GAN model, just a single discriminator neural networks as a classifier) that classifies an face image according whether there is a smiling person on the image or not.

  2. Give the image generated by the GAN model with the initial noise ( say z1) to the classifier.

  3. Compute gradients of the score with respect to z1. And finally Perform gradient ascent on z1 vector in the direction of improving the score.

I cannot understand how to gradients are used to change the z1 vector in a way that it now leads the GAN model to create "smiling" faces, could you please help ?

Improve this question
$\endgroup$
2
  • $\begingroup$ It's unclear whether you're only interested in the GAN or some other model. Can you please edit your post to clarify this? I would recommend that you focus on a specific model. $\endgroup$
    – nbro
    Commented Jul 10, 2022 at 11:17
  • $\begingroup$ What do you mean by "GAN or some other model", could you please explain a little bit I did not understand ? There is both a GAN model and another Discriminator model ( not the GAN's discriminator instead just a single discriminator neural network ) in the question. $\endgroup$
    – levitatmas
    Commented Jul 10, 2022 at 16:43

0

Reset to default

You must log in to answer this question.