Conditional Generative Adversarial Network output on non-trained condition

vision
1

I am new and currently learning about GANs, I have a query as follows:

Suppose I have trained my conditional GAN on MNIST dataset where I have conditioned on gaussian noise with different means. Ex:

Gaussian noise with mean 1 will give output as image of 1.

Gaussian noise with mean 2 will give output as image of 2. and so on.

After training , suppose I give gaussian noise with mean 1.5, what output image should I expect?

This logic I am extending for an Image to Image translation task. Suppose I have conditioned my GAN(pix2pix) model to add effect1 on input image for gaussian noise with mean 1, and add effect2 on input image for gaussian noise with mean 2. What output image can I expect if I give input image with gaussian noise with mean 1.5? will it be combination of effect1 and effect2 or can be anyone of them or some random output?

2

My guess is that the output might not be a clear interpolation between the two original classes, since your model got a new and unknown input.
Similar behavior can be seen, if you train a GAN with a latent vector of dim Z by sampling from a Gaussian distribution. If you try to directly interpolate between two valid points, your input would leave the “dense” part of the distribution (i.e. the high-dimensional bubble) and your model could output artifacts.

PS: let us know what the outcome is, in case you are running these experiments! :slight_smile: