Introduction to Conditional GANs

The MNIST GANs we developed earlier generated a wide variety of output images. This was good because a constant challenge for designing GANs is avoiding low diversity and mode collapse.

Why conditional GAN?

It would be useful if we could somehow encourage our GAN to produce images that were diverse, but also restricted to one class of the training data. We could then ask our GAN to produce different images of the digit 3 for example. If we were training with face images, we might be able to ask a GAN to produce only happy faces as long as emotion was a class in the training data.