Conditional GANs are an extension of the normal GANs, where both the discriminator and the generator are conditioned to some specific class y. This has interesting applications, because you can fix your generator to a specific class and then make it produce multiple different versions of a specific same class of our choosing. For example, if you set y to be the label corresponding to the digit 7 in MNIST, the generator will only produce images of sevens.

With conditional GANs, the minimax game becomes the following:

Here, we depend on the extra input y, which is the class label of the input image.

The simplest way to merge x and y, or z and y, is to just concatenate them together so that our input vector is longer. This creates a much more controlled dataset augmentation system. Here's an example of this in TensorFlow code: