Instead of taking noise as an input, the generator in stage II takes the image generated from the previous stage as an input and it is conditioned on the text description.

Here, implies that we are sampling from the . It basically means that we are sampling the image generated from stage I.

implies that we are sampling text from the given real data distribution, .

Then the generator loss can be given as follows:

Along with the regularizer, our loss function of the generator becomes: