Similar to the generator network, the discriminator network is a deep convolutional neural network and contains extra downsampling layers as the image is of a larger size than the discriminator network in Stage-I. The discriminator is a matching-aware discriminator (more information about which can be found at the following link: https://arxiv.org/pdf/1605.05396.pdf), which allows us to achieve better alignment between the image and the conditioning text. During training, the discriminator takes real images and their corresponding text descriptions as positive sample pairs, whereas negative sample pairs consist of two groups. The first group is real images with mismatched text embeddings, while the second is synthetic images with their corresponding text embeddings. Let's have a look at the architecture of the discriminator network, as shown in the following images:

More information about the architecture of the discriminator network can be found in the Keras implementation of StackGAN section.