The DCGAN authors proposed the following tips for constructing the structure of each model:
Guidelines by the authors of the DCGAN paper. Reference: https://arxiv.org/pdf/1511.06434.pdf
For the generator, you can add additional layers in the network at the cost of training time and the additional chance of divergence. For the discriminator, you can build a state-of-the-art binary classifier, but this will likely create an increased risk of divergence.
Essentially, you need to experiment with this architecture based on your own application's needs and figure out what the best parameters for your network are.