We know that the role of the generator is to generate a new image by learning the real data distribution. In DCGAN, a generator is composed of convolutional transpose and batch norm layers with ReLU activations.

The input to the generator is the noise, , which we draw from a standard normal distribution, and it outputs an image of the same size as the images in the training data, say, 64 x 64 x 3.

The architecture of the generator is shown in the following diagram:

First, we convert the noise, z, of a 100 x 1 shape into 1024 x 4 x 4 to have the shape of width, height, and feature map and it is called the project and reshape. Following this, we perform a series of convolutional operations with fractionally strided convolutions. We apply batch normalization to every layer except at the last layer. Also, we apply ReLU activations to every layer but the last layer. We apply the tanh activation function to scale the generated image between -1 and +1.