This is the repository containing the code for our master’s thesis. We study three multi-stage GANs: AttnGAN, MirrorGAN, DM-GAN, and contribute with the fusion of key ideas from these methods. Our final proposed model is called Dynamic Memory Mirrored Generative Adversarial Network (DMM-GAN). The DMM-GAN architecture is best illustrated via its modifications made in relation to AttnGAN. Figure 1 and Figure 2 together illustrate these modifications.

The details of STREAM and GLDM are shown in the following illustrations.

We credit Zhang et al. (https://github.com/hanzhanggit/StackGAN-v2) , Zu et al. (https://github.com/taoxugit/AttnGAN), Qiao et al. (https://github.com/qiaott/MirrorGAN), and Zhu et al. (https://github.com/MinfengZhu/DM-GAN) for providing implementations of their respective methods. Our code is based on the github repositories related to these methods.