Unpaired Image-to-Image Translation through CycleGAN

The term project of COMP511: Computer Vision with Deep Learning in Fall 2022 The CycleGAN model is built based on CycleGAN by Jun-Yan Zhu et al. Training backlogs are available in Weights & Biases.

General architecture of the CYCLEGAN model:

The difference between the CycleGAN and base CycleGAN:

Generator:

• There are nine residual blocks.
• LeakyReLU is implemented instead of ReLU.

Discriminator:

• Instance normalization is implemented after each convolutional layer.