A DiscoGAN is a generative adversarial network that generates images of products in domain B given an image in domain A. Illustrated in the following diagram is an architectural diagram of a DisoGAN network:
The images generated in domain B resemble the images in domain A in both style and pattern. This relation can be learned without explicitly pairing images from the two domains during training. This is quite a powerful capability, given that the pairing of items is a time-consuming task. On a high level, it tries to learn two generator functions in the form of neural networks GAB and GBA so that an image xA, when fed through the generator GAB, produces an image xAB, that looks realistic in domain B. Also, when this image xAB is fed through the other generator network GBA, it should produce an image xABA which should...
