2 discriminators in CycleGAN?

The standard CycleGan involves two generators and two discriminators (0, 1). And I am a little bit unsure about the choice of two discriminators if the discrimination is -1 (true X) <--> 0 (Fake) <---> 1 (true Y) and only using 1 discriminator. For such, this formulation can speed up the training and lowering the memory requirement, as it just eliminates an entire network.