Generating realistic electron microscopy (EM) images has been a challenging problem due to their complex global and local structures. Isola et al. proposed pix2pix, a conditional Generative Adversarial Network (GAN), for the general purpose of image-to-image translation; which fails to generate realistic EM images. We propose a new architecture for the discriminator in the GAN providing access to multiple patch sizes using skip patches and generating realistic EM images.