Training the discriminator with real images