A generative adversarial neural network (GAN) was made using a training dataset of 9780 facial images and trained in six configurations over 20,000 epochs in a batch-mode format generating fake facial images and attempting their detection. Learning rate, batch size andactivation function of hidden layers were altered to determine best model performance. Modest batch sizes of 2-3% of training data with smaller learning rate (0.0001) and Leaky-ReLU activation functions, which avoided the ‘dead neuron’ problem, outperformed other configurations.