Improving the training of generative adversarial networks by identifying the input and output layers in the generator network, assigning a gradient multiplier to each layer, and multiplying the gradient by the multiplier during back-propagation and before applying the product of the gradient and multiplier to node weights of the layers.