GANGogh: Creating Art with GANs – Towards Data Science – Medium


StackGAN uses feature information retrieved from a Recurrent Neural Net and a two phase image generation process -- -the first phase creates a low resolution image from a noise vector and the second phase uses an encoding of the first image to create high resolution image. They also both use gated multiplicative activation functions which seem to mesh well with this global conditioning (van den Oord 2016; van den Oord 2016). In papers such as A Neural Algorithm of Artistic Style, deep learning nets learn to 1) differentiate the style of a piece of art from its content and 2) to apply that style to other content representations. We could enforce this metric by adding a penalizing term to our discriminator's cost function that tries to minimize the cross-entropy in its prediction of genre versus the real genre of a given painting, and adding a penalizing term to our generator that tries to minimize the cross-entropy of the discriminator's prediction versus the genre it was instructed to make based on the conditioning vector.