In this work we consider the problem of generating aesthetically pleasing photography, sometimes termed photographic fine art (PFA).
We cast this problem as a generative modeling task and use a conditional GAN framework. Recent works have shown that conditioning based on semantic information is beneficial for improving photo-realism.
In this work we propose a novel GAN architecture which is able to generate photo-realistic images with a specified aesthetic quality by conditioning on both semantic and aesthetic information. To condition the generator, we propose a modified conditional batch normalization layer. To condition the discriminator, we use a joint probabilistic model of semantics and aesthetics to estimate the compatibility between an image (either real or generated) and the conditioning variable.
We show quantitatively and qualitatively that our model, called PFAGAN, is able to generate images conditioned on semantic categories and aesthetic scores.