Blending Two Styles: Generating Inter-Domain Images with MiddleGAN

Abstract

From celebrity faces to cats and dogs, humans enjoy pushing the boundaries of art by blending existing concepts together in new ways. With the rise of generative artificial intelligence, machines are increasingly capable of creating new images. Generative Adversarial Networks (GANs) generate images similar to their training data but struggle to blend images from distinct datasets. This paper introduces MiddleGAN, a novel GAN variant that blends inter-domain images from two distinct input sets. By incorporating a second discriminator, MiddleGAN forces the generator to create images that fool both discriminators, thus capturing the qualities of both input sets. We also introduce a blend ratio hyperparameter to control the weighting of the input sets and compensate for datasets of different complexities. Evaluating MiddleGAN on the CelebA dataset, we demonstrate that it successfully generates images that lie between the distributions of the input sets, both mathematically and visually. An additional experiment verifies the viability of MiddleGAN on handwritten digit datasets (DIDA and MNIST). We provide a proof of optimal convergence for the neural networks in our architecture and show that MiddleGAN functions across various resolutions and blend ratios. We conclude with potential future research directions for MiddleGAN.