Generate Novel Image Styles using Weighted Hybrid Generative Adversarial Nets

In recent years, Generative Adversarial Networks (GANs) have achieved significant improvements in image processing, especially image-to-image translation problem and image generation. But, few works are presented to creatively produce a novel domain from many training datasets with different domains. Inspired by creating a new calligraphic style, we propose a novel GAN model that supports creatively generate data domain, such as context, style and so on. In this paper, we call it as WHGAN. What is one key innovation is that WHGAN brings in a discriminative set (contains k discriminative models) that each one is responsible for a training dataset, in addition, the single generative model obtains feedbacks from discriminative models and produces a novel data distribution. Relatively, each discriminative model distinguishes the generated data distribution from its corresponding input dataset. Meanwhile, in order to make the generated data adjustable, we redesign the objective function with a set of variable weights that each one is responsible for a discriminator. For ease of presentation, we set k to be 2 in our implementation. Then, we conduct two evaluation on image dataset and synthesized 2D dataset respectively. Results show that WHGAN successfully generates oil-painting style images from photo-realistic and cartoon style inputs, furthermore, we also visually and objectively verify the impact of weights.