Generative Adversarial Networks (GANs) have a great performance in image
generation, but they need a large scale of data to train the entire framework,
and often result in nonsensical results. We propose a new method referring to
conditional GAN, which equipments the latent noise with mixture of Student's
t-distribution with attention mechanism in addition to class information.
Student's t-distribution has long tails that can provide more diversity to the
latent noise. Meanwhile, the discriminator in our model implements two tasks
simultaneously, judging whether the images come from the true data
distribution, and identifying the class of each generated images. The
parameters of the mixture model can be learned along with those of GANs.
Moreover, we mathematically prove that any multivariate Student's
t-distribution can be obtained by a linear transformation of a normal
multivariate Student's t-distribution. Experiments comparing the proposed
method with typical GAN, DeliGAN and DCGAN indicate that, our method has a
great performance on generating diverse and legible objects with limited data.