Conditional Invertible Neural Networks for Guided Image Generation
In this work, we address the task of natural image generation guided by a conditioning input. We introduce a new architecture called conditional invertible neural network (cINN). It combines the purely generative INN model with an unconstrained feed-forward network, which efficiently pre-processes the conditioning input into useful features. All parameters of a cINN are jointly optimized with a stable, maximum likelihood-based training procedure. Even though INNs and other normalizing flow models have received very little attention in the literature in contrast to GANs, we find that cINNs can achieve comparable quality, with some remarkable properties absent in cGANs, e.g. apparent immunity to mode collapse. We demonstrate these properties for the tasks of MNIST digit generation and image colorization. Furthermore, we take advantage of our bidirectional cINN architecture to explore and manipulate emergent properties of the latent space, such as changing the image style in an intuitive way.
PDF Abstract
