Capture and control content discrepancies via normalised flow transfer

Unsupervised Style Transfer (UST) has recently been a hot topic in Computer Vision, and this type of work has been exemplified by CycleGAN. Although the existing UST methods have proven to be useful, in many circumstances, we want to be able to manage not just the transformation of a single piece of instance, but also the morphological features of the two data sets’ underlying distributions. To this end, we propose a novel framework called Normalised Flow Transfer (NFT), where a reversible probability transform using the normalised flow method is developed to transfer the data in the first domain to the second, so as to exhibit their probabilities under both domains in the mapping process. A particularly interesting application under this framework is that when the data sets in two domains contain numerous clusters based on their finite class labels, we can control the distribution pattern across the two domains and apply any constraints on the underlying distributions of the same class. The experimental results show that not only can we devise many new complex style transfer functions, but also our framework has better image generation capabilities in terms of evaluation metrics, including mean square error, inception score and Frechet inception distance.