Liang Jie ; Zeng Hui ; Zhang Lei

Abstract

Existing image-to-image translation (I2IT) methods are either constrained tolow-resolution images or long inference time due to their heavy computationalburden on the convolution of high-resolution feature maps. In this paper, wefocus on speeding-up the high-resolution photorealistic I2IT tasks based onclosed-form Laplacian pyramid decomposition and reconstruction. Specifically,we reveal that the attribute transformations, such as illumination and colormanipulation, relate more to the low-frequency component, while the contentdetails can be adaptively refined on high-frequency components. We consequentlypropose a Laplacian Pyramid Translation Network (LPTN) to simultaneouslyperform these two tasks, where we design a lightweight network for translatingthe low-frequency component with reduced resolution and a progressive maskingstrategy to efficiently refine the high-frequency ones. Our model avoids mostof the heavy computation consumed by processing high-resolution feature mapsand faithfully preserves the image details. Extensive experimental results onvarious tasks demonstrate that the proposed method can translate 4K images inreal-time using one normal GPU while achieving comparable transformationperformance against existing methods. Datasets and codes are available:https://github.com/csjliang/LPTN.