Abstract

Modeling daytime changes in high resolution photographs, e.g., re-renderingthe same scene under different illuminations typical for day, night, or dawn,is a challenging image manipulation task. We present the high-resolutiondaytime translation (HiDT) model for this task. HiDT combines a generativeimage-to-image model and a new upsampling scheme that allows to apply imagetranslation at high resolution. The model demonstrates competitive results interms of both commonly used GAN metrics and human evaluation. Uniquely, thisgood performance comes as a result of training on a dataset of still landscapeimages with no daytime labels available. Our results are available athttps://saic-mdal.github.io/HiDT/.

 

Quick Read (beta)