Multiscale modeling of polymers exchanges information between coarse and fine representations of molecules to capture material properties over a wide range of spatial and temporal scales. Restoring details at a finer scale requires us to generate information following embedded physics and statistics of the models at two different levels of description. Techniques designed to address this persistent challenge balance among accuracy, efficiency, and general applicability. In this work, we present an image-based approach for structural backmapping from coarse-grained to atomistic models with cis-1,4 polyisoprene melts as an illustrative example. Through machine learning, we train conditional generative adversarial networks on the correspondence between configurations at the levels considered. The trained model is subsequently applied to provide predictions of atomistic structures from the input coarse-grained configurations. The effect of different data representation schemes on training and prediction quality is examined. Our proposed backmapping approach shows remarkable efficiency and transferability over different molecular weights in the melt based on training sets constructed from oligomeric compounds. We anticipate that this versatile backmapping approach can be readily extended to other complex systems to provide high-fidelity initial configurations with minimal human intervention.