Computer holography is a promising technique for reproducing 3D viewing experiences on-demand by modulating an amplitude or phase hologram. Yet, the huge computational burden of both acquiring 3D wave data and simulating diffraction propagation places an explicit challenge to the total runtime of 3D hologram generation, making dynamic holography impractical. Here we demonstrate a deep-learning based method capable of synthesizing a photorealistic 3D hologram in real-time directly from the input of a single 2D image....