We have developed and validated an ultra-quality 4D-MRI synthesis technique using deep learning-based deformable image registrations. The displacement vector fields between breathing frames were obtained from low-quality 4D-MRI. They were then applied to high-quality stationary T1, T2, and diffusion weighted images to generate ultra-quality 4D-MRI. The synthetic 4D-MRIs were verified in terms of tumor motion accuracy and image quality. All the motion errors were in a sub-voxel level, and the image quality was significantly improved. This technique holds great potential in volumetric tumor tracking with high accuracy.