Multiple acquisitions have to be averaged to achieve reasonable signal-to-noise ratio (SNR) in high-resolution diffusion tensor imaging (DTI). However, involuntary global/local motions during diffusion-sensitizing gradients create k-space shifts, and global/local phase differences between different acquisitions, complicating image reconstruction. In this work, we propose a phase-correcting joint non-local means reconstruction that effectively prevents phase cancellations and reduces noise. This technique jointly utilizes the images from different diffusion-encoding directions to preserve the fractional anisotropy (FA) map. Results are demonstrated for in vivo spinal cord DTI and on a simulated DTI dataset.