Diffusion magnetic resonance imaging (DMRI) is a key tool for non-invasively studying brain microstructure and connectivity, but its clinical utility is limited by low signal-to-noise ratio, long scan times, and hardware constraints. While recent deep learning methods have improved DMRI quality through denoising and non-linear mapping, they lack flexibility and efficiency. In this paper, we present FastDTI, a 3D scale-arbitrary super-resolution autoencoder with a residual dense network architecture for diffusion tensor imaging. Using curriculum learning, FastDTI enables non-integer upscaling and produces high-fidelity clinical maps—mean diffusivity (MD), fractional anisotropy (FA), and principal diffusion direction (D-Maps) from just six diffusion-weighted images. On paired 3T/7T HCP Young Adult data, FastDTI delivers a 35% PSNR improvement over traditional methods and achieves a 6x inference speedup over state-of-the-art deep learning models via its single-pass design, combining both accuracy and efficiency.