Dynamic 31P-MRS/MRSI is a promising tool for in vivo quantification of mitochondrial oxidative capacity. However, its practical utility is limited by the inherently low SNR of the 31P signal. This work is built upon our recent progress in accelerating dynamic 31P-MRSI using low-rank tensor models. We extended this method by learning the temporal priors with deep generative models and then incorporating them into the reconstruction via an information theoretical framework. This approach enabled high-resolution dynamic 31P-MRSI with 1.5x1.5x2 mm3 nominal spatial resolution and 5.1-sec temporal resolution in capturing the kinetics of metabolite changes in rat hindlimb during a stimulation-recovery protocol.