PurposeThe goal of this work is to increase the spatial resolution of training data, used by reconstruction methods such as k–t SENSE in order to calculate the missing data in a series of dynamic images, without compromising their temporal resolution or acquisition time.TheoryThe k‐t SENSE method allows dynamic imaging at high acceleration factors with high reconstruction quality. However, the low resolution training data required by k‐t SENSE may cause undesired temporal filtering effects in the reconstructed images.MethodsIn this work, a feedback regularization approach is applied to realize auto‐calibration of the k‐t SENSE algorithm. To that end, a full resolution training data set is calculated from the accelerated data itself using a TSENSE reconstruction. The reconstructed training data are then fed back for the actual k‐t SENSE reconstruction. For evaluation of our approach, temporal filtering effects are quantified by calculating the modulation transfer function and noise measurements are done by Monte‐Carlo simulations.ResultsComputer simulations and cardiac imaging experiments demonstrate an improved temporal fidelity of auto‐calibrated k‐t SENSE compared to standard k‐t SENSE.ConclusionAuto‐calibrated k‐t SENSE provides high quality reconstructions for dynamic imaging applications. Magn Reson Med 71:1123–1129, 2014. © 2013 Wiley Periodicals, Inc.