The main challenge for cardiac MRI is motion. To account for respiratory motion, k-space-based self- navigation approaches have recently been introduced [1]. We took self-navigation to the next level by exploiting an image-based beat-to-beat respiratory motion correction algorithm [2] for coronary MRI. For each heartbeat, under-sampled sub-images (sub-sets of data used for final image reconstruction) are reconstructed and multi-dimensional respiratory motion parameters extracted. Non-linear reconstruction (related to Compressed Sensing) is proposed to generate the motion-corrected sub-images and the results are compared to those of a more conventional approach [3]. These studies include computer simulations and initial 3T human in vivo data.