A modified reconstruction is proposed for highly accelerated dMRI. The method employs machine learning in a model-based setting. The current work improves the generalizability of the deep-learned plug-and-play prior employed in the reconstruction, for enabling utilization of the reconstruction from a wide range of acquisition settings. This is achieved by learning a compact q-space representation in a rotation invariant space. The method is tested for the reconstruction of combined multi-band and in-plane accelerated data from both single-shell and multi-shell experiments. The reconstruction error is shown to be less than 3% for net acceleration of R=12 for single- and multi-shell cases.