In medical imaging, balancing scan duration, image quality and patient comfort is crucial. Magnetic resonance (MR) imaging provides high soft-tissue contrast, but scan time increases with higher spatial resolution. Acquiring multiple low-resolution (LR) scans in different orientations can mitigate this. To enable easier downstream analysis in a unified representation, fusing these images is valuable, yet combining them remains challenging. Because high-resolution (HR) ground-truth data are largely unavailable in clinical practice, we introduce a novel fully selfsupervised super-resolution (SR) approach to combine two anisotropic LR scans into one high-resolution image [1]. To train our multi-view neural network, we propose a sparse coordinate-based loss that allows the integration of arbitrarily scaled LR images. We evaluated our method on two publicly available brain MRI datasets. We simulated anisotropic LR coronal and axial scans with a slice thickness of 2, and 4mm for the LR images. HR data was just used for evaluation purposes only and was not available during training. Our SR pipeline consists of two phases: a patient-agnostic offline phase, where the model learns cross-patient similarities, and a patient-specific online phase, where additional subject-specific information can be learned. Our results show that our approach produces SR images of equal or higher quality compared to current reference approaches. Additionally, we demonstrate that the proposed method generalizes well across different datasets. A key advantage lies in the substantial speed-up achieved in the patient-specific reconstruction, up to tenfold, while maintaining or improving image quality, making the method more applicable in clinical scenarios.