Motion artifacts are a longstanding obstacle in MRI, degrading image quality and leading to diagnostic uncertainty or costly re-scans. Despite the promise of deep learning solutions, most current approaches rely on supervised train- ing, which demands paired motion-free and motion-corrupted images, a type of data that is extremely rare in clinical practice. This scarcity presents a major roadblock for applying these methods broadly, particularly in routine hospital workflows where patient motion is unpredictable and acquisition pa- rameters vary. In this talk, I’ll introduce a framework designed specifically to overcome this data bottleneck. Instead of requiring matched image pairs or k-space access, it uses diffusion models to simulate realistic motion arti- facts on clean images, producing synthetic pairs that can then be used to train correction models. This unsupervised pipeline sidesteps the need for controlled acquisition experiments and is compatible with a wide range of clinical scans and hardware setups. I’ll present key findings from our eval- uation across datasets and acquisition planes, and discuss why solving the data availability challenge is crucial for making AI-based motion correction a practical reality in MRI.