Patient motion during MRI acquisition causes image artifacts that distort the scans. Retrospective motion correction (MoCo) is a post-processing method that reduces such artifacts. Deep learning models are a promising approach towards retrospective MoCo. However, these models require large datasets of motion-free and motion-corrupted scans, which are often unavailable. Current methods simulate motion artifacts using random transformations in the spatial or frequency domain of the scan. Such random transformations can result in artifacts that are not physically realistic. An alternative approach would use synthetic motion curves that mimic patient movement. This project evaluates two generative models, TimeVAE and Fourier Flow, trained on motion curves from the ABCD study. We assess their ability to generate realistic motion curves using statistical metrics and a novel benchmarking protocol (TSGBench). Both models capture key distribution characteristics but struggle with abrupt motions such as coughing. Nonetheless, the results are promising for future research.