Motion estimation is a core task in computer vision necessary for understanding and interacting with dynamic environments. Currently, the best-performing motion estimation methods are often supervised neural networks that require a great amount of annotated data. As annotating real-world data with motion information is difficult and error-prone, several dataset generators were created to produce synthetic motion estimation data with rich ground-truth annotations. However, these generators usually focus on specific dataset aspects such as photo-realism and data diversity paying no attention to aspects essential for researching the influence of training data on network performance. In this work, we present our Dataset Generation Framework with emphasis on data reproducibility, high controllability of data parameters and extendability to new tasks and requirements. We show that our framework is able to produce training data that is more efficient than the established training datasets, create datasets with specific features for testing purposes and generate datasets that are identical except for a single parameter change useful for isolating and examining the influence of the parameter.