Previous work has not shown if learned components for lensless image reconstruction can generalize to physical modifications. In this work, we train a reconstruction approach on an open-sourced, multi-mask dataset, and demonstrate improved generalizability.