Is Synthetic Data all We Need? Benchmarking the Robustness of Models Trained with Synthetic Images