Statistical Framework to Evaluate Synthetic Data Utility with Minimal Real Samples

This paper introduces two methods, including a symmetry-based paired loss difference test, to determine if training on synthetic datasets improves real-world performance. The goal is to minimize the number of real-world data points required to validate the sim-to-real transition for engineering AI models.