In this talk, we introduce a fresh perspective on the domain gap between synthetic and real-world data. Challenging the traditional dichotomy, we argue that both are merely models of reality and should be treated as such. Using human pose estimation in driver monitoring as a case study, the talk presents a method to quantify the domain gap by comparing classifier performance on perceptually identical real and synthetic images. The approach leverages high-fidelity digital twins and controlled data capture to isolate imperfections in both domains. Rather than choosing between synthetic or real data, we propose a hybrid strategy: record what’s hard to model (e.g., human behavior) and simulate what’s governed by known laws (e.g., light transport). We are closing the talk with recent insights on results and actionability of the proposed approach.