Deep learning-based gaze estimation methods exhibit significant performance degradation in cross-domain scenarios. Unlike general cross-domain tasks in computer vision, the eye region contains the most critical information for gaze estimation, making general-purpose methods fail to perform effectively. Furthermore, existing domain generalization approaches for gaze estimation overlook this unique characteristic. While these approaches focus on generalization constraints, they ignore the performance disparities between different features (global vs. local) under cross-domain conditions. To resolve this limitation, we propose a Domain-Invariant Global-Local feature learning framework (DIGL), which enhances the robustness of fused features by fully leveraging the transferability of local information. Specifically, we design a top-down eye attention module that achieves deep integration of global and local information, employing local fine-grained details to restore eye features obscured in global representations. Moreover, to address the limitations of current contrastive learning sample selection strategies, we propose an efficient contrastive learning method tailored explicitly for regression problems. Experimental results show that the DIGL framework outperforms state-of-the-art methods on multiple cross-domain metrics, achieving improvements of up to 30% over baseline methods and up to 11% over existing SOTA approaches in cross-domain scenarios.