<p>The practical deployment of human motion prediction models faces a fundamental challenge: when a model trained on data from a single subject is required to predict motions of unseen subjects with potentially different height, body shape, and movement habits, its performance exhibits significant instability and may even completely fail on certain individuals. This performance volatility and worst-case collapse under the zero-shot cross-subject setting severely hinder the reliable application of the technology. To address this challenge, this study presents an integrated framework for out-of-distribution stability, introducing three novel metrics—inter-subject error discrepancy, training-subject generalization stability, and worst-subject performance ratio—that collectively reveal instability risks overlooked by traditional geometric errors. We further construct a Multi-granularity Adaptive Fusion Gated Network (MG-AFGNet), grounded in the core principles of distribution alignment and structured dependency modeling. The network aligns inter-subject distributions via input normalization and output inverse projection modules, while explicitly modeling structured dependencies across joints and time through a dynamic adaptive fusion module. Experiments on three benchmark datasets demonstrate that our approach achieves a decisive advantage in cross-subject stability metrics over multiple baselines, with significant improvements in training-subject generalization stability. Crucially, comparisons under traditional protocols with seven mainstream methods confirm that this stability gain does not come at the cost of prediction accuracy, challenging the conventional trade-off assumption. This work promotes a paradigm shift in human motion prediction from pursuing single-subject accuracy toward ensuring cross-subject generalization stability, providing a solid theoretical and experimental benchmark for future research.</p>