Storage I/O traces serve as a critical foundation for system performance evaluation, storage optimization, resource management, and diagnostic troubleshooting. However, due to proprietary constraints or privacy concerns, obtaining large-scale datasets from real-world production environments remains challenging. To mitigate this, existing solutions typically rely on small-scale real samples to generate high-fidelity synthetic traces using statistical or machine learning techniques. Yet, most existing methods adopt a single, unified model to generate all fields, overlooking their distinct distributional patterns. To address this limitation, we propose a differentiated generation method, DiffGen, which employs a classifier to categorize fields and then synthesize each category using tailored models. Experiments on real-world traces demonstrate its superior accuracy and efficiency.