<p>In dynamic scene modeling, a central challenge is to simultaneously achieve high reconstruction quality and capture rich semantic detail. High fidelity is necessary for visual realism, and semantic understanding is key for intelligent interaction and editing. However, existing methods often struggle to deliver both concurrently, particularly when dealing with complex object interactions and fast motions, which can degrade rendering quality and semantic clarity. To overcome this, we introduce label-guided 4D Gaussian splatting, a novel method that fuses a spatiotemporal Gaussian representation with semantics-guided motion modeling. We conducted qualitative and quantitative experiments on real-world datasets to validate our approach, demonstrating that our method excels at capturing scene details, achieving superior reconstruction quality. By incorporating category labels, our approach also supports scene editing, making it applicable for various downstream tasks.</p>