SharpTimeGS: Sharp and Stable Dynamic Gaussian Splatting via Lifespan Modulation

TL;DR

SharpTimeGS introduces a lifespan-aware 4D Gaussian framework that decouples static and dynamic content through a learnable lifespan per primitive. By replacing the traditional Gaussian temporal decay with a flat-top visibility and modulating motion via a lifespan-aware factor $f(\sigma_t, r)$, the method stabilizes long-lived structures while preserving fast dynamics. A velocity–lifespan–aware densification and a velocity-guided initialization further allocate capacity to dynamic regions and stabilize optimization, enabling real-time 4K rendering at 100 FPS on a single RTX 4090. Extensive experiments across Neural3DV, ENeRF-Outdoor, and SelfCap show state-of-the-art quality and robust handling of complex motions, with ablations confirming the contributions of each component. The work offers a practical, scalable approach for high-fidelity dynamic scene reconstruction and rendering, advancing real-time 4D Gaussian representations.

Abstract

Novel view synthesis of dynamic scenes is fundamental to achieving photorealistic 4D reconstruction and immersive visual experiences. Recent progress in Gaussian-based representations has significantly improved real-time rendering quality, yet existing methods still struggle to maintain a balance between long-term static and short-term dynamic regions in both representation and optimization. To address this, we present SharpTimeGS, a lifespan-aware 4D Gaussian framework that achieves temporally adaptive modeling of both static and dynamic regions under a unified representation. Specifically, we introduce a learnable lifespan parameter that reformulates temporal visibility from a Gaussian-shaped decay into a flat-top profile, allowing primitives to remain consistently active over their intended duration and avoiding redundant densification. In addition, the learned lifespan modulates each primitives' motion, reducing drift in long-lived static points while retaining unrestricted motion for short-lived dynamic ones. This effectively decouples motion magnitude from temporal duration, improving long-term stability without compromising dynamic fidelity. Moreover, we design a lifespan-velocity-aware densification strategy that mitigates optimization imbalance between static and dynamic regions by allocating more capacity to regions with pronounced motion while keeping static areas compact and stable. Extensive experiments on multiple benchmarks demonstrate that our method achieves state-of-the-art performance while supporting real-time rendering up to 4K resolution at 100 FPS on one RTX 4090.

Figures (5)

• Figure 1: (a) Temporal visibility in existing motion-based methods. A step-like lifespan (blue line) requires multiple Gaussian primitives for approximation. (b) With a learnable radius $r$, our visibility function allows a single Gaussian primitive to represent a step-like lifespan (blue line). (c) In existing motion-based methods (e.g., FreeTimeGS wang2025freetimegs), residual velocities accumulate over time, causing drift in static regions. (d) With our lifepan modulation term $f(\sigma_t, r)$, where $\sigma_t$ is the lifespan variance and $r$ is the lifespan radius, static primitive remain static without drift.
• Figure 2: The pipeline of our method. We represent a dynamic scene using Gaussian primitives whose temporal visibility adapts to the actual lifespan of each point. To achieve this, we introduce a lifespan-dependent parameter $r$ that modulates the temporal Gaussian, allowing a single primitive to accurately model its full lifespan. Moreover, through the modulation terms $f(\sigma_t, r)$ related to $\sigma_t$ and $r$, the static part can be completely static and still able to express dynamic parts (the static and fast dynamic regions will be transformed into equations of motion in red and blue boxes, respectively). Note that the formulas in the boxes are only approximations. During optimization, all Gaussian representations remain identical.
• Figure 3: Qualitative comparison on the SelfCap Dataset xu2024representing. Our method achieves the rendering quality compared with baseline methods, especially for distant static regions (e.g., books and wall) and fast-moving dynamic regions (e.g., hairs and ball).
• Figure 4: Qualitative comparison on the ENeRF-Outdoor Dataset lin2022efficient. Our method achieves the best rendering quality compared with baseline methods, especially for distant static regions and fast-moving dynamic regions.
• Figure 5: Ablation study on the SelfCap Dataset xu2024representing. Our full model achieves the best rendering quality, especially for distant static regions and fast-moving dynamic regions.