TNT-GS: Truncated and Tailored Gaussian Splatting

Gaussian Splatting (GS) is widely utilized for efficiently representing and rendering complex 3D scenes by modeling them as continuous distributions of Gaussians, enabling high-quality and flexible scene reconstruction. However, GS is limited in its ability to represent high-frequency details and sharp transitions due to its inherent low-pass filtering effect. As a result, GS often requires stacking multiple Gaussians to better approximate fine structures, which significantly increases both computational and memory overhead. To address this fundamental limitation in representation, we propose Truncated and Tailored Gaussian Splatting (TNT-GS), a novel approach designed to enhance shape complexity and preserve sharp boundaries. Our method truncates Gaussians to generate sharp edges without the need for excessive Gaussian stacking, thereby improving efficiency. Additionally, we introduce learnable parameters to dynamically tailor the receptive field of the primitives, providing fine-grained control over the balance between high-frequency details and smooth low-frequency regions, achieving an optimal trade-off between sharpness and smoothness. Furthermore, we employ specialized densification strategies to enhance model efficiency while computing the affected tiles for each primitive, ensuring computational efficiency. Experimental results demonstrate that our method surpasses state-of-the-art approaches in both storage efficiency and rendering speed, making it a highly effective solution for real-time rendering.