Gradient-Direction-Aware Density Control for 3D Gaussian Splatting

TL;DR

GDAGS tackles two key weaknesses of 3D Gaussian Splatting by introducing Gradient Coherence Ratio (GCR) to measure directional gradient consistency and a nonlinear dynamic weight to per-Gaussian gradient contributions. This direction-aware density control prioritizes splitting of Gaussians with conflicting gradients and promotes coherent densification for cloning, yielding sharper geometry with fewer Gaussians and lower memory. Across thirteen real-world scenes, GDAGS achieves competitive rendering quality while reducing memory compared to state-of-the-art baselines, and ablations confirm the superiority of the nonlinear weighting over linear alternatives. The approach generalizes to other 3DGS frameworks, improving SSIM/LPIPS in integrated variants and offering a practical path toward efficient, high-fidelity novel view synthesis.

Abstract

The emergence of 3D Gaussian Splatting (3DGS) has significantly advanced Novel View Synthesis (NVS) through explicit scene representation, enabling real-time photorealistic rendering. However, existing approaches manifest two critical limitations in complex scenarios: (1) Over-reconstruction occurs when persistent large Gaussians cannot meet adaptive splitting thresholds during density control. This is exacerbated by conflicting gradient directions that prevent effective splitting of these Gaussians; (2) Over-densification of Gaussians occurs in regions with aligned gradient aggregation, leading to redundant component proliferation. This redundancy significantly increases memory overhead due to unnecessary data retention. We present Gradient-Direction-Aware Gaussian Splatting (GDAGS) to address these challenges. Our key innovations: the Gradient Coherence Ratio (GCR), computed through normalized gradient vector norms, which explicitly discriminates Gaussians with concordant versus conflicting gradient directions; and a nonlinear dynamic weighting mechanism leverages the GCR to enable gradient-direction-aware density control. Specifically, GDAGS prioritizes conflicting-gradient Gaussians during splitting operations to enhance geometric details while suppressing redundant concordant-direction Gaussians. Conversely, in cloning processes, GDAGS promotes concordant-direction Gaussian densification for structural completion while preventing conflicting-direction Gaussian overpopulation. Comprehensive evaluations across diverse real-world benchmarks demonstrate that GDAGS achieves superior rendering quality while effectively mitigating over-reconstruction, suppressing over-densification, and constructing compact scene representations.

Figures (14)

• Figure 1: (a) illustrates the Gaussian ellipsoid splatting process, where arrows of different colors represent the gradient direction and magnitude of different Gaussians on the pixels. (b) shows the densification process of different methods. In 3DGS, a large Gaussian covering many pixels may fail to split because the combined gradient magnitude from different pixels falls below the threshold, leading to over-reconstruction as shown in the Rendered part of (c), which manifests as blurry areas. AbsGS forces all Gaussian gradients to be positive, causing the combined gradient magnitude from different pixels to increase significantly. This results in a substantial rise in the number of splitting Gaussians, leading to over-densification as shown in the Top view of the point cloud part of (c), which appears as a large number of Gaussian point clouds outside the scene. Our proposed method utilizes gradient direction information to effectively control the splitting of large Gaussians and the growth in the number of small and medium Gaussians. (c) also presents the performance (performance metrics are labeled in the top-left corner) and efficiency (FPS and total number of Gaussians in the scene, #G) of different methods.
• Figure 2: The pipeline of GDAGS. First, for each Gaussian, GDAGS computes the GCR to quantify the directional coherence of its subgradients. Subsequently, this GCR metric is mapped through a nonlinear dynamic weighting function to generate per-Gaussian gradient weights, which modulate the view-space positional gradient magnitudes and produce a refined decision metric. Finally, this decision metric is compared against a predefined threshold to dynamically regulate densification.
• Figure 3: Qualitative comparisons of different methods on scenes from Mip-NeRF360, Tanks&Temples and Deep Blending datasets. Enlarged images are displayed in the bottom right corner.
• Figure 4: Performance of different hyperparameters $p$ in multiple datasets.
• Figure 5: Visualization of different densification methods during the training process in bicycle sense.