DC4GS: Directional Consistency-Driven Adaptive Density Control for 3D Gaussian Splatting

TL;DR

DC4GS addresses inefficiencies in gradient-based adaptive density control for 3D Gaussian Splatting by introducing directional consistency (DC) of positional gradients. It defines DC via the circular mean of gradient directions within each Gaussian and uses a DC-weighted split criterion (DCC) plus DC-guided split (DCS) to decide when to split and where to place sub-primitives along the Gaussian's principal axis. Empirically, DC4GS yields up to a $\sim30\%$ reduction in primitive counts while maintaining or improving reconstruction fidelity across standard datasets, with modest training overhead but notable rendering speedups due to fewer primitives. The method is readily integrable with existing 3DGS pipelines and enhances high-frequency detail and structural boundary preservation, enabling more efficient real-time or memory-constrained rendering.

Abstract

We present a Directional Consistency (DC)-driven Adaptive Density Control (ADC) for 3D Gaussian Splatting (DC4GS). Whereas the conventional ADC bases its primitive splitting on the magnitudes of positional gradients, we further incorporate the DC of the gradients into ADC, and realize it through the angular coherence of the gradients. Our DC better captures local structural complexities in ADC, avoiding redundant splitting. When splitting is required, we again utilize the DC to define optimal split positions so that sub-primitives best align with the local structures than the conventional random placement. As a consequence, our DC4GS greatly reduces the number of primitives (up to 30% in our experiments) than the existing ADC, and also enhances reconstruction fidelity greatly.

Figures (14)

• Figure 1: Comparison of AbsGS ye2024absgs and our Directional Consistency-driven density control (DC4GS) in terms of (a) primitive counts during training and (b) reconstruction quality. DC4GS saturates much earlier (here, 6k iterations) than the AbsGS does. Upon convergence, it achieves a significant reduction in primitive counts (here, 30%) and high-quality splits as well, resulting in much higher reconstruction quality (b) for high-frequency details (here, see the red inset).
• Figure 2: Visualization of how DC reflects structural complexities. In (a), a Gaussian can be aligned with a single-peak ground truth (GT) using a simple shift, which is indicated by a low angular variance (i.e., high DC). In contrast, (b) requires splitting due to the mismatch from the two-peak GT. This can be reflected in its divergent gradients (i.e., low DC), but not in the gradient magnitudes. (c) visualizes the per-pixel DC values in a real-world image, distinguishing directionally coherent (e.g., the sky) and incoherent (the train) regions well.
• Figure 3: Comparison of the conventional ADC scheme (e.g., 3DGS kerbl23, AbsGS ye2024absgs, and Pixel-GS zhang2024pixel) and our DC4GS. (a) Gaussian primitives are selected using the positional gradient magnitude-based criterion, and the new Gaussians are randomly spawned within the pre-split Gaussian. (b) Our DC-based split Criterion (DCC) further integrates the DC (the circular mean of the positional gradients) into the previous ADC criterion. Also, our DC-guided Split (DCS) better places the new sub-primitives so that their DCs are best distinguished.
• Figure 4: Comparison of split candidates selected by the criterion of AbsGS and DCC from the identical training states. After 14,900 training steps on the Stump scene using the 3DGS, Gaussians selected for splitting are visualized in yellow (AbsGS) and red (DCC). The difference (green) shows (potentially redundant) 20,252 Gaussians are not split by the DCC. Most of these differences are already of tiny sizes, suggesting limited structural gain from further splitting.
• Figure 5: Overview and examples of the DCS. (a) Candidate points are uniformly sampled along the principal axis and projected onto the image plane. The split cost $J(x)=J_l(x)+J_r(x)$ is computed using the directional consistency and gradient magnitudes, and the polynomial regression is used to find the optimal split position $x_{\text{opt}}$ from the limited set of discrete samples. (b) In a 2D experimental example, our DCS well aligns the split along the regions structural change occurs, resulting in faithful reconstruction. The plot demonstrates quality statistics in over 10,000 randomized samples.