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

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.

GDAGS introduces a gradient-direction-aware adaptive density control framework with two key components:

Gradient Coherence Ratio (GCR): A normalized metric that quantifies directional consistency of subgradients for each Gaussian.
Nonlinear Dynamic Weighting: A mechanism that uses GCR to independently weight Gaussians, promoting effective densification while suppressing redundancy.

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.

We evaluate GDAGS on three real-world datasets spanning 13 scenes. Our method demonstrates superior visual quality while having a compact spatial representation that requires only 50% Gaussians and memory consumption compared to Pixel-GS.

Methods	Mip-NeRF360				Tanks&Temples				Deep Blending
Methods	SSIM↑	PSNR↑	LPIPS↓	Mem↓	SSIM↑	PSNR↑	LPIPS↓	Mem↓	SSIM↑	PSNR↑	LPIPS↓	Mem↓
Plenoxels	0.626	23.08	0.463	2.1GB	0.719	21.08	0.379	2.3GB	0.795	23.06	0.510	2.7GB
INGP	0.671	25.30	0.371	13MB	0.723	21.72	0.330	13MB	0.797	26.42	0.423	13MB
Mip-NeRF360	0.792	27.69	0.237	85MB	0.759	22.22	0.257	85MB	0.901	29.40	0.425	85MB
3DGS	0.815	27.21	0.214	734MB	0.841	23.14	0.183	411MB	0.903	29.41	0.743	676MB
Pixel-GS	0.832	27.72	0.178	1.2GB	0.853	23.74	0.150	1.05GB	0.896	28.91	0.248	1.1GB
AbsGS-0004	0.820	27.49	0.191	728MB	0.853	23.73	0.162	304MB	0.902	29.67	0.256	444MB
GDAGS (Ours)	0.839	28.02	0.145	515MB	0.854	23.79	0.165	226MB	0.905	29.70	0.235	388MB

Quantitative comparison on three datasets. SSIM↑ and PSNR↑ are higher-the-better; LPIPS↓ is lower-the-better.

GDAGS achieves high-quality visual fidelity across all scenes while effectively mitigating localized blur and detail loss, such as vegetation under benches, uneven concrete pavements, and textural details on train surfaces.

Qualitative comparisons of different methods on scenes from Mip-NeRF360, Tanks&Temples and Deep Blending datasets.

Comparison of over-densification phenomena across different methods on scenes from the Mip-NeRF360, Tanks & Temples, and Deep Blending datasets.

We conduct ablation experiments on the proposed weighting method. Split gradient weighting effectively optimizes SSIM and LPIPS indicators and reduces memory costs. Clone gradient weighting improves PSNR indicators but increases memory costs. Integrating both achieves a balance between performance and efficiency.

Methods	Mip-NeRF360				Tanks&Temples				Deep Blending
Methods	SSIM↑	PSNR↑	LPIPS↓	Mem↓	SSIM↑	PSNR↑	LPIPS↓	Mem↓	SSIM↑	PSNR↑	LPIPS↓	Mem↓
3DGS	0.815	27.21	0.214	744MB	0.841	23.14	0.183	411MB	0.904	29.41	0.243	676MB
GDAGS-L	0.814	27.55	0.248	713MB	0.849	23.74	0.179	321MB	0.893	29.62	0.241	397MB
GDAGS-S	0.819	27.52	0.240	441MB	0.846	23.54	0.178	195MB	0.905	29.67	0.238	328MB
GDAGS-C	0.812	27.46	0.217	615MB	0.847	23.71	0.180	305MB	0.904	29.70	0.240	452MB
GDAGS (Ours)	0.839	28.02	0.145	515MB	0.847	23.55	0.176	226MB	0.905	29.70	0.235	388MB

Ablation experiment on three datasets.

Our work builds upon recent advances in 3D Gaussian Splatting and novel view synthesis:

3D Gaussian Splatting introduces explicit Gaussian primitives for real-time radiance field rendering.

AbsGS addresses over-reconstruction by enforcing gradient direction uniformity.

Pixel-GS trades memory efficiency for geometric precision by weighting gradients via pixel coverage.

BibTeX

@article{zhou2025gdags,
  title={Gradient-Direction-Aware Density Control for 3D Gaussian Splatting},
  author={Zhou, Zheng and Xiong, Yu-Jie and Zhang, Jia-Chen and Xia, Chun-Ming and Qiu, Xihe and Zhan, Hongjian},
  journal={arXiv preprint arXiv:2508.09239},
  year={2025}
}

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

Abstract

Method

Results

Quantitative Evaluation

Qualitative Evaluation

Ablation Study

Related Work

BibTeX