Research map / 3D vision / Gaussian splatting

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

NeurIPS 2025 · 3D vision › Gaussian splatting · approach: Compact and efficient Gaussian splatting · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DC4GS, which adds a circular-mean directional consistency of positional gradients to 3DGS adaptive density control, using it both to decide splits (DCC) and to place sub-primitives (DCS), reducing primitive counts while improving fidelity.

Problem

Conventional adaptive density control in 3D Gaussian splatting splits primitives based only on positional-gradient magnitude and places sub-primitives randomly, causing redundant splits and misaligned sub-primitives.

Approach

Introduce directional consistency (circular mean norm of unit positional gradients across pixels within a Gaussian) to weight the split criterion (DCC) and to choose split locations along the principal axis by minimizing a DC-based cost (DCS), with polynomial regression over 5 candidates.

Key results

• Integrating DC4GS with AbsGS reduces primitives by about 20% on average (up to 30% in the Room scene) while improving quality, e.g., Mip-NeRF360 LPIPS 0.191 to 0.188 and Deep Blending SSIM 0.900 to 0.905.
• On Tanks and Temples Train, AbsGS+DC4GS raises PSNR from 21.58 to 22.38 with primitives reduced from 1008K to 787K.
• Ablation on AbsGS: DCC alone cuts primitives (3149K to 2861K on Mip-NeRF360) and DCS alone improves quality; combining them gives 2615K primitives and 27.625 PSNR on Mip-NeRF360.

Measured effects reported in the paper

Surprising findings

• Simply extending AbsGS training to 60K iterations yields only marginal or negative gains, suggesting the gains are not from longer training.
• Training overhead is substantial: DC4GS raises 3DGS training time on Mip-NeRF360 from 34m to 49m.

Limitations

• Per-iteration DC-based split cost computation adds notable training overhead (about 48 ms per iteration on bicycle scene).
• Gains on some scenes are small or negative for some metrics (e.g., 3DGS on bonsai/counter PSNR slightly lower).

Experimental setup

Datasets: Mip-NeRF 360 Tanks and Temples Deep Blending D-NeRF DyNeRF

Benchmarks: Mip-NeRF 360 Tanks and Temples Deep Blending

Metrics: PSNR SSIM LPIPS number of primitives memory training time rendering time

Baselines: 3DGS AbsGS Pixel-GS Scaffold-GS GES LPM Mip-NeRF360 iNGP-big Plenoxels

Builds on: 3D Gaussian Splatting AbsGS Pixel-GS Scaffold-GS

Closest papers by meaning

• AdpSplit: Error-Driven Adaptive Splitting for Faster Geometry Discovery in 3D Gaussian Splatting (NeurIPS 2026 · similarity 0.90)
• Frequency-Aware Density Control via Reparameterization for High-Quality Rendering of 3D Gaussian Splatting (AAAI 2025 · similarity 0.88)
• ReorgGS: Equivalent Distribution Reorganization for 3D Gaussian Splatting (NeurIPS 2026 · similarity 0.88)
• Metropolis-Hastings Sampling for 3D Gaussian Reconstruction (NeurIPS 2025 · similarity 0.87)
• Steepest Descent Density Control for Compact 3D Gaussian Splatting (CVPR 2025 · similarity 0.87)
• Improving Gaussian Splatting with Localized Points Management (CVPR 2025 · similarity 0.86)