# EDGS: Eliminating Densification for Efficient Convergence of 3DGS

arXiv: https://arxiv.org/abs/2504.13204 · HTML: https://mlsanity.com/paper/2504.13204/

## Key facts

- Cited by 22 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 15 Apr 2025.

## Built on by (most cited first, 22 of 22)

- [RoMa v2: Harder Better Faster Denser Feature Matching](https://mlsanity.com/paper/2511.15706/) (2025)
- [LODGE: Level-of-Detail Large-Scale Gaussian Splatting with Efficient Rendering](https://mlsanity.com/paper/2505.23158/) (2025)
- [Faster-GS: Analyzing and Improving Gaussian Splatting Optimization](https://mlsanity.com/paper/2602.09999/) (2026)
- [World Reconstruction From Inconsistent Views](https://mlsanity.com/paper/2603.16736/) (2026)
- [ConeGS: Error-Guided Densification Using Pixel Cones for Improved Reconstruction With Fewer Primitives](https://mlsanity.com/paper/2511.06810/) (2025)
- [OmniRaindrop: A Comprehensive Benchmark for Dual-Focused Raindrop Removal under Day and Night Scenes](https://mlsanity.com/paper/2510.17719/) (2025)
- [Speeding Up the Learning of 3D Gaussians with Much Shorter Gaussian Lists](https://mlsanity.com/paper/2603.09277/) (2026)
- [MoRe: Monocular Geometry Refinement via Graph Optimization for Cross-View Consistency](https://mlsanity.com/paper/2510.07119/) (2025)
- [ProFuse: Efficient Cross-View Context Fusion for Open-Vocabulary 3D Gaussian Splatting](https://mlsanity.com/paper/2601.04754/) (2026)
- [Ego-1K -- A Large-Scale Multiview Video Dataset for Egocentric Vision](https://mlsanity.com/paper/2603.13741/) (2026)
- [SatSplat: Geometrically-Accurate Gaussian Splatting for Satellite Imagery](https://mlsanity.com/paper/2606.28581/) (2026)
- [Learn2Splat: Extending the Horizon of Learned 3DGS Optimization](https://mlsanity.com/paper/2605.15760/) (2026)
- [SwiftNDC: Fast Neural Depth Correction for High-Fidelity 3D Reconstruction](https://mlsanity.com/paper/2602.22565/) (2026)
- [The Role of Initialization in 3D Gaussian Splatting](https://mlsanity.com/paper/2603.20714/) (2026)
- [CAdam: Context-Adaptive Moment Estimation for 3D Gaussian Densification in Generative Distillation](https://mlsanity.com/paper/2605.20872/) (2026)
- [G2SR: Geometric Methods for Fast and Memory-Efficient Gaussian-based Surface Reconstruction](https://mlsanity.com/paper/2607.14470/) (2026)
- [Exploration Matters for Escaping the Blur Trap in 3D Gaussian Splatting](https://mlsanity.com/paper/2607.17965/) (2026)
- [Stipple: Real-Time Incremental Gaussian Splatting with Visual-Inertial Tracking](https://mlsanity.com/paper/2608.00931/) (2026)
- [StreamSplat: Streaming Feed-Forward 3D Gaussian Splatting](https://mlsanity.com/paper/2608.01659/) (2026)
- [SkyAnchor: Updating Metric-scale Aerial 3D Gaussian Scenes from Unposed Ground-View Sequences](https://mlsanity.com/paper/2609.13903/) (2026)
- [CoCam4D: Geometry-Aware Cooperative 4D Perception for Camera-Only Autonomous Driving](https://mlsanity.com/paper/2610.11577/) (2026)
- [Efficient Confidence-Aware Densification for 3DGS](https://mlsanity.com/paper/2610.13541/) (2026)

## Abstract (arXiv)

3D Gaussian Splatting reconstructs scenes by starting from a sparse Structure-from-Motion initialization and refining under-reconstructed regions. This process is slow, as it requires multiple densification steps where Gaussians are repeatedly split and adjusted, following a lengthy optimization path. Moreover, this incremental approach often yields suboptimal renderings in high-frequency regions. We propose a fundamentally different approach: eliminate densification with a one-step approximation of scene geometry using triangulated pixels from dense image correspondences. This dense initialization allows us to estimate the rough geometry of the scene while preserving rich details from input RGB images, providing each Gaussian with well-informed color, scale, and position. As a result, we dramatically shorten the optimization path and remove the need for densification. Unlike methods that rely on sparse keypoints, our dense initialization ensures uniform detail across the scene, even in high-frequency regions where other methods struggle. Moreover, since all splats are initialized in parallel at the start of optimization, we remove the need to wait for densification to adjust new Gaussians. EDGS reaches LPIPS and SSIM performance of standard 3DGS significantly faster than existing efficiency-focused approaches. When trained further, it exceeds the reconstruction quality of state-of-the-art models aimed at maximizing fidelity. Our method is fully compatible with other acceleration techniques, making it a versatile and efficient solution that can be integrated with existing approaches.