AD-GS: Alternating Densification for Sparse-Input 3D Gaussian Splatting

Gurutva Patle1, Nilay Girgaonkar1, Nagabhushan Somraj1, Rajiv Soundararajan1,
1Indian Institute of Science (IISc), Bengaluru
Paper arXiv Video Code Data

AD-GS renders the artifact free novel views from sparse input views.

Abstract

3D Gaussian Splatting (3DGS) has shown impressive results in real-time novel view synthesis. However, it often struggles under sparse-view settings, producing undesirable artifacts such as floaters, inaccurate geometry, and overfitting due to limited observations.

We propose AD-GS, a novel alternating densification framework that interleaves high and low densification phases. During high densification, the model densifies aggressively, followed by photometric loss based training to capture fine-grained scene details. Low densification then primarily involves aggressive opacity pruning of Gaussians followed by regularizing their geometry through pseudo-view consistency and edge-aware depth smoothness.

This alternating approach helps reduce overfitting by carefully controlling model capacity growth while progressively refining the scene representation. Extensive experiments on challenging datasets demonstrate that AD-GS significantly improves rendering quality and geometric consistency compared to existing methods.

Method

AD-GS Method: The training begins with a Warm-Up Phase of NW iterations, where two 3DGS models are trained independently using photometric loss. This is followed by the Alternating Densification Phase, which alternates between low and high densification steps every NL and NH iterations, respectively. Low Densification applies strict gradient and opacity thresholds and includes additional supervision via pseudo-view consistency and edge-aware depth smoothness losses. The total loss used during this phase is defined as the "Combined Loss". High Densification enables aggressive Gaussian growth and uses only photometric loss to recover high-frequency details. Over iterations, this alternation progressively refines scene geometry while avoiding overfitting and floaters. The figure illustrates how Gaussian counts increase during high densification and decrease due to pruning during low densification.

3DGS vs AD-GS teaser image

Visuals

Drag and Check Novel frames

Below we can check the novel view renders. Use the slider here.

Loading...

Observations

Challenges of Densification in Sparse View Training

As training progresses in 3DGS, densification under sparse views leads to floating artifacts and inconsistent geometry.

3DGS vs AD-GS teaser image

Floaters Suppression with AD-GS

3DGS develops floaters (as seen in the red boxes) as we train for more iterations due to its uncontrolled densification. However, our AD-GS model is able to suppress floaters.

3DGS vs AD-GS teaser image

Detail Preservation with AD-GS

CoR-GS tends to employ aggressive smoothing to resolve floaters at the cost of loss of details, whereas AD-GS is able to recon- struct textures better.

3DGS vs AD-GS teaser image

Comaprision with other methods

3DGS vs AD-GS

Using AD-GS you can produces much cleaner and more stable reconstructions whereas other methods exhibits significant artifacts under sparse-view setting.

FSGS vs AD-GS

Using AD-GS you can produces much cleaner and more stable reconstructions whereas other methods exhibits significant artifacts under sparse-view setting.

CoR-GS vs AD-GS

Using AD-GS you can produces much cleaner and more stable reconstructions whereas other methods exhibits significant artifacts under sparse-view setting.

DropGaussian vs AD-GS

Using AD-GS you can produces much cleaner and more stable reconstructions whereas other methods exhibits significant artifacts under sparse-view setting.

More Visual Results

For more visual results, please check here.

BibTeX

@misc{patle2025adgs,
  author       = {Gurutva Patle and Nilay Girgaonkar and Nagabhushan Somraj and Rajiv Soundararajan},
  title        = {AD‑GS: Alternating Densification for Sparse‑Input 3D Gaussian Splatting},
  journal      = {arXiv preprint arXiv:2509.11003},
  year         = {2025},
  archivePrefix= {arXiv},
  primaryClass = {cs.GR}
}