VinUniversity Releases Paper and Code on Deciding 3DGS Gaussian Birth and Death via the First Variation of the Loss

What happened

Quan Tran Hong, Long Nguyen-Chi and Binh T. Nguyen (VinUniversity Center for AI Research) have posted a paper on arXiv, "Birth-Death Structural Learning for 3D Gaussian Splatting," that revisits density control in 3D Gaussian Splatting (3DGS). The paper is dated October 25, 2026. The implementation code has also been released on GitHub.

3DGS is a technique that can render 3D representations built from real-world captures in real time. The balance between quality and efficiency depends on adaptive density control (ADC), which clones, splits and removes Gaussians. Conventional rules use proxy metrics such as positional gradients, opacity and image-space error.

The authors point out that these metrics cannot directly answer questions such as "where would placing a new Gaussian reduce the loss?" and "which Gaussians can be removed with little impact?" They therefore propose "first-variation birth-death control," which decides additions and removals using a variational score derived from the reconstruction loss.

According to the paper, incorporating this score into 3DGS-MCMC on standard benchmarks yields state-of-the-art reconstruction quality, with larger improvements under tighter Gaussian-count budgets. Specific numbers are not given in the paper's abstract, within the scope of the available material.

How it works

At each structure-update step, the current set of Gaussians is held fixed. An auxiliary "mass" coordinate is attached to each Gaussian, and the multi-view loss is differentiated with respect to that coordinate. This is the score used for birth and death decisions.

The paper states that it shows this score to be the coordinate of the first variation of the Splat Regression Model. Gaussians with high scores are chosen as locations for new Gaussians (birth), while those with low scores and low opacity become targets for removal (death).

According to the GitHub README, the released code attaches the mass variable to each Gaussian's opacity. The score is q_i = -α_i ∂F/∂α_i, and a positive value means that increasing the contribution locally lowers the loss. Gaussians whose q_i exceeds a threshold τ are selected as parents for the standard version's cloning and splitting.

Removal based on opacity and size is retained. τ is a tunable hyperparameter, set to 1e-7 in the reported experiments. The README explains that the paper centers on the MCMC version, while the released code is an adaptation to standard 3DGS.

By the numbers

Why it matters

If quality metrics can be raised while reducing the number of Gaussians, this could help cut file size and rendering load. For practitioners working on web and mobile delivery and viewers, it could be one way to limit counts on the training side.

However, this is currently a research-stage report. For commercial use, the license of upstream 3DGS is a barrier. Using it in production or inspection settings would require confirming reproducibility on one's own data.

How to try it

1. Following the README, clone the repository with --recurse-submodules and set up the environment with scripts/setup.sh.
2. Prepare a scene in COLMAP format (images/ and sparse/0/). Mip-NeRF360 scenes can be used.
3. Train and evaluate the proposed method with bash scripts/train_eval.sh <scene> <output dir>. Specifying vanilla as the third argument runs the baseline for comparison.
4. Compare the metrics and Gaussian counts using results.json and per_view.json in the output directory.

Caveats and open questions

• The numbers are the authors' reports listed in the README; third-party reproduction cannot be confirmed from the available material.
• The README results are a 7-scene average at 7k iterations. Results at the standard 30k iterations and specific values for the MCMC version are not confirmed in the material.
• It is unconfirmed whether the paper has been peer reviewed. Treat it as a preprint on arXiv.
• According to the README, Gaussian counts differ by scene. Per-scene results are said to be in docs/results.md, but we have not checked them.
• The effects on training time and memory are not confirmed in the material.
• Whether commercial use is allowed must follow the upstream license (limited to research and non-commercial evaluation). The repository's own license is listed as NOASSERTION.

Glossary

Adaptive density control (ADC)

A mechanism that adjusts the number and placement of Gaussians during 3DGS training by cloning, splitting and removing them.

First variation

A quantity describing how a function's value changes when its input is slightly altered. Here it indicates the direction of change in the loss.

3DGS-MCMC

A 3DGS variant that treats the addition and removal of Gaussians through probabilistic sampling.

PSNR / SSIM / LPIPS

Image reproduction quality metrics. Higher is better for PSNR and SSIM; lower is better for LPIPS.

Sources

Primary sources (publisher, paper, repository)

• Birth-Death Structural Learning for 3D Gaussian SplattingarXiv（arxiv.org）・en
• quanth238/birth-death-3dgsGitHub（github.com）・en

The X post that surfaced this story

• View @CSVisionPapers's post on Xx.com

This article was generated in Japanese by AI from the sources listed (primarily primary sources) and published automatically, without prior human review, after automated checks such as source matching (how we use AI). Figures and specifications are as stated by the publishers and have not been independently verified by us. Please consult the original sources. To report an error, use our contact form; we will review it promptly and correct it. This page is an AI translation of the Japanese original. If anything differs, the Japanese version prevails. Japanese original