Geometry Re-Densification Module

• Geometry Re-Densification is a process that dynamically adjusts the density of geometric representations during optimization to improve reconstruction accuracy and eliminate redundant structures.
• It employs operations like Gaussian splitting, cloning, and pruning across diverse representations such as 3D Gaussian Splatting, LiDAR octrees, and template meshes.
• Techniques integrate gradient-based scoring, eigenentropy measures, and geometry-aware validation to balance photometric quality with geometric precision while managing computational load.

Geometry re-densification denotes, in the literature summarized here, a class of mechanisms that revise the spatial density of a geometric representation during optimization so that under-reconstructed regions receive additional support and redundant or implausible structure is removed. In 3D Gaussian Splatting (3DGS), these modules typically alter when Gaussians are split, cloned, pruned, or newly inserted; related formulations appear in transformer-based Gaussian generation, dynamic scene modeling, template-mesh deformation, LiDAR point-cloud compression, and mesh-density super-resolution. ResGS is a representative recent example: it introduces a residual split that adds a downscaled Gaussian as a residual, integrates a Gaussian image pyramid for progressive supervision, and implements a selection scheme that prioritizes the densification of coarse Gaussians over time (Lyu et al., 2024).

1. Scope and representative formulations

The term is not used for a single canonical algorithm. Instead, it refers to modules that change density in ways that are specific to the underlying representation. In static 3DGS, the representation is a set of anisotropic Gaussians with positions, covariances, opacities, and view-dependent colors; in template-based reconstruction it is a fixed-connectivity mesh; in LiDAR compression it is an octree; in SuperMeshingNet it is a low-density stress field that is mapped to a high-density stress field (Desiatov et al., 21 Mar 2026).

Work	Representation	Stated re-densification mechanism
ResGS (Lyu et al., 2024)	3DGS	residual split; Gaussian image pyramid; prioritizes densification of coarse Gaussians
EntON (Jäger et al., 6 Mar 2026)	3DGS	Eigenentropy-aware alternating densification and pruning
GeoTexDensifier (Jiang et al., 2024)	3DGS	geometry-aware splitting with Validation of Depth Ratio Change
ConeGS (Baranowski et al., 10 Nov 2025)	3DGS	error-guided Gaussian insertion along pixel cones
GRED (Yu et al., 28 Aug 2025)	LiDAR octree	re-densify at a denser scale, then re-sparsify for prediction
Mesh Density Adaptation (Jung et al., 2023)	template mesh	density adaptation energy without changing connectivity

For ResGS specifically, the supplied material reports the existence of residual split, the Gaussian image pyramid, and the Gaussian-selection scheme, but also states that the actual description or mathematical details of the Geometry Re-Densification module are not included in the excerpt; exact scaling factors, covariance adjustments, loss terms, and pseudocode are therefore unspecified in that material (Lyu et al., 2024).

2. Densification criteria and scoring functions

A central design choice is the criterion that decides which primitives should grow. In the conventional 3DGS formulation described by EntON, densification relies on the running average of the view-space position gradient magnitude and a threshold $\tau_{\mathrm{pos}}$; if the average exceeds the threshold, a Gaussian is marked for densification, with small-scale Gaussians cloned and large-scale Gaussians split (Jäger et al., 6 Mar 2026).

Several works replace that score with more explicit signals. "Revising Densification in Gaussian Splatting" defines a differentiable per-pixel error, for example

$$\mathcal E_\pi(u)=1-\mathrm{SSIM}\bigl(C_\pi(u),T_\pi(u)\bigr),$$

redistributes it to each primitive through its rendering weights,

$$E_k^\pi=\sum_{u\in\text{Pixels}} \mathcal E_\pi(u)\,\omega_k^\pi(u),$$

and then tracks

$E_k=\max_{\pi\in\Pi}E_k^\pi.$

Primitives with $E_k>\epsilon$ are marked for growth. The same work also adds budget-aware growth control and an opacity-bias correction for cloning, using

$\hat\alpha = 1-\sqrt{1-\alpha},$

so that cloning does not over-weight the duplicated primitive (Bulò et al., 2024).

ReAct-GS retains a gradient-based criterion but changes how gradients are aggregated across views. It argues that unweighted averaging causes gradient dilution when a primitive is important in only a few views. Its importance-aware score weights each view by the primitive’s $\alpha$-blending rendering weight:

$$g_i = \frac{\sum_{k=1}^M \omega_i^k \sqrt{\bigl(\tfrac{\partial L^k}{\partial \mu_{i,x}^k}\bigr)^2+\bigl(\tfrac{\partial L^k}{\partial \mu_{i,y}^k}\bigr)^2}}{\sum_{k=1}^M \omega_i^k}.$$

Densification is then triggered by $g_i>\tau_{\mathrm{pos}}$, with splitting for large-scale Gaussians and cloning otherwise (Cheng et al., 21 Oct 2025).

EntON introduces a geometry-aware alternative based on local neighborhood structure. For each Gaussian center, it computes a $k$-nearest-neighbor covariance matrix, extracts its eigenvalues, normalizes them, and defines Eigenentropy as

$$\mathcal E_\pi(u)=1-\mathrm{SSIM}\bigl(C_\pi(u),T_\pi(u)\bigr),$$0

The method alternates between standard gradient-based densification and an Eigenentropy-aware pass that splits Gaussians in low-Eigenentropy neighborhoods and prunes Gaussians in high-Eigenentropy neighborhoods if opacity is low (Jäger et al., 6 Mar 2026).

3. Geometry-aware validation, placement, and insertion

A second design axis concerns where new primitives are placed and how geometric plausibility is enforced. GeoTexDensifier modifies splitting directly. Each Gaussian stores a reference view index, a reference depth ratio, and a reference raster weight. Before splitting, the parent is checked by a self-Validation-of-Depth-Ratio-Change; candidate children are then checked by a child-Validation-of-Depth-Ratio-Change using

$$\mathcal E_\pi(u)=1-\mathrm{SSIM}\bigl(C_\pi(u),T_\pi(u)\bigr),$$1

with $$\mathcal E_\pi(u)=1-\mathrm{SSIM}\bigl(C_\pi(u),T_\pi(u)\bigr),$$2 as the stated threshold. Accepted child positions are further adjusted by projecting sampled points onto the tangent plane defined by a surface normal and modulating the offset by image gradients. This design uses monocular depth prior and surface normals in a scale-invariant way to prevent runaway growth in poorly constrained regions (Jiang et al., 2024).

ConeGS abandons clone-and-split heuristics entirely. It computes a per-pixel reconstruction error

$$\mathcal E_\pi(u)=1-\mathrm{SSIM}\bigl(C_\pi(u),T_\pi(u)\bigr),$$3

samples high-error pixels from the induced discrete distribution, queries a pre-trained iNGP model for the median depth along each selected ray, and inserts a new Gaussian at

$$\mathcal E_\pi(u)=1-\mathrm{SSIM}\bigl(C_\pi(u),T_\pi(u)\bigr),$$4

Its initial covariance is set from the viewing-cone radius,

$$\mathcal E_\pi(u)=1-\mathrm{SSIM}\bigl(C_\pi(u),T_\pi(u)\bigr),$$5

using an isotropic initialization

$$\mathcal E_\pi(u)=1-\mathrm{SSIM}\bigl(C_\pi(u),T_\pi(u)\bigr),$$6

This makes densification independent of existing scene geometry state, while a pre-activation opacity penalty and primitive budgeting control redundancy (Baranowski et al., 10 Nov 2025).

VAD-GS extends geometry-aware insertion to scenes with missing structure. It first detects unreliable regions by voxel-based visibility reasoning, comparing voxel depth with Gaussian-rendered depth. It then selects support views using a diversity score

$$\mathcal E_\pi(u)=1-\mathrm{SSIM}\bigl(C_\pi(u),T_\pi(u)\bigr),$$7

runs patch matching-based multi-view stereo, and initializes brand-new Gaussians from recovered points, normals, and colors. The paper explicitly motivates this as a remedy for the limitation that clone/split-only strategies are incapable of reconstructing missing structures (Zhang et al., 10 Oct 2025).

4. Scheduling, pruning, and capacity control

Re-densification modules also differ in optimization schedule. "Decomposing Densification in Gaussian Splatting for Faster 3D Scene Reconstruction" separates growth into a global-to-local schedule. Phase I, “Global Spread,” allows only split operations up to a boundary $$\mathcal E_\pi(u)=1-\mathrm{SSIM}\bigl(C_\pi(u),T_\pi(u)\bigr),$$8; Phase II, “Local Refinement,” allows both split and clone. The same method introduces an energy-guided coarse-to-fine multi-resolution framework, with resolution-specific densification iterations

$$\mathcal E_\pi(u)=1-\mathrm{SSIM}\bigl(C_\pi(u),T_\pi(u)\bigr),$$9

and an adaptive opacity-pruning rule based on sorted opacity values and an upper limit $$E_k^\pi=\sum_{u\in\text{Pixels}} \mathcal E_\pi(u)\,\omega_k^\pi(u),$$0 (Huang et al., 27 Jul 2025).

Mini-Splatting2 pushes scheduling in the opposite direction through aggressive densification. Between 500 and 3000 iterations, every 250 iterations it identifies critical Gaussians by the largest per-pixel blending weight,

$$E_k^\pi=\sum_{u\in\text{Pixels}} \mathcal E_\pi(u)\,\omega_k^\pi(u),$$1

takes the $$E_k^\pi=\sum_{u\in\text{Pixels}} \mathcal E_\pi(u)\,\omega_k^\pi(u),$$2 quantile of the corresponding maximal weights, and clones each selected Gaussian twice. At 2000 iterations it performs depth re-initialization, and at 3000 and 8000 iterations it applies simplification. It further precomputes per-view Gaussian importance and visibility masks for culling (Fang et al., 2024).

AD-GS formalizes alternation between aggressive growth and controlled geometric consolidation. After a warm-up of $$E_k^\pi=\sum_{u\in\text{Pixels}} \mathcal E_\pi(u)\,\omega_k^\pi(u),$$3 iterations, it alternates blocks of $$E_k^\pi=\sum_{u\in\text{Pixels}} \mathcal E_\pi(u)\,\omega_k^\pi(u),$$4 high-densification iterations and $$E_k^\pi=\sum_{u\in\text{Pixels}} \mathcal E_\pi(u)\,\omega_k^\pi(u),$$5 low-densification iterations. The low-densification phase prunes Gaussians with $$E_k^\pi=\sum_{u\in\text{Pixels}} \mathcal E_\pi(u)\,\omega_k^\pi(u),$$6, uses a stricter densification threshold $$E_k^\pi=\sum_{u\in\text{Pixels}} \mathcal E_\pi(u)\,\omega_k^\pi(u),$$7, and optimizes a combined objective

$$E_k^\pi=\sum_{u\in\text{Pixels}} \mathcal E_\pi(u)\,\omega_k^\pi(u),$$8

where $$E_k^\pi=\sum_{u\in\text{Pixels}} \mathcal E_\pi(u)\,\omega_k^\pi(u),$$9 enforces pseudo-view consistency between two parallel 3DGS models and $E_k=\max_{\pi\in\Pi}E_k^\pi.$0 is an edge-aware depth-smoothness term (Patle et al., 13 Sep 2025).

Budgeting appears repeatedly. "Revising Densification in Gaussian Splatting" limits the number of offspring by

$E_k=\max_{\pi\in\Pi}E_k^\pi.$1

and crops the candidate set to the top-$E_k=\max_{\pi\in\Pi}E_k^\pi.$2 primitives by descending error. ConeGS supports both a fixed budget and an adaptive growth rule, while IDHFR uses a dynamic $E_k=\max_{\pi\in\Pi}E_k^\pi.$3 that grows from the initialization size to a user-defined limit (Bulò et al., 2024).

5. Dynamic, generative, and cross-domain variants

In dynamic 3DGS, temporality changes the meaning of density control. The Visibility-Aware Densification framework computes a visibility-weighted gradient average

$E_k=\max_{\pi\in\Pi}E_k^\pi.$4

so that only frames in which a Gaussian actually contributes influence its densification score. It then lowers the threshold for short-lived Gaussians through Temporally-Adaptive Thresholding,

$E_k=\max_{\pi\in\Pi}E_k^\pi.$5

and increases local temporal modeling capacity around temporal centers through Temporal Offset Warping (Sandu et al., 22 Jun 2026).

GeoLRM uses the term in a different but related sense. Its two-stage pipeline first predicts occupied 3D anchors from posed images, then applies a Reconstruction Transformer that refines the geometry and retrieves textural details. Each refined token is decoded into $E_k=\max_{\pi\in\Pi}E_k^\pi.$6 Gaussians, producing up to $E_k=\max_{\pi\in\Pi}E_k^\pi.$7k Gaussians on $E_k=\max_{\pi\in\Pi}E_k^\pi.$8k tokens. The core operations are self-attention with 3D RoPE, deformable cross-attention, and a Gaussian decoder with sigmoid-activated offsets, scales, and opacities (Zhang et al., 2024).

Outside 3DGS, LiDAR compression uses re-densification as a feature-extraction device rather than as a scene-reconstruction primitive generator. The Geometry Re-Densification Module in "Re-Densification Meets Cross-Scale Propagation" folds sparse octree occupancy codes from a deep level to a shallower denser level, extracts features with a ResBlock, and then recursively upsamples and prunes them back to the target level for occupancy prediction (Yu et al., 28 Aug 2025).

Template-based shape reconstruction uses a continuous reallocation of vertex density without changing connectivity. Mesh Density Adaptation defines the local average edge-length vector $E_k=\max_{\pi\in\Pi}E_k^\pi.$9 and minimizes

$E_k>\epsilon$0

often through a diffusion re-parameterization

$E_k>\epsilon$1

This formulation allows uniform and curvature-driven re-densification while preserving mesh topology (Jung et al., 2023).

SuperMeshingNet applies the concept to numerical computation in a 2D domain. It learns a mapping

$E_k>\epsilon$2

from a low-density FEM stress field to a high-density stress field, using residual dense blocks and pixel-shuffle upsampling for factors $E_k>\epsilon$3. Here re-densification is not a runtime primitive-management procedure but a learned mesh-density boost model (Xu et al., 2021).

6. Empirical patterns, limitations, and open issues

The reported gains are substantial but heterogeneous. EntON reports up to $E_k>\epsilon$4 improvement in geometric accuracy and up to $E_k>\epsilon$5 improvement in rendering quality, while reducing the number of Gaussians by up to $E_k>\epsilon$6 and training time by up to $E_k>\epsilon$7; on DTU it reports Chamfer C2C from $E_k>\epsilon$8 mm to $E_k>\epsilon$9 mm and Gaussian count from $\hat\alpha = 1-\sqrt{1-\alpha},$0k to $\hat\alpha = 1-\sqrt{1-\alpha},$1k (Jäger et al., 6 Mar 2026). The global-to-local densification framework reports training time reductions from $\hat\alpha = 1-\sqrt{1-\alpha},$2 min to $\hat\alpha = 1-\sqrt{1-\alpha},$3 min on MipNeRF-360, Deep Blending, and Tanks & Temples, with more than $\hat\alpha = 1-\sqrt{1-\alpha},$4 speedup and fewer Gaussians (Huang et al., 27 Jul 2025).

Other works emphasize different trade-offs. Mini-Splatting2 reports convergence in approximately $\hat\alpha = 1-\sqrt{1-\alpha},$5 m $\hat\alpha = 1-\sqrt{1-\alpha},$6 s with approximately $\hat\alpha = 1-\sqrt{1-\alpha},$7 M final Gaussians and PSNR $\hat\alpha = 1-\sqrt{1-\alpha},$8 dB, versus approximately $\hat\alpha = 1-\sqrt{1-\alpha},$9 m, approximately $\alpha$0 M Gaussians, and PSNR $\alpha$1 dB for vanilla 3DGS; it characterizes this as a $\alpha$2 speed-up at less than $\alpha$3 dB loss (Fang et al., 2024). ReAct-GS reports, on Mip-NeRF360, PSNR from $\alpha$4 to $\alpha$5, SSIM from $\alpha$6 to $\alpha$7, and LPIPS from $\alpha$8 to $\alpha$9, while explicitly targeting gradient dilution and frozen primitives (Cheng et al., 21 Oct 2025).

Cross-domain results confirm that the same high-level idea can serve different objectives. GeoLRM reports $$g_i = \frac{\sum_{k=1}^M \omega_i^k \sqrt{\bigl(\tfrac{\partial L^k}{\partial \mu_{i,x}^k}\bigr)^2+\bigl(\tfrac{\partial L^k}{\partial \mu_{i,y}^k}\bigr)^2}}{\sum_{k=1}^M \omega_i^k}.$$0k Gaussians with $$g_i = \frac{\sum_{k=1}^M \omega_i^k \sqrt{\bigl(\tfrac{\partial L^k}{\partial \mu_{i,x}^k}\bigr)^2+\bigl(\tfrac{\partial L^k}{\partial \mu_{i,y}^k}\bigr)^2}}{\sum_{k=1}^M \omega_i^k}.$$1 input images in only $$g_i = \frac{\sum_{k=1}^M \omega_i^k \sqrt{\bigl(\tfrac{\partial L^k}{\partial \mu_{i,x}^k}\bigr)^2+\bigl(\tfrac{\partial L^k}{\partial \mu_{i,y}^k}\bigr)^2}}{\sum_{k=1}^M \omega_i^k}.$$2 GB GPU memory and a Chamfer Distance of $$g_i = \frac{\sum_{k=1}^M \omega_i^k \sqrt{\bigl(\tfrac{\partial L^k}{\partial \mu_{i,x}^k}\bigr)^2+\bigl(\tfrac{\partial L^k}{\partial \mu_{i,y}^k}\bigr)^2}}{\sum_{k=1}^M \omega_i^k}.$$3 versus $$g_i = \frac{\sum_{k=1}^M \omega_i^k \sqrt{\bigl(\tfrac{\partial L^k}{\partial \mu_{i,x}^k}\bigr)^2+\bigl(\tfrac{\partial L^k}{\partial \mu_{i,y}^k}\bigr)^2}}{\sum_{k=1}^M \omega_i^k}.$$4 for InstantMesh on Google Scanned Objects (Zhang et al., 2024). The LiDAR GRED framework reports state-of-the-art compression ratios and real-time performance, achieving $$g_i = \frac{\sum_{k=1}^M \omega_i^k \sqrt{\bigl(\tfrac{\partial L^k}{\partial \mu_{i,x}^k}\bigr)^2+\bigl(\tfrac{\partial L^k}{\partial \mu_{i,y}^k}\bigr)^2}}{\sum_{k=1}^M \omega_i^k}.$$5 FPS for both encoding and decoding at $$g_i = \frac{\sum_{k=1}^M \omega_i^k \sqrt{\bigl(\tfrac{\partial L^k}{\partial \mu_{i,x}^k}\bigr)^2+\bigl(\tfrac{\partial L^k}{\partial \mu_{i,y}^k}\bigr)^2}}{\sum_{k=1}^M \omega_i^k}.$$6-bit quantization (Yu et al., 28 Aug 2025). SuperMeshingNet reports an average relative stress error of $$g_i = \frac{\sum_{k=1}^M \omega_i^k \sqrt{\bigl(\tfrac{\partial L^k}{\partial \mu_{i,x}^k}\bigr)^2+\bigl(\tfrac{\partial L^k}{\partial \mu_{i,y}^k}\bigr)^2}}{\sum_{k=1}^M \omega_i^k}.$$7 and a $$g_i = \frac{\sum_{k=1}^M \omega_i^k \sqrt{\bigl(\tfrac{\partial L^k}{\partial \mu_{i,x}^k}\bigr)^2+\bigl(\tfrac{\partial L^k}{\partial \mu_{i,y}^k}\bigr)^2}}{\sum_{k=1}^M \omega_i^k}.$$8 speed-up for $$g_i = \frac{\sum_{k=1}^M \omega_i^k \sqrt{\bigl(\tfrac{\partial L^k}{\partial \mu_{i,x}^k}\bigr)^2+\bigl(\tfrac{\partial L^k}{\partial \mu_{i,y}^k}\bigr)^2}}{\sum_{k=1}^M \omega_i^k}.$$9 output compared with FEM alone (Xu et al., 2021).

The literature also identifies several misconceptions. "The Role and Relationship of Initialization and Densification in 3D Gaussian Splatting" states that current densification approaches are not able to take full advantage of dense initialization and are often unable to significantly improve over sparse SfM-based initialization. It further reports that MCMC and IDHFR are nearly initialization-invariant with $g_i>\tau_{\mathrm{pos}}$0 when moving from sparse SfM to EDGS or laser scan, while uniformly dense initializations can degrade MCMC or IDHFR performance on on-trajectory views by approximately $g_i>\tau_{\mathrm{pos}}$1 PSNR because too many small Gaussians are created in low-frequency regions and capacity is wasted (Desiatov et al., 21 Mar 2026). A plausible implication is that geometry re-densification is not reducible to adding more primitives; criterion design, placement, pruning, and scheduling are all first-order variables.

A second recurring tension is between geometry and photometry. EntON states that surface-focused reconstruction methods frequently sacrifice photometric accuracy, whereas its alternating densification aims to balance geometric accuracy, rendering quality, and efficiency (Jäger et al., 6 Mar 2026). Sparse-view methods such as AD-GS, depth-validated methods such as GeoTexDensifier, and visibility-reasoning methods such as VAD-GS all encode the same concern in different ways: uncontrolled growth can increase primitive count while worsening floaters, geometry, or generalization. This suggests that the modern geometry re-densification module is best understood as a control system for representation allocation rather than as a single split-or-clone heuristic.