Global-to-Local Densification Strategy

• Global-to-local densification strategy is a method that initially leverages global signals to form a coarse scene structure before engaging local cues for detailed refinement.
• It employs a coarse-to-fine scheduling mechanism to selectively allocate computational resources, balancing global consistency with localized accuracy.
• The approach improves reconstruction fidelity and efficiency in applications such as 3D modeling, dynamic scene analysis, and depth estimation.

Searching arXiv for papers on global-to-local densification and related methods. A global-to-local densification strategy is a class of adaptive refinement procedures in which a model first relies on broad, globally aggregated signals or coarse scene structure, and subsequently allocates additional representational capacity according to finer local cues. Across the literature, the phrase denotes related but not identical mechanisms: alternating between scene-wide photometric supervision and local geometric regularization in 3D Gaussian Splatting (Jäger et al., 6 Mar 2026), staged split-versus-clone scheduling for scene-wide coverage before local detail refinement (Huang et al., 27 Jul 2025), self-supervised conversion of global reconstruction gains into local primitive rewards for feedforward reconstruction (Song et al., 28 Jun 2026), temporally aggregated visibility statistics followed by per-Gaussian decisions in dynamic 3DGS (Sandu et al., 22 Jun 2026), and global-context extraction followed by local refinement in dense prediction for panoramic depth estimation (Bai et al., 2022). In its most general form, the strategy is defined by a directional allocation principle: global information establishes coverage, consistency, or feasibility, while local information determines where densification, refinement, or pruning should occur.

1. General definition and conceptual scope

In the most direct formulation, a global-to-local densification strategy begins from a global signal that is either scene-wide, temporally aggregated, or otherwise coarse in scale, and then converts that signal into localized structural updates. In 3D Gaussian Splatting, this global signal may be image reconstruction loss and accumulated view-space position gradients over many views (Jäger et al., 6 Mar 2026). In feedforward dynamic Gaussian reconstruction, it may be the reduction in rendering error after tentative densification, which is then attributed back to candidate primitives (Song et al., 28 Jun 2026). In dynamic 3DGS, the global component may instead be temporal aggregation over a densification interval, after which each Gaussian receives a local visibility-aware decision (Sandu et al., 22 Jun 2026). In panoramic depth estimation, the same principle appears as globally coherent transformer features that are later fused with local convolutional features for dense prediction (Bai et al., 2022).

This suggests that “global” and “local” should not be restricted to spatial scale alone. The global component can be scene-wide optimization, whole-route calibration, temporally accumulated evidence, or full-field contextual modeling, while the local component can be a neighborhood descriptor, a per-primitive reward, a per-segment operating point, or a boundary-conditioned local subproblem. A plausible implication is that the unifying feature is not the exact representation but the sequencing of refinement: coarse consistency first, selective detail allocation second.

Within Gaussian-based scene reconstruction, the notion is often motivated by a mismatch between what baseline densification optimizes directly and what it only captures indirectly. EntON explicitly frames densification in 3DGS as a mismatch between photometric rendering and geometric organization of Gaussians around true surfaces (Jäger et al., 6 Mar 2026). RobustSplat++ similarly argues that early position-gradient-driven growth in in-the-wild scenes overfits transient disturbances and illumination variation, so it delays Gaussian growth until static scene structure is better stabilized (Fu et al., 4 Dec 2025). Decomposed densification in Gaussian Splatting formalizes the distinction further by arguing that split supports global diffusion whereas clone supports local refinement, and that conflating them from the outset leads to inefficient primitive allocation (Huang et al., 27 Jul 2025).

2. Global signals as densification priors

A canonical global signal in 3DGS is the accumulated view-space position gradient. EntON writes the gradient accumulation as

$$\nabla_{\boldsymbol{\mu}} L = \sum_{i} \frac{\partial L}{\partial \mathbf{p}_i} \cdot \frac{\partial \mathbf{p}_i}{\partial \boldsymbol{\mu}},$$

and the densification statistic as

$$\bar{g} = \frac{1}{N} \sum_{j=1}^{N} \left\| \nabla_{\boldsymbol{\mu}^{(j)}} L \right\|_2.$$

In standard 3DGS, densification is triggered when $\bar g$ exceeds $\tau_{\mathrm{pos}}$, with default $\tau_{\mathrm{pos}}=0.0002$ (Jäger et al., 6 Mar 2026). These gradients are global because they arise from multi-view photometric supervision and are accumulated scene-wide over training. EntON retains this signal and alternates back to it so that high-contribution Gaussians are not lost even when local geometry estimates remain unreliable (Jäger et al., 6 Mar 2026).

In decomposed densification for Gaussian Splatting, the same baseline mechanism is reinterpreted through the distinct operational roles of split and clone. The paper argues that standard densification entangles global diffusion and local detail preservation: split is associated with scene-wide propagation, whereas clone mainly increases local capacity in already represented areas (Huang et al., 27 Jul 2025). This motivates a global-to-local schedule in which split is used first to establish spatial coverage and clone is reintroduced later for refinement (Huang et al., 27 Jul 2025).

Other works define the global signal differently. L2D2-GS uses the reduction in reconstruction error between coarse and densified scenes,

$\Delta E = E_c - E_f,$

as a global gain map that measures the utility of densification at the image level (Song et al., 28 Jun 2026). Dynamic 3DGS with temporally aware densification aggregates per-frame statistics over a densification interval $\mathcal N$, then uses the visibility-weighted average

$$\bar{g}_i = \frac{\sum_{t\in \mathcal{N}} \sigma_{i}(t)\, g_{i}(t)}{\sum_{t\in\mathcal{N}} \sigma_{i}(t)}$$

to avoid underestimating short-lived dynamic Gaussians (Sandu et al., 22 Jun 2026). GLPanoDepth uses a transformer branch whose self-attention is described as “an inherently global operation, as every embedding token can attend to and thus influence other tokens globally,” making global context the first stage of the depth-prediction pipeline (Bai et al., 2022).

A plausible implication is that the global component plays two recurring roles. First, it provides robustness by pooling evidence beyond the scale of any single local structure. Second, it establishes the initial support on which finer local decisions become meaningful. EntON states this explicitly: the first 3000 iterations are pure standard 3DGS densification so that local covariance and eigenvalue estimates become meaningful only after sufficient Gaussian density exists (Jäger et al., 6 Mar 2026).

3. Local criteria and selective refinement

The local stage distinguishes global-to-local densification from uniform capacity growth. EntON constructs a $k$-nearest-neighbor neighborhood around each Gaussian center, computes the covariance

$$C = \frac{1}{k+1} \sum_{i=0}^{k} (p_i - \bar{p})(p_i - \bar{p})^T, \qquad \bar{p} = \frac{1}{k+1} \sum_{i=0}^{k} p_i,$$

extracts eigenvalues $$\bar{g} = \frac{1}{N} \sum_{j=1}^{N} \left\| \nabla_{\boldsymbol{\mu}^{(j)}} L \right\|_2.$$0, normalizes them,

$$\bar{g} = \frac{1}{N} \sum_{j=1}^{N} \left\| \nabla_{\boldsymbol{\mu}^{(j)}} L \right\|_2.$$1

and defines Eigenentropy as

$$\bar{g} = \frac{1}{N} \sum_{j=1}^{N} \left\| \nabla_{\boldsymbol{\mu}^{(j)}} L \right\|_2.$$2

This quantity measures local structural order or disorder: very low values correspond to highly anisotropic linear structure, the ideal planar case gives $$\bar{g} = \frac{1}{N} \sum_{j=1}^{N} \left\| \nabla_{\boldsymbol{\mu}^{(j)}} L \right\|_2.$$3, and isotropic neighborhoods approach $$\bar{g} = \frac{1}{N} \sum_{j=1}^{N} \left\| \nabla_{\boldsymbol{\mu}^{(j)}} L \right\|_2.$$4 (Jäger et al., 6 Mar 2026). EntON then splits Gaussians in low-Eigenentropy neighborhoods, prunes them in high-Eigenentropy neighborhoods, and leaves intermediate cases unchanged, with

$$\bar{g} = \frac{1}{N} \sum_{j=1}^{N} \left\| \nabla_{\boldsymbol{\mu}^{(j)}} L \right\|_2.$$5

This is a local surface-seeking rule built on neighborhood geometry rather than view-space gradients alone (Jäger et al., 6 Mar 2026).

L2D2-GS derives local decisions differently. After computing global reconstruction gain $$\bar{g} = \frac{1}{N} \sum_{j=1}^{N} \left\| \nabla_{\boldsymbol{\mu}^{(j)}} L \right\|_2.$$6, it rasterizes the contribution mask of candidate densified Gaussians and defines

$$\bar{g} = \frac{1}{N} \sum_{j=1}^{N} \left\| \nabla_{\boldsymbol{\mu}^{(j)}} L \right\|_2.$$7

This produces a per-primitive rendering contribution score, which becomes part of a local utility

$$\bar{g} = \frac{1}{N} \sum_{j=1}^{N} \left\| \nabla_{\boldsymbol{\mu}^{(j)}} L \right\|_2.$$8

The densification policy then predicts per-candidate probabilities $$\bar{g} = \frac{1}{N} \sum_{j=1}^{N} \left\| \nabla_{\boldsymbol{\mu}^{(j)}} L \right\|_2.$$9 and optimizes

$\bar g$0

with binary entropy

$\bar g$1

Here the local decision is learned, but its supervision is generated from a global scene-level objective (Song et al., 28 Jun 2026).

Temporally aware densification in dynamic 3DGS also ends in local per-Gaussian decisions. After visibility aggregation, each Gaussian receives a temporally adaptive threshold

$\bar g$2

where $\bar g$3 is an average temporal lifespan statistic, with $\bar g$4 and $\bar g$5 in the paper (Sandu et al., 22 Jun 2026). This relaxes densification thresholds for short-lived Gaussians while preserving the same global functional form across the model.

Metamon-GS uses a different local statistic: the variance of per-pixel color gradients associated with each Gaussian. The online updates are

$\bar g$6

A Gaussian is selected when

$\bar g$7

with $\bar g$8 and $\bar g$9 (Su et al., 20 Apr 2025). The paper explicitly states that this is not an explicitly designed global-to-local schedule, but it does produce adaptive local density compensation in thinly initialized or high-detail regions (Su et al., 20 Apr 2025).

4. Scheduling mechanisms: coarse-to-fine, alternating, and delayed growth

A defining property of many global-to-local densification strategies is not only the choice of local descriptor but the schedule by which it is activated. EntON makes this especially explicit. The first 3000 iterations use standard gradient-based densification only. After that, densification is invoked every 100 iterations, alternating every 100 iterations between standard 3DGS densification and Eigenentropy-aware densification (Jäger et al., 6 Mar 2026). The paper summarizes the procedure as: if $\tau_{\mathrm{pos}}$0, then before iteration 3000 use gradient-based densification; afterwards alternate between gradient-based and Eigenentropy-aware steps based on $\tau_{\mathrm{pos}}$1 (Jäger et al., 6 Mar 2026). The method also lowers the gradient threshold in its 3DGS-like step to $\tau_{\mathrm{pos}}$2, and Gaussians are “divided further depending on their size into 2, 4 or 8 Gaussians,” inspired by Micro-splatting (Jäger et al., 6 Mar 2026).

Decomposed densification in Gaussian Splatting proposes a related but simpler schedule: early split-only densification for global spread, later split-plus-clone densification for local refinement (Huang et al., 27 Jul 2025). This is paired with an energy-guided coarse-to-fine multi-resolution training framework. For an image $\tau_{\mathrm{pos}}$3, the frequency-domain energy is

$\tau_{\mathrm{pos}}$4

and for a downsampling factor $\tau_{\mathrm{pos}}$5,

$\tau_{\mathrm{pos}}$6

The number of densification iterations assigned to scale $\tau_{\mathrm{pos}}$7 is

$\tau_{\mathrm{pos}}$8

with $\tau_{\mathrm{pos}}$9 and $\tau_{\mathrm{pos}}=0.0002$0 in the default setting (Huang et al., 27 Jul 2025). The transition from global spread to local refinement is aligned with this coarse-to-fine resolution schedule.

RobustSplat++ implements a different scheduling principle: delayed Gaussian growth. Instead of redefining the split/clone operator, it delays when densification is allowed. The delayed iteration start is set to 10K in a 30K training schedule, in contrast to vanilla 3DGS, which begins densification by default after 500 iterations (Fu et al., 4 Dec 2025). Before 10K, the Gaussian count is fixed while the model optimizes static scene structure, mask estimation, and, in the full variant, appearance embeddings (Fu et al., 4 Dec 2025). The paper describes this as prioritizing global static structure before allowing local splitting and cloning, and pairs it with scale-cascaded mask bootstrapping that uses low-resolution supervision before densification and high-resolution supervision after it starts (Fu et al., 4 Dec 2025).

GLPanoDepth expresses a closely related principle in dense prediction. It first extracts global features from Cubemap Vision Transformers and only then refines them with local CNN features through gated multi-scale fusion: $\tau_{\mathrm{pos}}=0.0002$1

$\tau_{\mathrm{pos}}=0.0002$2

The paper describes the output as “fine-grained and globally coherent,” making the global-to-local staging explicit even though the task is depth prediction rather than primitive growth (Bai et al., 2022).

5. Variants across dynamic scenes, retrieval, planning, and graph analysis

The concept generalizes beyond static 3DGS. In dynamic Gaussian reconstruction, L2D2-GS frames densification as a learned policy rather than a hand-crafted heuristic. Reconstruction alternates between continuous parameter updates

$\tau_{\mathrm{pos}}=0.0002$3

and a discrete densification event that selects new primitives from a large candidate set (Song et al., 28 Jun 2026). The global-to-local aspect is explicit in the attribution chain “scene-level gain $\tau_{\mathrm{pos}}=0.0002$4 pixel-level gain map $\tau_{\mathrm{pos}}=0.0002$5 visibility-weighted primitive attribution $\tau_{\mathrm{pos}}=0.0002$6 local densification reward” (Song et al., 28 Jun 2026).

Temporally aware densification in dynamic 3DGS uses a temporal rather than purely spatial global signal. The method keeps per-Gaussian temporal centers $\tau_{\mathrm{pos}}=0.0002$7, visibility proxies $\tau_{\mathrm{pos}}=0.0002$8, and Fourier-based motion centered at $\tau_{\mathrm{pos}}=0.0002$9 (Sandu et al., 22 Jun 2026). It also introduces Temporal Offset Warping,

$\Delta E = E_c - E_f,$0

with

$\Delta E = E_c - E_f,$1

to allocate more temporal modeling capacity near each Gaussian’s active temporal region (Sandu et al., 22 Jun 2026). This is a temporally global, Gaussian-local densification strategy rather than a purely spatial one.

In long-range 4D motion modeling, MoRel starts from a single Global Canonical Anchor and progressively specializes it into Key-frame Anchors, which serve as local canonical spaces. Feature-variance-guided Hierarchical Densification assigns each global anchor a level

$\Delta E = E_c - E_f,$2

where $\Delta E = E_c - E_f,$3, and modulates densification by

$\Delta E = E_c - E_f,$4

Low-frequency anchors are prioritized early, high-frequency anchors later, yielding a coarse-to-fine hierarchy inside temporally localized anchor spaces (Kwak et al., 10 Dec 2025).

Outside rendering, the same structural pattern appears in retrieval and planning. The retrieval paper on local-to-global search does not use the term “densification strategy” explicitly, but it transforms sparse local retrieval similarities into dense query-specific global embeddings via MDS, which can be read as a densification-like mechanism from local evidence to global representation (Aiger et al., 4 Sep 2025). GuILD for motion planning begins with the global Informed Set,

$\Delta E = E_c - E_f,$5

and then refines it into local subsets induced by an intermediate beacon $\Delta E = E_c - E_f,$6,

$\Delta E = E_c - E_f,$7

thereby converting global informed sampling into adaptive local informed sampling (Mandalika et al., 2021). In graph analysis, local density decomposes a global densest-subgraph objective into per-vertex quantities satisfying

$\Delta E = E_c - E_f,$8

which can be interpreted as a global-to-local density characterization rather than a constructive densification process (Christiansen et al., 2024).

6. Performance trade-offs, limitations, and recurring misconceptions

A recurring trade-off in global-to-local densification is between geometric fidelity, photometric quality, compactness, and computational cost. EntON makes this explicit through the neighborhood size $\Delta E = E_c - E_f,$9: small $\mathcal N$0 improves geometric accuracy, while larger $\mathcal N$1 preserves rendering quality (Jäger et al., 6 Mar 2026). On DTU, the best geometry is achieved at $\mathcal N$2, with mean C2C $\mathcal N$3 versus $\mathcal N$4 for 3DGS, while $\mathcal N$5 yields $\mathcal N$6 PSNR close to 3DGS’s $\mathcal N$7 (Jäger et al., 6 Mar 2026). The recommended balance is $\mathcal N$8, with $\mathcal N$9 C2C, $$\bar{g}_i = \frac{\sum_{t\in \mathcal{N}} \sigma_{i}(t)\, g_{i}(t)}{\sum_{t\in\mathcal{N}} \sigma_{i}(t)}$$0 PSNR, $$\bar{g}_i = \frac{\sum_{t\in \mathcal{N}} \sigma_{i}(t)\, g_{i}(t)}{\sum_{t\in\mathcal{N}} \sigma_{i}(t)}$$1 Gaussians, and $$\bar{g}_i = \frac{\sum_{t\in \mathcal{N}} \sigma_{i}(t)\, g_{i}(t)}{\sum_{t\in\mathcal{N}} \sigma_{i}(t)}$$2 minutes (Jäger et al., 6 Mar 2026). The same paper also shows that EntON can reduce mean Eigenentropy from about $$\bar{g}_i = \frac{\sum_{t\in \mathcal{N}} \sigma_{i}(t)\, g_{i}(t)}{\sum_{t\in\mathcal{N}} \sigma_{i}(t)}$$3 to about $$\bar{g}_i = \frac{\sum_{t\in \mathcal{N}} \sigma_{i}(t)\, g_{i}(t)}{\sum_{t\in\mathcal{N}} \sigma_{i}(t)}$$4 after 15K iterations and reduce Gaussian count from $$\bar{g}_i = \frac{\sum_{t\in \mathcal{N}} \sigma_{i}(t)\, g_{i}(t)}{\sum_{t\in\mathcal{N}} \sigma_{i}(t)}$$5 to $$\bar{g}_i = \frac{\sum_{t\in \mathcal{N}} \sigma_{i}(t)\, g_{i}(t)}{\sum_{t\in\mathcal{N}} \sigma_{i}(t)}$$6 at $$\bar{g}_i = \frac{\sum_{t\in \mathcal{N}} \sigma_{i}(t)\, g_{i}(t)}{\sum_{t\in\mathcal{N}} \sigma_{i}(t)}$$7 (Jäger et al., 6 Mar 2026).

L2D2-GS shows a different form of the same trade-off. Removing densification drops performance from $$\bar{g}_i = \frac{\sum_{t\in \mathcal{N}} \sigma_{i}(t)\, g_{i}(t)}{\sum_{t\in\mathcal{N}} \sigma_{i}(t)}$$8 to $$\bar{g}_i = \frac{\sum_{t\in \mathcal{N}} \sigma_{i}(t)\, g_{i}(t)}{\sum_{t\in\mathcal{N}} \sigma_{i}(t)}$$9 PSNR on PandaSet full-sequence reconstruction even when the final primitive budget is matched, indicating that the gain comes from where and when primitives are added rather than from primitive count alone (Song et al., 28 Jun 2026). The full method reaches $k$0 PSNR, $k$1 SSIM, and $k$2 LPIPS with $k$3M Gaussians, compared with reproduced G3R at $k$4, $k$5, and $k$6 using $k$7M Gaussians (Song et al., 28 Jun 2026).

Temporally aware densification shows that naive threshold lowering is not equivalent to a temporally aware strategy. The paper reports that globally lowering $k$8 increases model size quickly with limited gain, whereas visibility-aware and lifespan-aware densification produces better masked dynamic-region metrics at better model size (Sandu et al., 22 Jun 2026). On VRU Basketball, the full method improves M-PSNR from $k$9 to $$C = \frac{1}{k+1} \sum_{i=0}^{k} (p_i - \bar{p})(p_i - \bar{p})^T, \qquad \bar{p} = \frac{1}{k+1} \sum_{i=0}^{k} p_i,$$0 and M-SSIM from $$C = \frac{1}{k+1} \sum_{i=0}^{k} (p_i - \bar{p})(p_i - \bar{p})^T, \qquad \bar{p} = \frac{1}{k+1} \sum_{i=0}^{k} p_i,$$1 to $$C = \frac{1}{k+1} \sum_{i=0}^{k} (p_i - \bar{p})(p_i - \bar{p})^T, \qquad \bar{p} = \frac{1}{k+1} \sum_{i=0}^{k} p_i,$$2 relative to its baseline (Sandu et al., 22 Jun 2026).

Several misconceptions recur across the literature. One is that global-to-local densification is synonymous with hierarchical mesh refinement. The surveyed papers show otherwise: in 3DGS it may mean alternating between gradient-based and neighborhood-based densification (Jäger et al., 6 Mar 2026), in dynamic scenes it may mean temporal aggregation followed by per-Gaussian thresholding (Sandu et al., 22 Jun 2026), and in retrieval it may mean converting sparse local similarities into dense global embeddings (Aiger et al., 4 Sep 2025). Another misconception is that better densification is simply a matter of lowering thresholds or adding more primitives. RobustSplat++ and decomposed densification both argue that premature or poorly scheduled growth can actively harm reconstruction by overfitting transients or causing clone-heavy clustering (Fu et al., 4 Dec 2025, Huang et al., 27 Jul 2025).

Limitations also vary by method. EntON is especially suited to man-made, piecewise planar scenes and is less appropriate for highly curved, irregular, or organic scenes such as vegetation (Jäger et al., 6 Mar 2026). Its kNN computation can be time-consuming, especially in large scenes (Jäger et al., 6 Mar 2026). GLPanoDepth notes that the standalone transformer branch is data-hungry and can fail to converge on smaller or noisier panoramic datasets (Bai et al., 2022). MoRel’s feature-variance hierarchy depends on a variance proxy whose exact semantics are tied to learned anchor features rather than direct geometric measurements (Kwak et al., 10 Dec 2025). Map density selection for VPR is global-to-local in requirement structure, but its current formulation still selects one global density rather than region-specific densities (Hussaini et al., 25 Feb 2026). These limits suggest that the strategy is robust as a design pattern, but its effectiveness depends on how global evidence and local refinement are instantiated in a given domain.

7. Historical and methodological relations

The strategy has deep methodological ties to multiscale modeling and global-local coupling outside computer vision. In phase-field brittle fracture, a coarse global linear elastic problem is coupled to a fine local nonlinear phase-field fracture patch, which is dynamically enlarged as the crack approaches its boundary (Noii et al., 2019). The local enlargement criterion is phase-field based: expand if $$C = \frac{1}{k+1} \sum_{i=0}^{k} (p_i - \bar{p})(p_i - \bar{p})^T, \qquad \bar{p} = \frac{1}{k+1} \sum_{i=0}^{k} p_i,$$3 on the local interface $$C = \frac{1}{k+1} \sum_{i=0}^{k} (p_i - \bar{p})(p_i - \bar{p})^T, \qquad \bar{p} = \frac{1}{k+1} \sum_{i=0}^{k} p_i,$$4, with values such as $$C = \frac{1}{k+1} \sum_{i=0}^{k} (p_i - \bar{p})(p_i - \bar{p})^T, \qquad \bar{p} = \frac{1}{k+1} \sum_{i=0}^{k} p_i,$$5 used in examples (Noii et al., 2019). In adaptive ductile phase-field fracture, a global elastic-plastic model is augmented either with one local elastic-plastic-fracture zone or with two nested local domains, one elastic-plastic and one fracture-only, using indicators based on global plastic flow and local phase-field state (Aldakheel et al., 2021). These formulations make the global-to-local logic explicit: use a coarse global constitutive model, then activate richer local physics where the evolving state indicates that it is required.

A similar pattern appears in structural time integration. The global/local coupling strategy for viscoplastic structures starts from a coarse global analysis and allows local models to adopt finer adaptive time grids than the global model, effectively performing densification in both space and time (Blanchard et al., 2019). The preferred practical strategy is weak time coupling, where local time refinement remains local unless the global model itself requires smaller steps (Blanchard et al., 2019). This suggests that the “global-to-local” idea extends naturally from spatial densification to temporal discretization and to model hierarchy more broadly.

Across these domains, the central methodological relation is to selective model enrichment. Global-to-local densification differs from uniform refinement by assuming that only a subset of the domain, trajectory, sequence, or representation budget should receive fine-scale treatment. It differs from one-shot local modeling by preserving a global scaffold that supports consistency, stability, and long-range interaction. This suggests that the strategy occupies an intermediate position between fully global coarse models and fully local fine models: it is a mechanism for moving information and capacity between them.

8. Synthesis

Global-to-local densification strategy denotes a family of adaptive procedures in which broad, globally aggregated evidence establishes a coarse but stable representation, after which local descriptors, local rewards, or local operating conditions determine where refinement, splitting, cloning, or pruning should occur. In 3DGS, this may mean alternating between multi-view gradient densification and neighborhood Eigenentropy, with thresholded local splitting and pruning (Jäger et al., 6 Mar 2026). In accelerated Gaussian Splatting, it may mean using split first for global diffusion and clone later for local refinement, coordinated with coarse-to-fine image supervision (Huang et al., 27 Jul 2025). In feedforward dynamic reconstruction, it may mean translating scene-level rendering gains into local per-primitive utilities (Song et al., 28 Jun 2026). In dynamic 3DGS, it may mean aggregating temporal visibility globally over an interval, then making per-Gaussian local decisions with adaptive thresholds (Sandu et al., 22 Jun 2026). In panoramic depth estimation, it may mean global transformer context followed by local convolutional refinement through gated fusion (Bai et al., 2022).

The concept is therefore best understood as a structural principle rather than a single algorithm. Its recurring components are a global scaffold, a localization mechanism, and a schedule that controls when and how local refinement is activated. Its main promise is improved allocation of representational capacity: broad coverage or consistency is established before expensive fine detail is introduced. Its main difficulty is that the local criterion must be reliable enough not to suppress important structure or amplify unstable regions. The surveyed literature suggests that when these components are aligned, global-to-local densification can improve accuracy, compactness, and efficiency simultaneously; when they are not, it degenerates into premature local overfitting or redundant growth (Jäger et al., 6 Mar 2026, Huang et al., 27 Jul 2025, Fu et al., 4 Dec 2025).