View-Adaptive Dynamic Splatting

• View-adaptive dynamic splatting is a technique that adjusts Gaussian parameters based on view and time to enhance dynamic scene rendering.
• It employs target-view-conditioned residual refinement and multi-view consistent densification to optimize performance and visual quality.
• Several methods like ViewSplat and MoE-GS illustrate adaptive Gaussian updates that balance rendering fidelity with computational efficiency.

View-adaptive dynamic splatting is a family of Gaussian-splatting formulations in which the effective representation, optimization policy, or refinement behavior changes as a function of viewpoint, time, or both. In the strongest sense, the term denotes methods that make Gaussian attributes depend on the target view during rendering, rather than relying on a single rigid primitive set for all cameras. In broader usage, it also covers dynamic splatting pipelines that adapt densification, expert selection, or geometric supervision according to view-conditioned evidence during training or refinement. Recent work therefore uses the phrase across several distinct regimes: target-view-conditioned feed-forward rendering, viewpoint-targeted optimization for dynamic scenes, multi-view-consistent adaptive densification, and bandwidth- or geometry-adaptive decomposition of dynamic Gaussian representations (Jeong et al., 26 Mar 2026).

1. Terminological scope and historical framing

The most direct formulation of view-adaptive dynamic splatting appears in "ViewSplat" (Jeong et al., 26 Mar 2026). That work argues that prior feed-forward 3D Gaussian splatting pipelines regress one static set of Gaussian primitives that must explain all possible viewpoints simultaneously, and attributes a major fidelity bottleneck to that assumption. Its proposed alternative is a view-adaptable latent representation in which base Gaussian primitives are predicted once, while scene-conditioned dynamic MLPs take target-view coordinates as input and output residual updates for 3D position, scale, rotation, opacity, and color during rendering. In that formulation, the rendered primitive set itself depends on the requested camera (Jeong et al., 26 Mar 2026).

Other papers use the same words more narrowly. "4D Neural Voxel Splatting" (Wu et al., 1 Nov 2025) is described as view-adaptive mainly because it adds a dedicated view refinement stage that restricts training to underperforming viewpoints and lowers densification thresholds for those views, while its core representation remains a compact dynamic voxel scaffold with learned deformation. "MVG-Splatting" (Li et al., 2024) is relevant when view-adaptivity is interpreted as multi-view-consistent adaptive densification: it recomputes near/mid/far depth quantiles per rendered view and uses those view-specific statistics to decide where to project additional 2D Gaussian primitives during optimization. "MoE-GS" (Jin et al., 22 Oct 2025) adopts perhaps the broadest dynamic interpretation by blending multiple specialized dynamic Gaussian experts using a Volume-aware Pixel Router whose weights depend on splatted volumetric cues, view direction, and time.

This distribution of meanings makes the term partly polysemous. A common misconception is to treat any dynamic 3DGS method as view-adaptive merely because it renders arbitrary novel cameras. The literature distinguishes more sharply. Several dynamic splatting systems are time-adaptive or motion-adaptive but not explicitly viewpoint-conditioned in their representation. "A Compact Dynamic 3D Gaussian Representation for Real-Time Dynamic View Synthesis" (Katsumata et al., 2023) models positions and rotations as functions of time while keeping scale, color, and opacity invariant; "SWinGS" (Shaw et al., 2023) adapts temporal window size to motion complexity; and "Hybrid 3D-4D Gaussian Splatting" (Oh et al., 19 May 2025) adaptively converts temporally invariant 4D Gaussians into 3D Gaussians. These methods are dynamic, but not view-adaptive in the same sense as ViewSplat or MoE-GS.

This suggests that encyclopedia usage benefits from a layered definition. In the narrow sense, view-adaptive dynamic splatting refers to target-view-conditioned Gaussian parameter updates at rendering time. In a broader but still research-grounded sense, it also includes dynamic Gaussian systems whose training-time refinement, densification, routing, or geometry supervision is conditioned on viewpoint-specific evidence.

2. Core representational paradigms

A first paradigm is target-view-conditioned residual refinement over canonical Gaussians. ViewSplat predicts canonical/base 3D Gaussians and a scene-conditioned view-dependent head that generates pixel-wise dynamic MLPs through a hypernetwork architecture. Each view MLP receives a compact 4D target-view encoding composed of direction and log-distance to the target camera center and outputs residuals for all Gaussian attributes. The refined Gaussians are then rendered with a standard differentiable 3DGS rasterizer. The paper reports that this mechanism improves every SPFSplat backbone variant on RE10K and ACID, while preserving fast inference: 17 FPS and real-time rendering: 154 FPS (Jeong et al., 26 Mar 2026).

A second paradigm is view-conditioned dynamic scene modeling with compact canonical scaffolds. In 4D-NVS, the scene is represented by persistent neural voxels rather than per-frame Gaussian clouds. Visible voxels generate Gaussians on demand; a HexPlane-based 4D feature field predicts geometric deformation only, while color and opacity remain view-conditioned outputs of anchor-feature decoders. The paper’s explicitly view-adaptive component is a third-stage refinement that samples only crude viewpoints, weights sampling probability by severity, and lowers thresholds from $\tau_g = 0.0002$ to $0.0001$ and from $\tau_\alpha = 0.05$ to $0.03$ (Wu et al., 1 Nov 2025).

A third paradigm is view- and time-conditioned expert routing. MoE-GS combines heterogeneous dynamic Gaussian experts—such as MLP-based, polynomial, and interpolation-based dynamic splatting models—through a router that begins from per-Gaussian routing weights, splats them into image space, and computes per-pixel softmax gates. The final rendered image is

$I_{MoE} = \sum_{k=1}^{N} G'_k \cdot I_{E_k},$

with

$G'_k = \text{Softmax}(R'_k),$

so the contribution of each expert varies across pixels, viewpoints, and times (Jin et al., 22 Oct 2025). This is not a single Gaussian representation with one explicit view-conditioned deformation rule; it is a conditional composition of full dynamic renderers.

A fourth paradigm, weaker with respect to view-adaptation but still relevant, is view-conditioned motion transformation without a full 4D canonical field. DBMovi-GS predicts transformed Gaussian copies from Gaussian attributes and the camera view $p$,

$$\{(\delta \mu_{ji}, \delta r_{ji}, \delta s_{ji})\}_{i=1}^{M} = \mathcal{F}_\theta(\gamma(\mu_j), r_j, s_j, \gamma(p)),$$

to handle object motion blur, while separately estimating frame-to-frame camera motion with rigid $SE(3)$ transforms (Song et al., 26 Jun 2025). The view-aware component is therefore real, but narrower than the full target-view-conditioned residual update of ViewSplat.

A plausible implication is that current work separates into two structural camps: methods that change the effective primitive attributes for a requested target view, and methods that keep a largely fixed representation but adapt training or blending policies according to viewpoint difficulty or view-conditioned evidence.

3. Mechanisms of adaptation

The most explicit mechanism is per-view residual prediction. ViewSplat writes the method conceptually as

$\theta_i^{(t)} = \theta_i^{\text{base} + \Delta \theta_i(t),$

where $0.0001$0 denotes the attributes of Gaussian $0.0001$1 and $0.0001$2 the target view. The residuals cover position, rotation, scale, opacity, and color / SH coefficients. The paper’s ablation shows that omitting $0.0001$3 causes catastrophic failure, while the full joint update performs best. It also shows that merely increasing static SH degree does not reproduce the gains of view-adaptive residual refinement (Jeong et al., 26 Mar 2026).

A second mechanism is view-targeted training refinement. In 4D-NVS, difficult views are detected by PSNR or gradient criteria relative to an EMA, then pushed into a refinement stack with failure type, severity score, and temporal consistency flags. Refinement is a dedicated third stage of 14k iterations in which only crude viewpoints are sampled and density control becomes more aggressive. Reported quality gains include 25.8 PSNR without refinement vs 28.5 with refinement in one HyperNeRF table, and 28.61 without refinement vs 31.28 with full method on HyperNeRF-Chicken (Wu et al., 1 Nov 2025).

A third mechanism is volumetric, view-dependent router construction. MoE-GS defines per-Gaussian routing attributes

$0.0001$4

splats them into image-space features, and refines them as

$0.0001$5

where $0.0001$6 is the pixel ray direction. The router then computes per-pixel expert weights and blends expert renderings. An ablation on N3V reports 31.12 PSNR for a Pixel Router, 32.05 PSNR for a Volume Router, and 33.23 PSNR for the proposed Volume-aware Pixel Router (Jin et al., 22 Oct 2025).

A fourth mechanism is view-conditioned densification or geometric supervision during training. MVG-Splatting does not perform per-novel-view adaptive inference; rather, it recomputes view-specific depth quantiles and uses MVS-style geometric consistency masks to decide where new surfels should be inserted. The paper therefore describes the method more accurately as view-conditioned adaptive training rather than dynamic viewpoint-adaptive inference (Li et al., 2024). Similarly, VAD-GS in dynamic urban scenes selects different supporting views per unreliable region using a diversity-aware score, then reconstructs missing geometry via patch matching-based MVS; the adaptation is region- and visibility-specific, not renderer-specific (Zhang et al., 10 Oct 2025).

This diversity of mechanisms also clarifies a second misconception: view-adaptivity need not mean a single explicit viewpoint-conditioned MLP over Gaussian attributes. The literature also treats view-targeted optimization, per-pixel expert routing, and view-specific geometric refinement schedules as legitimate adaptive behaviors when those behaviors materially alter which primitives are refined, trusted, or blended.

4. Dynamic-scene formulations and efficiency trade-offs

Dynamic splatting introduces a second axis of adaptation: time. One major design question is whether to represent motion through per-frame Gaussian sets, continuous-time parameterizations, canonical-plus-deformation models, or native 4D Gaussians. The answer strongly affects how view-adaptive mechanisms can be integrated.

The compact continuous-time approach of "A Compact Dynamic 3D Gaussian Representation for Real-Time Dynamic View Synthesis" models positions by Fourier approximation and rotations by linear quaternion trends while keeping scale, color, and opacity invariant. The method reduces storage from $0.0001$7 to $0.0001$8 and reports 150 FPS on D-NeRF, 118 FPS on DyNeRF, and 188 FPS on HyperNeRF (Katsumata et al., 2023). However, it does not introduce viewpoint-conditioned deformation, per-view adaptive density control, or view-conditioned appearance correction beyond ordinary SH-based view dependence.

"SWinGS" instead uses one dynamic 3DGS model per sampled temporal window, with a motion-adaptive window schedule based on optical-flow magnitude and a cross-window consistency stage on sampled novel views. Its reported average on Neural 3D Video is 32.05 PSNR, 0.949 SSIM, 0.093 LPIPS, and 71.51 FPS (Shaw et al., 2023). This is dynamic and adaptive in time, but again not view-adaptive in the stronger representational sense.

A different route is hybrid structural adaptation between static and dynamic primitives. Hybrid 3D-4DGS starts from full 4D Gaussians and converts those with sufficiently large temporal scale into static 3D Gaussians using

$0.0001$9

On N3V, the paper reports 11m 53s training, 208 FPS, and 273 MB storage, compared with 5.5 hours, 114 FPS, and 2.1 GB for 4DGS (Oh et al., 19 May 2025). This is adaptive over space-time complexity rather than viewpoint.

A related systems-level extension is PD-4DGS, which decomposes a dynamic Gaussian pipeline into a static scaffold, global deformation, and local refinement so that any transmitted prefix is renderable. On Dycheck, it reports reducing first-frame latency from 73.5 s for MoDec-GS and 314.0 s for 4DGS to 1.74 s when only the 0.436 MB base layer is streamed at 2 Mbps (Li et al., 12 May 2026). This is bandwidth-adaptive rather than view-adaptive, but it indicates that dynamic splatting increasingly interacts with systems constraints beyond geometry and rendering fidelity.

This suggests that view-adaptive dynamic splatting is developing under multiple resource regimes. Some papers spend complexity on per-view residual refinement; some spend it on dynamic expert ensembles; others reduce temporal redundancy to make further adaptation computationally feasible.

5. Geometry, densification, and multi-view consistency

A substantial branch of the literature uses “adaptive” to mean where and when new Gaussians are created, often under multi-view geometric supervision. These methods are especially relevant because geometry errors often dominate view-dependent failures in dynamic or sparse-input settings.

MVG-Splatting addresses geometry problems in 2D Gaussian surfels by refining depth and normals, then performing adaptive quantile-based densification. Depth maps are segmented into near, mid, and far regions by per-view quantiles, and geometric consistency tolerances are adjusted asymmetrically so that under-reconstructed near and far regions receive more projected points. The supplementary schedule runs 15k iterations of original adaptive densification, 5k optimization iterations, then 10k depth projection densification iterations with densification every 100 iterations (Li et al., 2024). On Mip-NeRF 360 it reports 27.84 PSNR, 0.834 SSIM, and 0.196 LPIPS.

Temporally Aware Densification for Dynamic 3DGS pushes this logic into the time domain. Its Visibility-Aware Densification replaces fixed-interval gradient averaging with

$\tau_\alpha = 0.05$0

so short-lived dynamic Gaussians are judged by actual temporal visibility rather than being diluted by frames in which they are absent. Temporally-Adaptive Thresholding and Temporal Offset Warping further adapt densification to temporal lifespan and motion complexity. The full model reports 32.42 / 24.68 / 0.863 / 0.059 on N3DV, 34.14 / 28.87 / 0.901 / 0.044 on Interdigital, and 28.28 / 25.39 / 0.881 / 0.090 on VRU Basketball (Sandu et al., 22 Jun 2026). This is not explicitly view-adaptive, but it resolves dynamic reconstruction failures that would otherwise degrade novel-view quality.

GC-4DGS addresses sparse-input dynamic view synthesis by injecting geometry consistency into 4DGS. It filters MVS depth using dynamic consistency checking across views and frames, supervises rendered depth with a masked structure loss, and adds global ordinal plus local normalized monocular depth regularization. In the 3-view setting on N3DV, it reports 27.69 PSNR, 0.907 SSIM, 0.074 LPIPS, 0.034 AVGE, and 190 FPS, outperforming RF-DeRF by 2.62 dB and 4DGS by 1.58 dB (Li et al., 28 Nov 2025).

VAD-GS is another geometry-recovery variant oriented to dynamic urban scenes. It identifies unreliable geometry through voxel-based visibility reasoning, selects informative supporting views via a diversity-aware score, and reconstructs missing structure through patch matching-based MVS. On Waymo it reports 35.59 PSNR, 31.31 PSNR* on dynamic objects, 0.950 SSIM, and 0.047 LPIPS; on nuScenes it reports scene-level gains over StreetGaussians and notes especially strong benefits when initialization is sparse (Zhang et al., 10 Oct 2025). Here “view-adaptive” means adaptive support-view selection for densification, not target-view-conditioned rendering.

A plausible implication is that view-adaptive dynamic splatting increasingly depends on geometry-aware support policies. When the representation itself is explicit and rasterized, misallocated density or incomplete structure cannot be fully hidden by view-dependent appearance modeling. Consequently, many “adaptive” methods first solve which regions or views deserve more geometry, and only then solve how appearance should vary with view.

6. Empirical patterns, limitations, and open interpretations

Across papers, several empirical patterns recur. First, target-view-conditioned refinement helps most when static feed-forward prediction is the bottleneck. ViewSplat improves SPFSplat from 25.484 / 0.847 / 0.153 to 26.317 / 0.857 / 0.144 on RE10K, and SPFSplatV2-L from 25.668 / 0.855 / 0.137 to 26.798 / 0.870 / 0.124 (Jeong et al., 26 Mar 2026). Second, view-targeted dynamic refinement can rescue hard viewpoints without global over-densification. 4D-NVS reports 28.5 PSNR, 0.872 MS-SSIM, 13 min, 44 FPS, and 3050 MiB on HyperNeRF, and attributes a substantial portion of the gain to the difficult-view refinement stage (Wu et al., 1 Nov 2025). Third, router-based mixtures outperform any single dynamic expert when scene behavior is heterogeneous. MoE-GS improves N3V average PSNR from 32.33 for E-D3DGS and 32.10 for Ex4DGS to 33.23 for MoE-GS with $\tau_\alpha = 0.05$1, while pruning can recover much of the speed loss (Jin et al., 22 Oct 2025).

The literature also makes clear what view-adaptive dynamic splatting is not. It is not automatically equivalent to static view-dependent appearance, ordinary novel-view rendering, or generic time-varying deformation. The static 3D-4D hybrid model (Oh et al., 19 May 2025), the compact dynamic Fourier representation (Katsumata et al., 2023), and motion-adaptive windowed training (Shaw et al., 2023) all improve dynamic splatting without making primitive attributes depend on the queried target view. Conversely, Camera Splatting optimizes view placement using camera splats and point cameras, but it remains a static-scene view-optimization method rather than a dynamic Gaussian scene representation (Lee et al., 19 Sep 2025).

Several limitations recur across the surveyed methods. A fully view-adaptive representation can be expensive in memory or runtime, as seen in the rendering slowdown from baseline SPFSplat to ViewSplat’s dynamic refinement (Jeong et al., 26 Mar 2026). Mixture-of-experts routing improves quality but introduces substantial overhead before pruning or distillation (Jin et al., 22 Oct 2025). Dynamic urban densification methods often rely on rigid or piecewise-planar assumptions in their MVS subroutines (Zhang et al., 10 Oct 2025). Blur-aware dynamic Gaussian systems such as DBMovi-GS are only weakly view-adaptive and provide limited evidence specifically isolating the value of view conditioning (Song et al., 26 Jun 2025). Even methods that are explicitly “view-adaptive” at training time, such as MVG-Splatting, may not be adaptive at inference time in the sense of altering the representation online (Li et al., 2024).

The field therefore remains internally differentiated. One line pursues view-conditioned rendering-time corrections; another pursues view-aware optimization schedules or expert selection; another pursues geometry-aware adaptive densification; and another pursues resource-adaptive decomposition that could support future viewpoint-aware delivery. Taken together, these works suggest that “view-adaptive dynamic splatting” is less a single architecture class than a converging research agenda: explicit Gaussian scene representations are being made progressively more conditional on where the camera is, when the scene is observed, and which parts of the representation are actually needed (Jeong et al., 26 Mar 2026).