OpenNovelty.org

MoE-GS: Mixture of Experts for Dynamic Gaussian Splatting

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.0Download Report PDF
Gaussian SplattingDynamic Scene Reconstruction
Abstract:

Recent advances in dynamic scene reconstruction have significantly benefited from 3D Gaussian Splatting, yet existing methods show inconsistent performance across diverse scenes, indicating no single approach effectively handles all dynamic challenges. To overcome these limitations, we propose Mixture of Experts for Dynamic Gaussian Splatting (MoE-GS), a unified framework integrating multiple specialized experts via a novel Volume-aware Pixel Router. Unlike sparsity-oriented MoE architectures in large language models, MoE-GS is designed to improve dynamic novel view synthesis quality by combining heterogeneous deformation priors, rather than to reduce training or inference-time FLOPs. Our router adaptively blends expert outputs by projecting volumetric Gaussian-level weights into pixel space through differentiable weight splatting, ensuring spatially and temporally coherent results. Although MoE-GS improves rendering quality, the increased model capacity and reduced FPS are inherent to the MoE architecture. To mitigate this, we explore two complementary directions: (1) single-pass multi-expert rendering and gate-aware Gaussian pruning, which improve efficiency within the MoE framework, and (2) a distillation strategy that transfers MoE performance to individual experts, enabling lightweight deployment without architectural changes. To the best of our knowledge, MoE-GS is the first approach incorporating Mixture-of-Experts techniques into dynamic Gaussian splatting. Extensive experiments on the N3V and Technicolor datasets demonstrate that MoE-GS consistently outperforms state-of-the-art methods with improved efficiency. Video demonstrations are available at https://huggingface.co/spaces/moegs/MoE-GS.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper proposes MoE-GS, a mixture-of-experts framework for dynamic Gaussian splatting that combines multiple deformation priors via a volume-aware pixel router. It resides in the 'Canonical Space with Deformation Fields' leaf, which contains six papers including Deform3DGS and DESR. This leaf represents a moderately populated research direction within the broader deformation modeling branch, indicating established interest in canonical-space approaches but not extreme saturation. The taxonomy shows five sibling papers in this exact category, suggesting a competitive yet not overcrowded niche.

The taxonomy reveals neighboring leaves addressing alternative motion representations: 'Spacetime and 4D Representations' (five papers) treats time as a unified dimension, while 'Trajectory and Motion Curve Modeling' (two papers) uses explicit motion paths. The 'Sparse Control and Decomposed Motion' leaf (two papers) shares MoE-GS's goal of handling heterogeneous dynamics but through control-point decomposition rather than expert routing. The 'Scene Decomposition and Disentanglement' branch (five papers total) tackles static-dynamic separation, a complementary problem that MoE-GS addresses implicitly through adaptive expert blending rather than explicit segmentation.

Among sixteen candidates examined across three contributions, none clearly refute the core claims. The MoE-GS framework contribution examined six candidates with zero refutations, suggesting limited direct overlap in the mixture-of-experts approach for Gaussian splatting. The volume-aware router received no candidate examination, indicating potential novelty or insufficient semantic matches in the search. Efficiency improvements examined ten candidates without refutations, though this may reflect the limited search scope rather than absolute novelty. The statistics suggest the core MoE architecture and routing mechanism appear less explored in prior work within the examined candidate set.

Based on the limited search of sixteen semantically similar papers, the work appears to introduce a relatively fresh angle within canonical-space deformation methods. The taxonomy context shows this is an active but not saturated area, and the absence of refutations among examined candidates suggests the expert-routing approach is underexplored. However, the search scope remains narrow, and a broader literature review might reveal closer precedents in computer vision or neural rendering beyond the Gaussian splatting domain.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
16
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: dynamic scene reconstruction using Gaussian splatting. The field has organized itself around several complementary research directions. Deformation Modeling and Motion Representation focuses on how to encode temporal changes, often through canonical spaces with learned deformation fields or direct spacetime parameterizations, as seen in works like 4D Gaussian Splatting[3] and Deformable 3D Gaussians[24]. Scene Decomposition and Disentanglement tackles separating static backgrounds from dynamic objects, a critical challenge for urban and multi-object scenarios exemplified by Street Gaussians[1] and DrivingGaussian[11]. Domain-Specific Dynamic Reconstruction addresses specialized settings such as endoscopic procedures (Deformable Endoscopic[6], Endo-4DGS[8]) or thermal imaging (ThermalGS[2]). Meanwhile, Rendering Quality and Temporal Consistency emphasizes smooth frame-to-frame coherence, Efficiency and Scalability pursues real-time performance and memory reduction, and Specialized Reconstruction Techniques explores novel representations like mesh-adsorbed Gaussians or Bézier curves. A particularly active line of work centers on canonical-space deformation models, which establish a reference frame and warp Gaussians over time. MoE-GS[0] sits squarely within this branch, employing a mixture-of-experts strategy to handle complex, heterogeneous motions more effectively than single-field approaches like Deform3DGS[13] or DESR[14]. This contrasts with methods that directly parameterize spacetime coordinates without an explicit canonical frame, trading off interpretability for potentially simpler optimization. Another vibrant area involves disentangling static and dynamic components, where MoE-GS[0] shares motivations with SC-GS[9] and HUGS[16], yet differs by leveraging expert gating to adaptively route different motion patterns rather than relying on uniform segmentation. Open questions remain around balancing model capacity with generalization, especially when scenes exhibit rare or abrupt motion modes that challenge both canonical-field and direct spacetime representations.

Claimed Contributions

MoE-GS framework for dynamic Gaussian splatting

The authors introduce MoE-GS, the first framework that applies Mixture-of-Experts techniques to dynamic Gaussian splatting. This unified framework integrates multiple specialized expert models to adaptively handle diverse dynamic scene reconstruction challenges by combining heterogeneous deformation priors rather than reducing computational cost.

6 retrieved papers
Volume-aware Pixel Router

The authors propose a novel routing mechanism that projects Gaussian-level gating weights into pixel space using differentiable weight splatting. This router combines volumetric awareness with pixel-level blending to ensure spatially and temporally coherent expert integration while maintaining stable optimization.

0 retrieved papers
Efficiency improvements and distillation strategy

The authors develop two complementary approaches to address computational overhead: (1) single-pass multi-expert rendering and gate-aware Gaussian pruning to improve runtime efficiency within the MoE framework, and (2) a knowledge distillation strategy that transfers MoE performance to individual experts for lightweight deployment without architectural changes.

10 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

MoE-GS framework for dynamic Gaussian splatting

The authors introduce MoE-GS, the first framework that applies Mixture-of-Experts techniques to dynamic Gaussian splatting. This unified framework integrates multiple specialized expert models to adaptively handle diverse dynamic scene reconstruction challenges by combining heterogeneous deformation priors rather than reducing computational cost.

Contribution

Volume-aware Pixel Router

The authors propose a novel routing mechanism that projects Gaussian-level gating weights into pixel space using differentiable weight splatting. This router combines volumetric awareness with pixel-level blending to ensure spatially and temporally coherent expert integration while maintaining stable optimization.

Contribution

Efficiency improvements and distillation strategy

The authors develop two complementary approaches to address computational overhead: (1) single-pass multi-expert rendering and gate-aware Gaussian pruning to improve runtime efficiency within the MoE framework, and (2) a knowledge distillation strategy that transfers MoE performance to individual experts for lightweight deployment without architectural changes.

Table of Contents