MoE-GS: Mixture of Experts for Dynamic Gaussian Splatting
In-Hwan Jin, Hyeongju Mun, Joonsoo Kim, Kugjin Yun, Kyeongbo Kong
TL;DR
MoE-GS introduces a novel mixture of experts framework for dynamic Gaussian splatting, enhancing 3D scene reconstruction and view synthesis by adaptively blending specialized experts for improved quality across diverse scenes.
Contribution
This work is the first to incorporate Mixture-of-Experts techniques into dynamic Gaussian splatting, addressing performance inconsistencies across scenes.
Findings
Outperforms state-of-the-art methods on N3V and Technicolor datasets.
Achieves higher rendering quality with adaptive expert blending.
Provides strategies for efficient deployment and lightweight models.
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.…
Peer Reviews
Decision·ICLR 2026 Poster
Very good results supported by comprehensive quantitative experiments (more in the appendix!), images, and an easy-to-read demo webpage. MoE-GS runs with higher speeds and lower memory requirements (with pruning enabled) compared to only using one of the methods the experts are based on [Table 3], while offering state-of-the-art image quality. Good applicability/adaptability. Theoretically one can use any dynamic GS method as an "expert" so long as it returns a set of 3D Gaussians at each time
Minor weakness: unlike in MoE LLMs where most experts are not executed, the MoE mode here (i.e. without distillation) still runs all experts at all times. The inference speed has potential to be improved further if, we can combine experts with sparsity. Is there a reason this is not done in this paper?
1. The paper provides a solid diagnostic study (Fig. 1) showing how different dynamic Gaussian Splatting models reach performance peaks at different scenes, spatial regions, and time steps. This empirical observation is valuable to the community and highlights an important open problem of model generalization in dynamic reconstruction. 2. Even if the current MoE-GS realization focuses on image-level blending, the paper raises the broader idea of combining specialized dynamic reconstruction mode
1. Has conceptual mismatch between its stated motivation and its technical realization. In 3D reconstruction, **novel view synthesis is merely a means to validate that the reconstructed geometry is correct, not the ultimate goal itself**. However, MoE-GS operates entirely at the image level, where the proposed Volume-aware Pixel Router blends rendered 2D outputs from multiple experts using per-pixel softmax weighting. Each expert maintains an independent 3D Gaussian representation, and **the rou
MoE-GS is the first work to introduce a Mixture-of-Experts architecture to dynamic Gaussian splatting, offering a new solution to the problem that a single model cannot handle scene-specific diversity. The proposed MoE-GS model consistently achieves higher rendering quality than existing SOTA methods on complex dynamic scenes and demonstrated its superiority across diverse datasets (N3V, Technicolor). The authors acknowledge the drawbacks of the MoE architecture, namely increased computa
Major weaknesses are as below: In LLMs, MoE is typically used at inference time to reduce the number of activated parameters [1]. By contrast, this paper uses MoE as a mechanism for mixing multiple models. Nevertheless, it suggests that as the number of models increases, computational resource demands grow, while the performance gains are unlikely (at a fine-grained level, e.g. a patch of novel view image) to surpass the best performance of any single model. [1] Dai et al. “DeepSeekMoE: Towa
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsComputer Graphics and Visualization Techniques · Advanced Vision and Imaging · 3D Shape Modeling and Analysis