MVFusion-GS: Motion-Variance Guided Temporal Attention for High-Quality Dynamic Gaussian Splatting

2026-07-02 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Expert, quick

Summary

MVFusion-GS, a novel method, enhances 3D Gaussian Splatting (3DGS) for high-quality dynamic scene reconstruction by addressing limitations in existing deformation networks. Current approaches often fail to capture long-term motion intensity or short-term temporal coherence, resulting in inaccurate foreground deformation and background artifacts. MVFusion-GS integrates two motion-aware mechanisms: Motion-Variance Guided Refinement, which aggregates per-Gaussian deformation statistics to estimate motion variance for dynamic-static separation, and MotionFormer Temporal Attention, which employs Transformer self-attention across neighboring timesteps to model local motion dependencies and improve temporal consistency. This approach achieves leading performance on both dynamic scene reconstruction and distractor-free reconstruction benchmarks, demonstrating improved foreground motion modeling and static background reconstruction.

Key takeaway

For Computer Vision Engineers developing real-time novel view synthesis for dynamic scenes, MVFusion-GS offers a significant advancement. You should consider integrating explicit motion awareness into your deformation networks, specifically by utilizing motion variance for dynamic-static separation and temporal attention for consistency. This approach can resolve issues like inaccurate foreground deformation and pseudo-static background residuals, leading to higher quality dynamic 3DGS outputs.

Key insights

MVFusion-GS improves dynamic 3D Gaussian Splatting by explicitly integrating motion awareness through variance-guided refinement and temporal attention.

Principles

• Explicit motion awareness improves dynamic scene reconstruction.
• Aggregating deformation statistics aids dynamic-static separation.
• Transformer self-attention enhances temporal consistency in motion.

Method

MVFusion-GS enhances deformation networks with Motion-Variance Guided Refinement for dynamic-static separation and MotionFormer Temporal Attention using Transformer self-attention for local motion dependencies.

In practice

• Reconstruct dynamic scenes with higher fidelity.
• Improve distractor-free reconstruction quality.
• Enhance foreground motion modeling accuracy.

Topics

• 3D Gaussian Splatting
• Dynamic Scene Reconstruction
• Novel View Synthesis
• Temporal Attention
• Motion Variance
• Computer Vision

Best for: Research Scientist, AI Scientist, Computer Vision Engineer, Machine Learning Engineer

Related on AIssential

• GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens — GlobalSplat uses a global latent scene representation for efficient, compact, and consistent 3D Gaussian Splatting.
• TurboGS: Accelerating 3D Gaussian Splatting via Error-Guided Sparse Pixel Sampling and Optimization — TurboGS accelerates 3DGS by focusing optimization on perceptually informative pixels using error-guided sampling and dynamic density control.
• MoRe: Motion-aware Feed-forward 4D Reconstruction Transformer — MoRe efficiently reconstructs dynamic 4D scenes from monocular video by disentangling motion from static structure.
• Deep Learning Approaches for 3D Medical Scene Completion: From Geometric Modeling to Generative Paradigms — The field of 3D medical scene completion has rapidly evolved from voxel-based methods to generative diffusion and Gaussian splatting.
• Fast and Lightweight Novel View Synthesis with Differentiable Multiplane Image — MPI, enhanced with diffusion and point maps, offers a fast, lightweight alternative for novel view synthesis.
• PanoPlane: Plane-Aware Panoramic Completion for Sparse-View Indoor 3D Gaussian Splatting — Panoramic scene completion with layout-anchored attention steering significantly improves sparse-view 3D reconstruction.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.