Feature Consistent 4D Gaussian Splatting for Realistic Dynamic View Synthesis

ICLR 2026 Conference Submission21991 Authors

Keywords: Novel View Synthesis, Dynamic Scene, Gaussian Splatting

Abstract: Dynamic novel view synthesis remains challenging due to the complexity of diverse motion patterns. In 4D Gaussians, the temporal dimension further complicates constraint formulation, making temporally consistent rendering difficult. To address this, we introduce 4D Feature Gaussian Splatting (F4DGS), a dynamic rendering algorithm that introduces feature consistency regularization to enable realistic rendering. This regularization jointly synchronizes hierarchical semantic features, velocity, and depth, ensuring coherent motion and appearance. We further extend the regularization beyond static alignment to capture temporal associations over continuous unit time intervals. F4DGS is the first rendering algorithm to explicitly couple velocity and depth for learning motion-consistent 4D representations, enabling high-fidelity, physically plausible rendering of dynamic content. Through comprehensive evaluations on monocular and multi-view dynamic datasets, F4DGS achieves real-time, high-resolution rendering and consistently outperforms existing methods across both quantitative and qualitative benchmarks. Notably, F4DGS achieves a 3.51 PSNR improvement on the Plenoptic dataset with comparable rendering speed and training time.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 21991