Research Milestone

4D Gaussian Splatting for Real-Time Dynamic Scene Rendering

Models dynamic scenes with 4D neural voxels and a deformation field over a single canonical Gaussian set for real-time training and storage efficiency.

Authors / Team

Guanjun Wu · Researcher

Year

2024

Deep Dive

The method maintains one set of canonical 3D Gaussians and predicts their deformation over time using features built from 4D neural voxels and a compact MLP, avoiding per-frame Gaussian duplication. Rendering stays in the differentiable splatting regime, with reported real-time performance on dynamic benchmarks. The project page and code release support reproduction and comparison with static 3DGS pipelines.

What we learn

01
The key for dynamic Gaussians is a compact temporal encoding instead of per-frame point blow-up.
02
Canonical space plus a deformation field remains a practical bridge between dynamic NeRFs and explicit splatting.

Verbatim quote

"Representing and rendering dynamic scenes has been an important but challenging task."— source ↗