Instant4D: 4D Gaussian Splatting in Minutes

TL;DR

Instant4D tackles dynamic 3D reconstruction from uncalibrated monocular video by decoupling geometry from photometry through deep visual SLAM-based initialization, grid pruning, and a compact 4D Gaussian representation. It introduces a motion-aware, isotropic 4D Gaussian formulation and a depth-consistent back-projection pipeline to efficiently model dynamic scenes, achieving minutes-scale reconstructions with substantial memory savings. The approach delivers competitive rendering quality on benchmarks like NVIDIA Dynamic Scene and DyCheck while dramatically reducing training time and memory compared to prior methods, and demonstrates generalization to in-the-wild videos. The combination of grid pruning and motion-aware Gaussians is shown to be a key driver of efficiency and robustness, with ablations confirming the importance of each component for temporal coherence and occlusion handling.

Abstract

Dynamic view synthesis has seen significant advances, yet reconstructing scenes from uncalibrated, casual video remains challenging due to slow optimization and complex parameter estimation. In this work, we present Instant4D, a monocular reconstruction system that leverages native 4D representation to efficiently process casual video sequences within minutes, without calibrated cameras or depth sensors. Our method begins with geometric recovery through deep visual SLAM, followed by grid pruning to optimize scene representation. Our design significantly reduces redundancy while maintaining geometric integrity, cutting model size to under 10% of its original footprint. To handle temporal dynamics efficiently, we introduce a streamlined 4D Gaussian representation, achieving a 30x speed-up and reducing training time to within two minutes, while maintaining competitive performance across several benchmarks. Our method reconstruct a single video within 10 minutes on the Dycheck dataset or for a typical 200-frame video. We further apply our model to in-the-wild videos, showcasing its generalizability. Our project website is published at https://instant4d.github.io/.

Figures (9)

• Figure 1: Part (a):Instant4D achieves better rendering performance with fewer training iterations against the original 4D Gaussian Splatting (4DGS) yang2023real. Part (b): Visualization on detailed dynamic object like a "spinning" apple. After 40-minute optimization, the rendering result of 4DGS remains blurry, while our method achieves better visual quality by 0.8 dB PSNR, faster optimization for convergence by 85%, and lower GPU memory by 69%. Part (c): Bubble chart comparing with most recent art. Note that the bubble size indicates the size of an optimized model.
• Figure 2: Pipeline of Instant4D. We use Deep Visual SLAM model and Unidepth piccinelli2024unidepth to obtain camera parameters, and metric depth. The metrics depth would be further optimized to consistent video depth. After that we back project from consistent depth to get dense point cloud, further voxel filtered to sparse point cloud, as discuss in Section \ref{['subsec_slam']}. Based on the 4d Gaussians Initialization, we can reconstruct a scene in 2 minutes. More details about optimization are described in Section \ref{['subsec_opt']}.
• Figure 3: Visual comparison on the NVIDIA dataset. yoon2020novel
• Figure 4: Visual Comparison on the Dycheck dataset.gao2022dynamic
• Figure 5: Visualization on the DAVIS Dataset.