4C4D: 4 Camera 4D Gaussian Splatting

Junsheng Zhou; Zhifan Yang; Liang Han; Wenyuan Zhang; Kanle Shi; Shenkun Xu; Yu-Shen Liu

cs.CVApr 5, 2026

4C4D: 4 Camera 4D Gaussian Splatting

Authors: Junsheng Zhou, Zhifan Yang, Liang Han, Wenyuan Zhang, Kanle Shi, Shenkun Xu, Yu-Shen Liu

Organizations: School of Software, Tsinghua University · Kuaishou Technology

Abstract

This paper tackles the challenge of recovering 4D dynamic scenes from videos captured by as few as four portable cameras. Learning to model scene dynamics for temporally consistent novel-view rendering is a foundational task in computer graphics, where previous works often require dense multi-view captures using camera arrays of dozens or even hundreds of views. We propose 4C4D, a novel framework that enables high-fidelity 4D Gaussian Splatting from video captures of extremely sparse cameras. Our key insight lies that the geometric learning under sparse settings is substantially more difficult than modeling appearance. Driven by this observation, we introduce a Neural Decaying Function on Gaussian opacities for enhancing the geometric modeling capability of 4D Gaussians. This design mitigates the inherent imbalance between geometry and appearance modeling in 4DGS by encouraging the 4DGS gradients to focus more on geometric learning. Extensive experiments across sparse-view datasets with varying camera overlaps show that 4C4D achieves superior performance over prior art. Project page at: https://junshengzhou.github.io/4C4D.

Figures & tables

Figure 2: Overview of 4C4D. We introduce a Neural Decaying Function $f_{\theta}$ , implemented as a lightweight neural network, to adaptively control the opacity decay of Gaussians. Given key Gaussian attributes as input, $f_{\theta}$ predicts a factor that controls the decay of Gaussian opacities. During training, both the Neural Decaying Function and the 4D Gaussians are jointly optimized via gradient backpropagation under a photometric rendering loss.

Figure 3: Rendered RGB and depth results from 4DGS trained with and without our Neural Decaying Function, evaluated on both training and unseen views. Without neural decaying, 4DGS can overfit the training views but fails to generalize to unseen viewpoints, due to the learned poor geometry.

Figure 4: Visual comparisons under Neural3DV dataset.

Figure 5: Visual comparisons under ENeRF-Outdoor dataset.

Method	PSNR $\uparrow$	DSSIM 1 $\downarrow$	DSSIM 2 $\downarrow$	LPIPS $\downarrow$
4DGaussians	20.82	0.117	0.077	0.190
STGS	17.70	0.158	0.107	0.325
4DGS	20.60	0.143	0.094	0.244
Ex4DGS	19.33	0.149	0.099	0.239
Ours	22.29	0.098	0.062	0.146

Table 1: Quantitative comparison under Neural3DV dataset.

Figure 6: Visual comparisons under Mobile-Stage dataset.

Method	PSNR $\uparrow$	DSSIM 1 $\downarrow$	DSSIM 2 $\downarrow$	LPIPS $\downarrow$
4DGaussians	18.21	0.300	0.183	0.456
STGS	15.51	0.315	0.186	0.688
4DGS	23.52	0.183	0.112	0.151
Ex4DGS	21.89	0.224	0.138	0.263
Ours	24.32	0.162	0.097	0.121

Table 2: Quantitative comparisons under ENeRF-Outdoor dataset.

Method	PSNR $\uparrow$	DSSIM 1 $\downarrow$	DSSIM 2 $\downarrow$	LPIPS $\downarrow$
4DGaussians	20.15	0.218	0.144	0.226
STGS	14.30	0.321	0.229	0.658
4DGS	22.15	0.199	0.135	0.180
Ex4DGS	17.85	0.272	0.193	0.260
Ours	22.36	0.194	0.129	0.121

Table 3: Quantitative comparison under Mobile-Stage dataset.

Figure 7: Visual comparisons under the self-captured Dyn4Cam dataset.

Method	PSNR $\uparrow$	DSSIM 1 $\downarrow$	DSSIM 2 $\downarrow$	LPIPS $\downarrow$
w/o Neural Decay	22.60	0.097	0.062	0.147
w/o Vis Detect	24.49	0.075	0.045	0.127
Full	24.68	0.070	0.042	0.115

Table 4: Ablation study on the framework designs.

Method	PSNR $\uparrow$	DSSIM 1 $\downarrow$	DSSIM 2 $\downarrow$	LPIPS $\downarrow$
None	22.60	0.097	0.062	0.147
Constant	24.31	0.075	0.046	0.125
Pow Decay	24.35	0.074	0.045	0.124
Exp Decay	24.32	0.077	0.047	0.135
Ours	24.68	0.070	0.042	0.115

Table 5: Ablation study on decaying functions.

Figure 8: More visual comparisons under Neural3DV dataset.

	Coffee Martini				Cook Spinach				Cut Roasted Beef
Method	PSNR	DSSIM 1	DSSIM 2	LPIPS	PSNR	DSSIM 1	DSSIM 2	LPIPS	PSNR	DSSIM 1	DSSIM 2	LPIPS
STGS [ 20 ]	16.62	0.189	0.129	0.373	18.72	0.134	0.090	0.278	18.65	0.135	0.089	0.286
4DGS [ 47 ]	17.50	0.174	0.118	0.289	21.96	0.106	0.067	0.205	21.90	0.102	0.064	0.191
Ex4DGS [ 17 ]	17.45	0.179	0.124	0.281	20.61	0.138	0.089	0.224	20.33	0.144	0.092	0.228
Ours	19.57	0.140	0.093	0.203	24.44	0.075	0.046	0.111	23.22	0.083	0.051	0.124

Table 6: Quantitative comparison on Neural3DV dataset (Part 1).

	Flame Salmon				Flame Steak				Sear Steak
Method	PSNR	DSSIM 1	DSSIM 2	LPIPS	PSNR	DSSIM 1	DSSIM 2	LPIPS	PSNR	DSSIM 1	DSSIM 2	LPIPS
STGS [ 20 ]	14.54	0.227	0.162	0.474	18.96	0.131	0.086	0.270	18.72	0.136	0.090	0.271
4DGS [ 47 ]	17.81	0.163	0.110	0.283	22.55	0.212	0.146	0.294	21.86	0.102	0.064	0.202
Ex4DGS [ 17 ]	15.40	0.193	0.135	0.308	21.39	0.123	0.077	0.200	20.80	0.119	0.078	0.196
Ours	19.15	0.136	0.091	0.205	23.72	0.075	0.046	0.114	23.66	0.078	0.048	0.118

Table 7: Quantitative comparison on Neural3DV dataset (Part 2).

	Actor1_4				Actor2_3				Actor5_6
Method	PSNR	DSSIM 1	DSSIM 2	LPIPS	PSNR	DSSIM 1	DSSIM 2	LPIPS	PSNR	DSSIM 1	DSSIM 2	LPIPS
STGS [ 20 ]	15.50	0.314	0.186	0.690	15.55	0.313	0.184	0.686	15.44	0.316	0.187	0.682
4DGS [ 47 ]	23.24	0.194	0.119	0.154	23.64	0.186	0.114	0.154	23.71	0.145	0.087	0.139
Ex4DGS [ 17 ]	20.98	0.229	0.142	0.269	22.02	0.236	0.145	0.283	22.68	0.208	0.125	0.235
Ours	24.07	0.182	0.109	0.129	24.39	0.175	0.104	0.124	24.50	0.130	0.077	0.111

Table 8: Quantitative comparison on ENeRF-Outdoor dataset.

Figure 9: More visual comparisons under ENeRF-Outdoor dataset.

Figure 10: More visual comparisons under Dyn4Cam dataset (Part 1).

Figure 11: More visual comparisons under Dyn4Cam dataset (Part 2).

Sparse-View 4D Reconstruction

See all 13 papers

D^2-4DGS: Dual-Depth Guided Sparse-Camera 4D Gaussian Splatting
Aug 3, 2026Jijian Zhao3D Gaussian Splatting Sparse-View 4D Reconstruction
Huazhong University of Science and Technology
4DGS-Fixer: Generative Sparse-View 4D Gaussian Splatting with Iterative Refinement Guided by Video Diffusion Priors
Sep 18, 2026Haitao Huang, Shenghao Zhao, Boyuan Tian +73D Gaussian Splatting Sparse-View 4D Reconstruction
Goertek Alpha Labs Shanghai, China · Singapore Institute of Technology Singapore, Singapore · Goertek Alpha Labs Santa Clara, USA +1
Ground4D: Consistency-Aware 4D Reconstruction from Monocular Video
Jun 27, 2026Qing Zhao, Weijian Deng, Pengxu Wei +1Multi-View Consistency 3D Gaussian Splatting
School of Computer Science and Engineering, Sun Yat-sen University, China · Tsinghua Shenzhen International Graduate School, Tsinghua University, China

Aug 3, 2026J/K move · Enter open · S save

D^2-4DGS: Dual-Depth Guided Sparse-Camera 4D Gaussian Splatting

Jijian Zhao

Huazhong University of Science and Technology

3D Gaussian Splatting Sparse-View 4D Reconstruction Dynamic Scene Reconstruction Dynamic Novel View Synthesis

Open paper AskPDF arXiv

Dynamic 4D Gaussian Splatting has emerged as an efficient representation for dynamic novel view synthesis through explicit scene modeling and real-time rendering. However, existing methods typically require dense multi-view videos for sufficient geometric constraints, making capture expensive and limiting sparse-camera deployment. Reducing input views lowers acquisition cost but weakens geometry supervision, often causing missing structures and floating Gaussians. Depth priors provide geometric cues, yet no single source offers both dense coverage and reliable geometry. Monocular depth provides dense structure but is scale-ambiguous and locally biased, whereas multi-view geometric depth provides incomplete anchors consistent with the reconstruction coordinate system. To exploit their complementarity, we propose D

^2

-4DGS, a sparse-camera dynamic 4D Gaussian Splatting framework guided by dual-source depth priors. We align monocular estimates with valid multi-view geometric depths and verify their consistency to identify reliable geometric anchors. These verified anchors support consistency-aware pruning and depth supervision, while verified geometric depths and aligned mono-only estimates provide candidate geometry for densification in under-reconstructed regions. Finally, RGB-D joint optimization improves appearance fidelity and geometric consistency under sparse-view supervision. Across all nine dataset--view settings, D

^2

-4DGS achieves the highest PSNR, improving by 1.33 dB on average over the best competing method in each setting.

Abstract

Sparse-View 4D Reconstruction

Explore similar work

D^2-4DGS: Dual-Depth Guided Sparse-Camera 4D Gaussian Splatting

4DGS-Fixer: Generative Sparse-View 4D Gaussian Splatting with Iterative Refinement Guided by Video Diffusion Priors

Ground4D: Consistency-Aware 4D Reconstruction from Monocular Video