CVPR 2026 3D Vision Gaussian Splatting Multi-Period Scene Neural Rendering Temporal Modeling

ChronoGS: Disentangling Invariants and Changes in Multi-Period Scenes

Zhongtao Wang1   Jiaqi Dai2   Qingtian Zhu3   Yilong Li1   Mai Su1   Fei Zhu1*   Meng Gai1   Shaorong Wang2   Chengwei Pan4   Yisong Chen1*   Guoping Wang1

1 Peking University
2 Beijing Forestry University
3 The University of Tokyo
4 Beihang University

PDF Code Dataset

TL;DRThe first unified differentiable Gaussian-splatting framework for reconstructing multi-period scenes, jointly handling non-continuous geometry and appearance changes, accompanied by the ChronoScene benchmark dataset.

Abstract

ChronoGS introduces a temporally modulated Gaussian representation for reconstructing multi-period scenes. It aggregates all periods within a unified anchor scaffold of union geometry, and disentangles stable and evolving components via temporally modulated features (period-invariant base, local period-varying, and global period-varying) together with a geometry activation mechanism.

Method

Existing static methods assume a single time-invariant geometry shared by all views, while dynamic methods rely on smooth and continuous motion over time.

Factorize a multi-period scene into a shared canonical geometry and period-specific variations. A unified anchor scaffold encodes the union geometry across all periods; each anchor maintains a period-invariant base feature and local/global period-varying features, which are decoded into Gaussian attributes by a lightweight MLP. A geometry activation mechanism automatically deactivates anchors whose predicted opacity falls below zero at a given period, enabling adaptive geometry expression while preserving global structural consistency.

Anchor Scaffold

Anchor Scaffold

A voxelized uniform grid initialized from merged sparse point clouds of all periods, covering both stable and varying geometry as a consistent backbone.

Base Feature

Base Feature

Period-invariant base feature f_i^base that describes shared geometry and appearance across periods.

Local Var. Feature

Local Var. Feature

Local period-varying feature pool f_i^var storing period-specific information per anchor.

Global Var. Feature

Global Var. Feature

Global period-varying feature g(t) modeling scene-level factors such as illumination and seasonal appearance.

Temporal Encoding

Temporal Encoding

One-hot encoding over period indices with linear interpolation for inter-period positions, preserving exact representations for observed periods.

Geometry Activation

Geometry Activation

Automatically excludes Gaussian primitives with negative decoded opacity from compositing and backprop, deactivating structures absent in specific periods.

Fig. 2. Overview of ChronoGS. For a given camera, visible anchors are selected from the learned scaffold of union geometry. Per-anchor features and global features are temporally modulated by encoding e(t) and decoded by a lightweight MLP into Gaussian attributes, which are rendered via differentiable splatting.

Fig. 2. Overview of ChronoGS. For a given camera, visible anchors are selected from the learned scaffold of union geometry. Per-anchor features and global features are temporally modulated by encoding e(t) and decoded by a lightweight MLP into Gaussian attributes, which are rendered via differentiable splatting.

Anchor Scaffold of Union Geometry: Initialized by voxelizing merged sparse point clouds from all periods, covering both stable and varying geometry to provide a consistent geometric backbone for cross-period reconstruction.

Results

ChronoScene Real
22.16
PSNR↑
+1.25 vs realtime4DGS Averaged over 6 real-world scenes
ChronoScene Real
0.6533
SSIM↑
+0.1409 vs realtime4DGS Averaged over 6 real-world scenes
ChronoScene Real
0.3390
LPIPS↓
−0.1553 vs realtime4DGS Averaged over 6 real-world scenes
ChronoScene Synthetic
28.80
PSNR↑
+6.51 vs realtime4DGS Averaged over 6 synthetic scenes
ChronoScene Synthetic
0.8562
SSIM↑
+0.0788 vs realtime4DGS Averaged over 6 synthetic scenes
ChronoScene Synthetic
0.2509
LPIPS↓
−0.0842 vs realtime4DGS Averaged over 6 synthetic scenes
Memory
0.52GB
Real
−90.6% vs realtime4DGS (5.53GB) Model storage footprint
Memory
0.65GB
Synthetic
−91.2% vs realtime4DGS (7.41GB) Model storage footprint
MethodMem.↓PSNR↑SSIM↑LPIPS↓
3DGS1.01GB18.290.46580.4862
Scaffold-GS0.36GB16.270.37490.5363
GS-W0.23GB20.330.40180.5638
4DGS0.18GB19.030.46740.6085
realtime4DGS5.53GB20.910.51240.4943
ChronoGS ours0.52GB22.160.65330.3390

Fig. 7 shows temporal evolution renderings at continuous period indices. Compared to dynamic baselines such as 4DGS that assume smooth motion and often hallucinate intermediate structures, ChronoGS cleanly switches between periods with more plausible intermediate states.

Qualitative Results

Fig. 6. Qualitative comparison on ChronoScene. Static models produce ghosting and appearance blending when trained on mixed multi-period data, while dynamic methods fail under large temporal gaps due to their smooth-motion assumption. ChronoGS faithfully reconstructs period-specific geometry and appearance.

Fig. 6. Qualitative comparison on ChronoScene. Static models produce ghosting and appearance blending when trained on mixed multi-period data, while dynamic methods fail under large temporal gaps due to their smooth-motion assumption. ChronoGS faithfully reconstructs period-specific geometry and appearance.

Fig. 7. Temporal evolution at a test viewpoint. ChronoGS queries continuous period indices and renders from the same camera, cleanly switching between periods with more plausible intermediate states compared to 4DGS.

Fig. 7. Temporal evolution at a test viewpoint. ChronoGS queries continuous period indices and renders from the same camera, cleanly switching between periods with more plausible intermediate states compared to 4DGS.

Results

Takeaway

By decomposing multi-period scenes into shared canonical geometry and period-specific variations through a unified scaffold + temporally modulated features + geometry activation, faithful period-specific reconstruction with cross-period consistency can be achieved in a single model.

Citation

@inproceedings{wang2026chronogs,
  title={ChronoGS: Disentangling Invariants and Changes in Multi-Period Scenes},
  author={Wang, Zhongtao and Dai, Jiaqi and Zhu, Qingtian and Li, Yilong and Su, Mai and Zhu, Fei and Gai, Meng and Wang, Shaorong and Pan, Chengwei and Chen, Yisong and Wang, Guoping},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2026}
}