Video Stabilization through Rendering-Aware Trajectory Optimization<br>in 3DGS-Reconstructed Scenes
StabiGS:Video Stabilization through Rendering-Aware Trajectory Optimization<br>in 3DGS-Reconstructed Scenes
Souheib Ben Mabrouk1,2
Jean-Emmanuel Deschaud1
Eva Coupeté2
Thomas Derbanne2
Nicolas Rahmouni2
1Mines Paris PSL
2GoPro
CVPR 2026 (Findings)
Paper
arXiv
-->
Poster
Synthetic Data
Results
Code (coming soon)
TL;DR: Full-frame video stabilization through global 3D scene reconstruction and rendering-aware trajectory optimization.
Abstract
Existing video stabilization methods based on 2D warping or local 3D cues often struggle to preserve full-frame content and geometric consistency, particularly under large camera motions or complex scene geometry. We present StabiGS , a novel approach that formulates video stabilization as a rendering-aware view synthesis problem. Our method reconstructs a global 3D scene representation and jointly optimizes the camera smoothness and the rendering quality to compute the stabilized trajectory. Thus, unlike filtering-based approaches, StabiGS produces stable videos while maintaining geometric consistency. To enable a comprehensive evaluation, we present a synthetic benchmark with controllable camera shakiness and accurate ground-truth poses, enabling a reliable assessment of stabilization and rendering quality. Extensive experiments and a user study demonstrate that StabiGS achieves state-of-the-art performance in stabilization and rendering quality.
Shaky
StabiGS
Comparison
Here, we compare StabiGS with state-of-art video stabilization methods. Our method enables cinematic full-frame stabilization while preserving geometric consistency, even under intense camera motion and complex scene geometry.
Comparison Method:
Shaky vs StabiGS sequence (set 1).
-->
Shaky
StabiGS
RStab vs StabiGS sequence (set 2).
-->
RStab
StabiGS
Deep3D vs StabiGS sequence (set 3).
-->
Deep3D
StabiGS
Meshflow vs StabiGS sequence (set 4).
-->
Meshflow
StabiGS
FuSta vs StabiGS sequence (set 5).
-->
FuSta
StabiGS
DIFRINT vs StabiGS sequence (set 6).
-->
DIFRINT
StabiGS
GaVS vs StabiGS sequence (set 7).
-->
GaVS
StabiGS
Notice: This website is currently under reconstruction.
Related Work
- The real-world videos are based on the following works:
Progressively Optimized Local Radiance Fields for Robust View Synthesis
GaVS: 3D-Grounded Video Stabilization via Temporally-Consistent Local Reconstruction and Rendering
Deep Online Fused Video Stabilization
- The synthetic videos are created using Blender and the assets are from the following works:
NVIDIA ORCA: Open Research Content Archive
Blenderkit Scene: Tropical forest
Blenderkit Scene: Scandinavian decoration
BibTeX
@InProceedings{Ben_Mabrouk_2026_CVPR,<br>author = {Ben Mabrouk, Souheib and Deschaud, Jean-Emmanuel and Coupet\'e, Eva and Derbanne, Thomas and Rahmouni, Nicolas},<br>title = {StabiGS: Video Stabilization through Rendering-Aware Trajectory Optimization in 3DGS-Reconstructed Scenes},<br>booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Findings},<br>month = {June},<br>year = {2026},<br>pages = {8481-8491}