cs.CVOct 9, 2026

Make It Flow: Animating Water in 3D Gaussian Splatting Scenes

Authors: Kjell Keune, Florian Hahlbohm, Jannis Möller, Susana Castillo, Martin Eisemann, Marcus Magnor

Organizations: Computer Graphics Lab, TU Braunschweig, Germany

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Explicit 3D scene representations such as 3D Gaussian Splatting (3DGS) have recently enabled high-quality, real-time novel view synthesis. However, these methods assume static scenes, limiting realism in the presence of dynamic phenomena ubiquitous in real-world environments. While coherent motion (e.g., humans or vehicles) can be captured and reconstructed, high-frequency effects such as flowing water remain intractable under casual capture conditions, where dense multi-view, high-speed acquisition is not feasible. We present a method to animate water in static 3DGS scenes, enabling visually plausible motion without requiring dynamic capture or physical simulation. Our approach estimates 2D optical flow from reference images or model renderings and lifts it into a consistent 3D Eulerian motion field. We propose a cinemagraph-inspired algorithm that applies this motion field to deform Gaussian primitives, producing temporally seamless looping animations. The resulting representation retains real-time rendering performance, significantly enhances perceptual realism, and preserves free-viewpoint navigation within the scene. We evaluate our method on a newly captured dataset of in-the-wild waterfall scenes and demonstrate clear improvements over prior work in dynamic Gaussian splatting and 3D cinemagraph generation in visual quality and computational efficiency.

Figures & tables

Figure 1: We generate optical flow depicting fluid motion from multi-view observations of a static 3DGS reconstruction by refining user-defined motion hints with segmentation masks and GAN-based flow estimation. Utilizing depth renderings from the 3DGS model, we reproject optical flow into a Eulerian motion field and create seamlessly looping animations by transforming multiple instances of dynamic Gaussians using a symmetric animation technique.
Figure 2: Qualitative comparison with baselines for novel-view renderings. Our method creates a visually plausible time-consistent fluid animation, while 3D-Cinemagraphy [ LCS∗23 ] introduces artifacts, 4DGS [ YYPZ24 ] fails to capture high-frequency water motion and generates a non-plausible fluid animation, and LoopGaussian [ LCW∗24 ] creates motion that does not follow realistic fluid flow (cf. our video).
Figure 3: Qualitative comparison between LoopGaussian [ LCW∗24 ] and our method on four scenes from our dataset where we visualize scene flow by coloring Gaussians according to their current motion [ BRS∗07 ] . We show that our method creates time-consistent motion that resembles the natural flow of water, while the motion synthesized by LoopGaussian points in arbitrary directions. The results were rendered from a novel viewpoint.
Scene#ImagesResolutionCamera∣G∣\left|\mathcal{G}\right|∣Gdyn∣|\mathcal{G}^{\textit{dyn}}|∣H∣\left|\mathcal{H}\right|
Creek1036223×4080DSLR1.59M51K11
Frogfall1866223×4080DSLR6.72M26K12
Silverfalls673811×2144Drone5.62M51K10
Smallfall746223×4080DSLR5.56M87K8
Town1596223×4080DSLR3.11M12K8
Waterfall1586223×4080DSLR5.26M182K14
Table 1: Properties of all dataset scenes including number of total Gaussians ∣G∣\left|\mathcal{G}\right| , dynamic Gaussians ∣Gdyn∣|\mathcal{G}^{\textit{dyn}}| and 3D motion hints ∣H∣\left|\mathcal{H}\right| .
Figure 4: Ground truth images along with renderings of the static and animated 3DGS reconstruction, depth renderings, dense flow hints Fndense\mathcal{F}^{\text{dense}}_{n} , refined optical flow Fn\mathcal{F}_{n} , and scene flow by coloring Gaussians according to their current motion [ BRS∗07 ] . Depicted camera poses are used as closest input poses for novel-view renderings (cf. Fig. 2 and Fig. 3 ).
Figure 5: Ablation of opacity blending and state offset. We show that both techniques are required for creating a seamless loop without noticeable cuts (top) and reoccurring patterns (middle).
Figure 6: Ablation of motion field refinements. The top row shows a voxel grid visualization, and the bottom row Gaussians colored based on their current motion. By refining static regions and view hit thresholding, we can eliminate motion in static regions (left) and wrong motion (middle).
Figure 7: We show that on some scenes (top) a single directional 3D motion hint can produce nearly the same motion as many hints. On other scenes (bottom), one hint does not suffice to model the transition from a falling waterfall into a flowing river. Hints are illustrated as red lines.
 3D-Cinema.4DGSLoopGaussianOurs3DGS
SceneVRAMFPSVRAMFPSVRAMFPSVRAMFPSVRAMFPS
Creek10.5 GiB<12.6 GiB4312.1 GiB4291.3 GiB4570.9 GiB650
Frogfall10.8 GiB<12.8 GiB3508.7 GiB2755.5 GiB2822.9 GiB283
Silverfalls13.6 GiB<12.7 GiB3017.3 GiB2824.6 GiB2832.4 GiB328
Smallfall10.5 GiB<12.7 GiB4307.2 GiB3494.5 GiB3422.3 GiB389
Town10.4 GiB<12.6 GiB4324.0 GiB4092.5 GiB4071.4 GiB569
Table 2: Quantitative comparison for rendering one animation loop.
 LoopGaussianOurs
SceneTotal TimeTotal TimeMasksFlowFusionStorage
Creek3m42s40s22s16s2s5.25.2 MB
Frogfall1m55s1m03s38s17s8s4444 KB
Silverfalls3m26s37s22s13s2s0.20.2 MB
Smallfall6m25s28s17s9s2s0.70.7 MB
Town1m01s55s33s16s6s1.11.1 MB
Table 3: Quantitative comparison of motion field preprocessing. For our method, we also show storage cost and sub-task timings.
Figure 8: Our method generalizes to fluids other than water, such as lava. We show the animation from a novel camera trajectory in our video.
Figure 9: Failure case on “Weir”: inaccurate Gaussian geometry (depth, left) and a poorly reconstructed water surface (middle) lead to wrongly reprojected flow and thus an inaccurate motion field, visible as animation artifacts (right).
Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Sceneinitial ∣C∣|C|refined ∣C∣|C|refined ρ\rhoaverage wvieww^{\text{view}}rviewr^{\text{view}}
Creek75 M14 M2.7 %95
Frogfall50 M2.6 M0.1 %378
Silverfalls70 M0.3 M4.7 %128
Smallfall50 M1.4 M3.6 %167
Town60 M2.0 M4.1 %196
Waterfall60 M34 M0.8 %238
Appendix
Table 4: Properties of per scene motion fields including initial and refined number of voxels ∣C∣|C| , density ρ\rho of the refined grid, average view hits per voxel wvieww^{\text{view}} and view hit threshold rviewr^{\text{view}} used.
Figure 10: Illustration of our symmetric animation technique for one splatted Gaussian. Forward instance Afw\mathcal{A}^{\text{fw}} (blue) and backward instance Abw\mathcal{A}^{\text{bw}} (red) are illustrated at different keyframes. Additionally, the Gaussian contour is drawn independent of Gaussian opacity, where Ofw\mathcal{O}^{\text{fw}} and Obw\mathcal{O}^{\text{bw}} are plotted. The Gaussian primitive is deformed by the underlying motion field (see Sec. 3.3.1 ).
Figure 11: Ablation of Gaussian transformation. Only displacing Gaussian means with constant covariance creates noisy artifacts (top row). A small α\alpha (see Eq. 7 ) results in blurry unrealistic deformations (middle row), while a larger α\alpha creates convincing transformations following fluid motion.
Figure 12: We demonstrate the use of additional positional offset instances. While without positional offset (left) the waterfall is sparsely reconstructed, additional instances (right) create more voluminous and denser fluid regions.
More in this topic

3D Gaussian Splatting

3D Gaussian Splatting represents scenes as optimizable 3D Gaussian primitives for fast, photorealistic rendering from new viewpoints.

See all 1,027 papers

Explore similar work

CardsList
Sep 29, 2026

Eulerian Motion Reconstruction for Water Scenery

Reconstructing and animating water scenery from nature produces compelling and immersive visual experiences. Previous work examined this task from the perspective of 2D video textures, with the goal of creating a looping video. In our work, we tackle the problem from a 3D perspective, creating a looping 4D dynamic reconstruction which can be interactively rendered from novel viewpoints from a single non-looping 2D source video. We represent motion as a 3D static Eulerian motion field that advects canonical Gaussian splats that are cyclically reborn at fixed time periods, supervised using rendering losses. To model non-periodic and stochastic dynamics present in real-world scenes, we add a non-periodic, time-varying residual term to capture deviations from the static Eulerian motion field. We show quantitatively and qualitatively that our framework enables photorealistic animation of water scenes better than prior art.

Chuhan Chen, Yen-Chi Cheng, Ayush Saraf +8

Carnegie Mellon University · University of Illinois Urbana-Champaign · Meta

4D Reconstruction3D ReconstructionDynamic Scene ReconstructionMotion Representation Learning
AskPDF
Sep 29, 2026

DispFlow-GS: Displacement Flow Supervision with Motion Disentangling for Monocular Deformable 3D Gaussian Splatting

Accurate dynamic scene reconstruction is important for robotic perception, where temporally consistent representations of dynamic environments are essential. Deformable 3D Gaussian Splatting (3DGS) models dynamic scenes through deformation fields, and recent methods incorporate motion supervision by aligning rendered Gaussian flow with optical flow. However, we find that such Gaussian-flow-based supervision provides only limited improvements in motion modeling. We identify a fundamental limitation of this supervision paradigm, namely a domain gap between rendered Gaussian flow and optical flow. To address this limitation, we propose a motion supervision framework built on Displacement Flow, which splats per-Gaussian 3D displacements onto the image plane to provide direct and stable optimization signals. We further disentangle scene motion from camera motion via intermediate-view rendering, enabling more reliable motion priors and targeted constraints on deformation and geometry. We also observe a discrepancy between motion fidelity and image-based evaluation, where improved motion awareness does not necessarily translate into better rendered image quality or higher image-based metric scores. Motivated by this mismatch, we introduce Deformation-Rendering Consistency (DRC), a motion-aware metric that measures the alignment between predicted deformation and rendering improvement. Experiments on dynamic scene benchmarks show substantial improvements in motion localization and motion--rendering consistency, reaching up to 39% and 6%, respectively, while image-based metrics change by only about 0.1%. These results confirm the observed mismatch between motion fidelity and image-based evaluation, demonstrating the significance of DRC for motion-aware evaluation.

Thai Duy Nguyen, Haitian Zhang, Addison Lin Wang

Nanyang Technological University, Singapore.

3D Gaussian SplattingMonocular 3D ReconstructionDynamic Scene Reconstruction
AskPDF
Jul 20, 2026

AniGS: Bridging Rendering and Diffusion Prior for 3D Scene Animation

Novel view rendering of large and complex reconstructed scenes is becoming increasingly photorealistic. However, most reconstructions remain static and lack the ambient motion that makes environments immersive. We present AniGS, a method for scene-level animation of 3D Gaussian Splatting (3DGS) reconstructions that adds subtle, distributed dynamics, e.g., vegetation motion, while preserving rigid structures. Unlike existing 3D animation techniques which are limited to object-centric subjects or small regions, AniGS is designed for large, cluttered, navigable scenes. AniGS represents the scene with a canonical 3DGS and models motion using a time-conditioned deformation field. To animate the entire scene, we leverage a pretrained video diffusion model and introduce an iterative dataset--model update strategy that progressively expands viewpoint coverage and repeatedly updates camera-fixed training videos using a render-and-refine scheme. To prevent artifacts from unintended motion in static areas, we further introduce a composed video-to-video refinement scheme that restricts motion to desired regions. Experiments on five real-world, large-scale outdoor scenes demonstrate that AniGS produces natural ambient dynamics and high-quality novel view videos, enabling more immersive viewing experiences of reconstructed environments.

Yen-Chi Cheng, Chen Gao, Chuhan Chen +8

University of Illinois Urbana-Champaign, Champaign-Urbana, Illinois, USA · Meta · Waymo, Bellevue, Washington, USA +2

3D Gaussian SplattingDynamic Scene Reconstruction
AskPDF