DimFlow: Bidirectional Grid-Expert Learning for Dynamic Gaussian Reconstruction

Gabriel Adineera; Jinoh Kim

cs.CVOct 26, 2026

DimFlow: Bidirectional Grid-Expert Learning for Dynamic Gaussian Reconstruction

Authors: Gabriel Adineera, Jinoh Kim

Organizations: Gabriel Adineera and Jinoh Kim are with Texas A&M University-Commerce, USA.

Abstract

Dynamic Gaussian reconstruction requires compact scene features to support coordinated changes in geometry and appearance. A factorized grid provides efficient local queries, but accurate interpolation alone does not ensure that the resulting dynamic attributes remain compatible across space and time. We introduce DimFlow, a framework that learns these features through bidirectional interaction between a multi-resolution grid and a coordinate-based expert branch. Feature-level adversarial alignment connects their distributions, adaptive fusion transfers complementary information, and activation-masked residual feedback refines the grid during training. Four attribute experts organize deformation, rotation, opacity, and temporal information before Gaussian rendering. In the aggregate comparison across D-NeRF, HyperNeRF, and DAVIS, DimFlow achieves 39.73 dB PSNR, improving on 4D-GS by 1.92 dB while matching its reported 85 FPS rendering speed. The model occupies 18.9 MB, compared with 18.3 MB for 4D-GS. Component and attribute-expert ablations reveal complementary contributions from feature alignment, adaptive exchange, and temporal processing.

Figures & tables

Fig. 1: Two feature views of the same scene. Two-dimensional projections of grid and coordinate-expert features without and with adversarial alignment. Each panel contains 350 samples from each branch. Distributional overlap is interpreted together with the reconstruction ablation.

Panel of Figure 2 — Fig. 2: DimFlow overview. Grid and coordinate features are adaptively fused before four gated attribute experts predict dynamic Gaussian updates. Canonical Gaussians provide the reference state, and differentiable splatting produces the requested view. The lower band summarizes distribution alignment, activation-masked feature feedback, and temporal regularity used during training. All illustrated planes and Gaussian clouds are schematic.

Fig. 3: Temporal averaging and trend compensation. Analytical frequency response of Equation 17 for unit-spaced time indices. Increasing

\eta

preserves more rapid variation, illustrating why trend compensation and temporal regularization serve different roles. The curves are derived from the stated operator.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Train $\downarrow$	FPS $\uparrow$	Storage $\downarrow$
	dB			min $\pm$ s		MB
K-Planes [ 6 ]	$35.71\,{\scriptscriptstyle\pm 0.14}$	$0.97\,{\scriptscriptstyle\pm 0.00}$	$0.05\,{\scriptscriptstyle\pm 0.01}$	$49\,{\scriptscriptstyle\pm 42}$	$0.97\,{\scriptscriptstyle\pm 0.00}$	$418.2\,{\scriptscriptstyle\pm 0.00}$
TiNeuVox-B [ 4 ]	$35.42\,{\scriptscriptstyle\pm 0.27}$	$0.97\,{\scriptscriptstyle\pm 0.01}$	$0.03\,{\scriptscriptstyle\pm 0.01}$	$26\,{\scriptscriptstyle\pm 39}$	$1.5\,{\scriptscriptstyle\pm 0.00}$	$48.1\,{\scriptscriptstyle\pm 0.00}$
HexPlane [ 5 ]	$35.21\,{\scriptscriptstyle\pm 0.12}$	$0.97\,{\scriptscriptstyle\pm 0.00}$	$0.04\,{\scriptscriptstyle\pm 0.00}$	$10\,{\scriptscriptstyle\pm 27}$	$2.5\,{\scriptscriptstyle\pm 0.00}$	$37.9\,{\scriptscriptstyle\pm 0.00}$
4D-GS [ 9 ]	$37.81\,{\scriptscriptstyle\pm 0.11}$	$\mathbf{0.98}\,{\scriptscriptstyle\pm 0.01}$	$\mathbf{0.02}\,{\scriptscriptstyle\pm 0.00}$	$7\,{\scriptscriptstyle\pm 39}$	$\mathbf{85}\,{\scriptscriptstyle\pm 0.00}$	$\mathbf{18.3}\,{\scriptscriptstyle\pm 0.00}$
MSTH [ 7 ]	$34.88\,{\scriptscriptstyle\pm 0.09}$	$\mathbf{0.98}\,{\scriptscriptstyle\pm 0.00}$	$0.03\,{\scriptscriptstyle\pm 0.01}$	$\mathbf{5}\,{\scriptscriptstyle\pm 18}$	–	–

TABLE I: Aggregate reconstruction quality and efficiency across D-NeRF, HyperNeRF, and DAVIS. Bold marks the best value, including ties at the displayed precision. Training time is in minutes with variation in seconds; missing entries are unreported.

Panel of Figure 5 — Fig. 4: Qualitative reconstruction views. Cat and Dancer comparisons highlight local appearance, silhouettes, and motion boundaries. Ground-truth views can include motion blur. The complete eight-method comparison is available in the supplementary material.

Fig. 5: Quality, speed, and component effects. (a) Aggregate PSNR and frame rate from Table I ; vertical error bars show the reported PSNR variation. (b) PSNR decrease for component removals in Table II . The panels summarize separate evaluation settings. MSTH is absent from (a) because its frame rate was not reported.

Variant	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$\Delta$
Full model	$\mathbf{39.73}\,{\scriptscriptstyle\pm 0.06}$	$\mathbf{0.98}\,{\scriptscriptstyle\pm 0.00}$	$\mathbf{0.02}\,{\scriptscriptstyle\pm 0.00}$	0.00
w/o adversarial alignment	$37.85\,{\scriptscriptstyle\pm 0.10}$	$0.97\,{\scriptscriptstyle\pm 0.01}$	$0.04\,{\scriptscriptstyle\pm 0.01}$	1.88
w/o adaptive weighting	$38.12\,{\scriptscriptstyle\pm 0.08}$	$0.97\,{\scriptscriptstyle\pm 0.00}$	$0.03\,{\scriptscriptstyle\pm 0.01}$	1.61
w/o temporal feature flow	$37.94\,{\scriptscriptstyle\pm 0.09}$	$0.97\,{\scriptscriptstyle\pm 0.01}$	$0.04\,{\scriptscriptstyle\pm 0.00}$	1.79
w/o bidirectional exchange	$38.02\,{\scriptscriptstyle\pm 0.07}$	$0.97\,{\scriptscriptstyle\pm 0.01}$	$0.03\,{\scriptscriptstyle\pm 0.01}$	1.71
w/o coordinate expert	$37.70\,{\scriptscriptstyle\pm 0.11}$	$0.96\,{\scriptscriptstyle\pm 0.01}$	$0.05\,{\scriptscriptstyle\pm 0.01}$	2.03

TABLE II: Component ablations on D-NeRF.

\Delta

is the PSNR decrease from the full model, calculated from the displayed means.

Metric	Full model	No deformation	No temporal
PSNR $\uparrow$	38.79	36.53	37.12
SSIM $\uparrow$	0.98	0.95	0.96
LPIPS $\downarrow$	0.03	0.05	0.04
FlowsME $\downarrow$	0.12	0.31	0.58
TSV $\downarrow$	0.015	0.028	0.030
EDE $\downarrow$	0.08	0.21	0.16

TABLE III: Attribute-expert ablations. The full-model baseline in this setting is 38.79 dB.

\Delta

PSNR is calculated relative to this baseline.

Fig. 6: Expert-specific sensitivity. PSNR and optical-flow consistency error from Table III . “No def.” removes the deformation expert; “No temp.” removes the temporal expert. The full-model baseline here is distinct from the aggregate comparison.

Dynamic Scene Reconstruction

Dynamic scene reconstruction builds time-varying 3D models of moving objects and environments from video or multi-view images.

See all 288 papers

DSD-GS: Dynamic-Static Decomposition of Gaussian Splatting for Efficient and High-Fidelity Dynamic Scene Reconstruction
May 29, 2026Youngtae Han, Sung-hwan Han, Youngmin Yi3D Gaussian Splatting Dynamic Scene Reconstruction
Department of Artificial Intelligence Engineering, Sogang University
Learning Global Motion with Compact Gaussians for Feed-Forward 4D Reconstruction
May 29, 2026Mungyeom Kim, Minkyeong Jeon, Honggyu An +10Point Tracking 3D Gaussian Splatting
KAIST AI · ETH Zürich, ETH AI Center · Sony AI +1
Eulerian Gaussian Splatting using Hashed Probability Pyramids
May 27, 2026Mia Gaia Polansky, George Kopanas, Stephan Garbin +23D Gaussian Splatting Neural Rendering
Harvard University · Google DeepMind · Google

May 29, 2026J/K move · Enter open · S save

DSD-GS: Dynamic-Static Decomposition of Gaussian Splatting for Efficient and High-Fidelity Dynamic Scene Reconstruction

Youngtae Han, Sung-hwan Han, Youngmin Yi

Department of Artificial Intelligence Engineering, Sogang University

3D Gaussian Splatting Dynamic Scene Reconstruction Dynamic Novel View Synthesis

Open paper AskPDF arXiv

Dynamic scene reconstruction and novel view synthesis are fundamental to next-generation visual intelligence applications such as virtual reality, robotics, and digital twins. However, high-fidelity reconstruction of complex, time-varying scenes from arbitrary viewpoints remains a significant challenge. Existing dynamic 3DGS methods suffer from computational inefficiency, since they model all Gaussians as dynamic components. While recent decomposition-based approaches address this issue, they still struggle with degraded reconstruction quality and prolonged training time. To mitigate these limitations, we propose a novel dynamic reconstruction framework built upon an efficient static-dynamic decomposition strategy using a Feed-Forward Gaussian Splatting encoder and an optical flow model. By eliminating redundant computations on static regions, our method achieves state-of-the-art performance, outperforming existing baselines across rendering quality, training and rendering speed, and storage efficiency. Notably, on the Neural 3D dataset, our framework requires only 10 minutes for training and achieves a rendering speed of over 700 FPS on a single NVIDIA RTX 5090 GPU at resolution of 1352x1014. Furthermore, our decomposition strategy eliminates the need for COLMAP preprocessing and enables deterministic initialization, thereby enhancing both efficiency and reproducibility.

Abstract

Dynamic Scene Reconstruction

Explore similar work

DSD-GS: Dynamic-Static Decomposition of Gaussian Splatting for Efficient and High-Fidelity Dynamic Scene Reconstruction

Learning Global Motion with Compact Gaussians for Feed-Forward 4D Reconstruction

Eulerian Gaussian Splatting using Hashed Probability Pyramids