Pith. sign in

REVIEW 4 major objections 5 minor 53 references

STDR: Spatio-Temporal Decoupling for Real-Time Dynamic Scene Rendering

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper argues that spatio-temporal incoherence during Gaussian initialization, not the deformation model itself, is a key bottleneck in dynamic 3D Gaussian Splatting reconstruction, and proposes a plug-and-play module that resolves it…

desk verdict A useful plug-and-play module for dynamic 3DGS with consistent gains across four baselines, but the spatial-smoothness regularizer may fight the method's own decoupling at motion boundaries, and the paper needs error bars. read the letter →

arxiv 2505.22400 v1 pith:IJKFBINR submitted 2025-05-28 cs.GR cs.CV

classification cs.GRcs.CV
keywords dynamicscenereconstruction3DGaussianSplattingspatio-temporaldecouplingmaskdeformationfieldtemporalconsistencyreal-timerenderingnovelviewsynthesis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Dynamic scene reconstruction with 3D Gaussian Splatting usually builds a canonical set of Gaussians by merging all training frames as though they shared one timestamp. The paper argues that this initialization choice is itself a source of error: moving objects become smeared across positions, producing 'ghost' Gaussians that a deformation field must later untangle without any temporal cue. To fix this, STDR gives each Gaussian a learnable spatio-temporal mask that weights its opacity per timestamp, feeds the resulting temporal probability distribution into a separated deformation field, and adds temporal-smoothness and spatial-awareness losses. When plugged into four existing 3DGS-based dynamic pipelines, the module improves PSNR, SSIM, and LPIPS on synthetic and real-world benchmarks while keeping rendering real-time. The point, if true, is that a model-agnostic fix to initialization can clean up dynamic reconstruction without redesigning the deformation machinery.

What carries the argument

The load-bearing mechanism is a learnable spatio-temporal mask $m_i \in \mathbb{R}^K$ attached to each Gaussian, where $K$ is the number of timestamps. During rendering the mask rescales the Gaussian's opacity as $\alpha_i^{\mathrm{new}} = m_i^j \alpha_i$ at timestamp $j$, and after softmax normalization $\tilde{m}_i^j = \exp(m_i^j) / \sum_{k=1}^K \exp(m_i^k)$ it acts as a spatio-temporal probability distribution fed into a separated deformation field $f_{\mathrm{sep}}$, which extracts spatial and temporal features $z_s, z_t$ used by $f_{\mathrm{def}}(x_\gamma, z_s, z_t, t_\tau)$. The mask is trained in a warm-up phase with the original opacity frozen, so temporal gradients are forced through the mask; two regularizers, temporal smoothness $L_{\mathrm{temp}}$ and a KL-based spatial-awareness term $L_{\mathrm{spatial}}$, then encourage the masks to vary smoothly in time and to agree among spatial neighbors. This is what lets the method claim to separate 'spatially close but temporally distant' Gaussians before the deformation field sees them.

What would settle it

Render a synthetic scene where two objects pass through the same location at different times and share a motion boundary, then freeze the deformation field and inspect the learned masks; if the masks do not concentrate on each object's true visible timestamps and instead blend across the boundary, the decoupling has not actually resolved the temporal ambiguity.

Watch

Extended reading notes

Core claim

The paper's central claim is that the 'spatio-temporal incoherence' created during warm-up initialization is a principal bottleneck for 3DGS-based dynamic scene reconstruction. It proposes STDR, a plug-and-play module with three pieces: a per-Gaussian spatio-temporal mask that modulates opacity so each Gaussian can express which timestamps it belongs to; a separated deformation field that consumes the normalized mask as a spatio-temporal probability distribution and factorizes spatial and temporal features; and a consistency regularization that penalizes abrupt mask changes over time and pulls each Gaussian's temporal distribution toward its spatial neighbors'. The authors report that inserting STDR into DeformGS, SC-GS, SPGS, and 4DGS yields consistent gains in PSNR, SSIM, and LPIPS across D-NeRF, NeRF-DS, and HyperNeRF, and that rendering speed remains real-time. The intended takeaway is that an explicit temporal identity per Gaussian, learned during initialization, resolves the ambiguity that otherwise causes ghosting, motion blur, and temporally inconsistent deformation.

Load-bearing premise

The central assumption is that a simple learned weight per Gaussian per timestamp, trained only against reconstruction error during warm-up, can correctly sort the temporally smeared canonical Gaussians into the right frames, and that nearby points genuinely move in synchrony.

Editorial extensions

If this is right

  • Inserting STDR into 4DGS, DeformGS, SPGS, or SC-GS improves PSNR, SSIM, and LPIPS on D-NeRF, NeRF-DS, and HyperNeRF relative to the same baselines without STDR.
  • The learned spatio-temporal masks separate static from dynamic content on their own: static Gaussians converge to uniform masks across time, while dynamic ones develop sharp, time-specific activations, so the mask can serve as an interpretable motion cue.
  • Freezing the original opacity during warm-up and routing gradients solely through the mask is sufficient to resolve initialization ambiguity, and removing either the temporal smoothness or spatial awareness loss degrades the results.
  • Real-time rendering survives the addition: reported frame rates on D-NeRF scenes remain in the tens of FPS after integration, so the decoupling does not come at the cost of interactivity.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: since the mask distributions emerge without explicit supervision, they could be recycled as free static/dynamic segmentation labels or as a motion-saliency prior for downstream editing tasks; the paper describes the emergent property but does not train or evaluate such a use.
  • Editorial inference: the KL spatial-awareness term assumes neighboring Gaussians share temporal behavior, so it may over-smooth masks exactly at motion boundaries; a weighted or edge-aware version of the neighbor term is a natural testable extension that the paper does not explore.
  • Editorial inference: if the identified failure is truly an initialization artifact, a similar temporal-identity mechanism could benefit NeRF-based dynamic renderers and other voxel or mesh representations, though the paper only demonstrates the module on 3DGS pipelines.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes STDR, a plug-and-play module for 3D Gaussian Splatting (3DGS) based dynamic scene reconstruction. The method augments each Gaussian with a learnable spatio-temporal mask that modulates opacity per timestamp, introduces a separated deformation field that consumes spatial and temporal features derived from the mask, and adds temporal smoothness and spatial KL consistency losses. The module is integrated into four existing baselines (DeformGS, SC-GS, SPGS, 4DGS) and evaluated on D-NeRF, NeRF-DS, and HyperNeRF, reporting consistent PSNR/SSIM/LPIPS improvements. The paper's central claim is that the module reduces spatio-temporal incoherence from canonical initialization and improves reconstruction quality and temporal consistency.

Significance. If the reported improvements are robust, STDR is a practically useful contribution: it is a lightweight, architecture-agnostic module that can be dropped into existing deformable 3DGS pipelines. The controlled integration into four baselines is a strength, and the motivation regarding temporally entangled initialization is well articulated. However, the significance is tempered by the absence of ablation on the two central components (mask and separated deformation field), the lack of error bars or multiple seeds, and the partial HyperNeRF evaluation. The paper does not provide code, which limits reproducibility of the exact training schedule and network details.

major comments (4)
  1. [Section 3.4, Eq. (9)] The spatial-awareness KL regularizer encourages each Gaussian's temporal mask to match its K spatial neighbors. This directly conflicts with the paper's stated goal of distinguishing 'Gaussians that are spatially close but temporally distant' (Sections 1 and 3.2). At motion boundaries, a foreground Gaussian and a background Gaussian are KNN neighbors but should have maximally different temporal activations; Eq. (9) pulls their masks together. The aggregate ablation in Table 3 shows that removing Lspatial hurts the mean PSNR, but it does not isolate boundary behavior. Please provide an analysis or experiment that evaluates learned masks and rendering at object boundaries (e.g., boundary-region metrics or mask visualizations) to show that the regularizer does not undermine the decoupling in the exact regime where the method claims to help, or explain why the asymmetric KL direction and the weight lambda2=0.2 prevent this failure mode.
  2. [Section 4.3, Table 3] The ablation only removes the two regularization terms Ltemp and Lspatial. The two other core components introduced in Section 3.3, namely the spatio-temporal mask and the separated deformation field, are not ablated. Since the paper's central contribution is the joint combination of these components, the reader cannot tell whether the mask, the separated deformation field, or the regularizers are responsible for the improvements. Please add ablations that remove the mask (falling back to the baseline deformation field), remove the separated deformation field (feeding only the original position or a standard time input), and combine all components, so that the contribution of each part is quantified.
  3. [Tables 1, 2, and 4] All quantitative results are reported as single runs with no standard deviations or number of seeds. Some improvements are small, e.g., SSIM differences of 0.001-0.005 in Table 1 and PSNR gains below 0.3 dB in Table 2, and it is unclear whether these are significant relative to run-to-run training variance. Please report mean +/- standard deviation over at least three seeds for the main comparisons, or justify why the improvements are robust despite the absence of variance information.
  4. [Section 4.2, Table 4] The HyperNeRF evaluation reports only PSNR and SSIM, without LPIPS, and covers only five scenes, while the abstract claims improvements in perceptual fidelity across real-world benchmarks. Please either add LPIPS results for HyperNeRF or explicitly state in the main text that the real-world perceptual evaluation is limited to NeRF-DS, so that the scope of the claim is precise.
minor comments (5)
  1. [Section 3.3] There is a typo in 'sptio-temporal mask'; it should be 'spatio-temporal mask'.
  2. [Section 3.4, Eq. (9) and Section 3.3] The symbol K is used both for the number of time frames in the mask m_i in R^K and for the number of neighbor Gaussians in Eq. (9). Please use distinct symbols, e.g., T for timestamps and K for neighbors, to avoid confusion.
  3. [Eq. (10) and Appendix A.1] The final loss in Eq. (10) contains lambda, lambda1, and lambda2, but Appendix A.1 only specifies lambda1=0.1 and lambda2=0.2. Please report the value of lambda or clarify how the reconstruction loss weights are set.
  4. [Section 4.2, Table 1] The baseline name 'Deformable3D' is used in the table, but the text refers to 'DeformGS'; please make the naming consistent. Also, 'TiNeuV ox' contains a spacing typo.
  5. [Appendix A.1] The separated deformation field is described as a multi-branch network taking the spatio-temporal mask vector as input, while Eq. (7) writes z_s, z_t = fsep(x_gamma, m~_gamma). Please clarify whether fsep also takes the Gaussian position x_gamma as input and how the two branches produce z_s and z_t, since this is central to reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: STDR is a learned internal module whose improvements are evaluated on held-out benchmark views; no derivation step reduces to its own inputs.

full rationale

The paper's core mechanism is a learnable per-Gaussian spatio-temporal mask m_i (Eq. 5), normalized by softmax (Eq. 6), fed into a separated deformation field (Eq. 7), and regularized by temporal smoothness and spatial-awareness terms (Eqs. 8-9). Nothing here is a prediction derived from first principles; the mask is an internal latent representation optimized by the reconstruction loss and then used to condition the deformation field. The central empirical claim, that adding STDR to DeformGS, SC-GS, SPGS, or 4DGS improves PSNR/SSIM/LPIPS on D-NeRF, NeRF-DS, and HyperNeRF, is evaluated on held-out test views of standard benchmarks against external baselines, so the result does not reduce to a fitted parameter renamed as a prediction. The only author self-citations ([3], [4], [20]) appear in related-work context and are not load-bearing: no uniqueness theorem is imported, no ansatz is justified solely by self-citation, and no known result is renamed as new. The appendix limitation (A.4) concedes that fixed temporal resolution may over-smooth under non-uniform dynamics, which is a robustness caveat rather than a circularity. The Eq. 9 spatial-KL assumption may be violated at motion boundaries, but that is a modeling-risk concern, not a circular derivation. Accordingly, the appropriate finding is no significant circularity (score 0).

Assumptions & free parameters 6 free parameters · 5 assumptions · 3 invented entities

The central claim depends on a handful of hand-chosen hyperparameters (λ1, λ2, λ, M, N, training schedule) and several domain assumptions about the nature of dynamic scene artifacts and the learnability of temporal masks. The method introduces three internal latent/module entities, none of which have independent empirical evidence.

free parameters (6)
  • λ1 = 0.1
    Weight for temporal smoothness loss (Eq. 8), chosen by hand; no sensitivity analysis reported.
  • λ2 = 0.2
    Weight for spatial-awareness KL loss (Eq. 9), chosen by hand; no sensitivity analysis reported.
  • λ = not reported
    Weight balancing L1 and D-SSIM in Eq. 10; value is not specified in the paper.
  • M = 1000
    Number of sampled Gaussians for the KL divergence in Eq. 9; set for computational efficiency.
  • N = 5
    Number of KNN neighbors used in the spatial-awareness regularizer; chosen by hand.
  • warm-up and freeze iterations = 3000 and 6000
    Training schedule: mask-only optimization for 3000 iterations, regularization until 6000, then mask freezing; affects the learned representations.
assumptions (5)
  • standard math 3D Gaussian Splatting rendering equations (Eqs. 1-3) are valid for the scenes.
    Used as the backbone representation and rendering model, assumed from prior work [14].
  • domain assumption The reconstruction loss is a weighted combination of L1 and D-SSIM (Eq. 10).
    Standard practice in 3DGS baselines; no derivation or justification is given beyond common usage.
  • domain assumption Spatio-temporal incoherence during initialization is the main cause of observed artifacts.
    The paper's motivating observation, argued in Section 3.2 but not proven; it underpins the need for the proposed module.
  • ad hoc to paper A per-Gaussian temporal mask can be learned via backpropagation and remains stable after freezing.
    This is the core mechanism of STDR; its convergence and stability are assumed and not formally analyzed.
  • ad hoc to paper Spatially adjacent Gaussians share similar temporal behavior (Eq. 9).
    Used in the spatial-awareness regularizer; this assumption may fail at motion boundaries or with fast-moving objects.
invented entities (3)
  • Spatio-temporal mask
    purpose: Modulate each Gaussian's opacity across timestamps to disentangle temporal identity.
    A learnable latent variable with no external falsifiable handle; its 'probability distribution' is used only internally.
  • Separated deformation field
    purpose: Factorize spatial and temporal features to guide Gaussian deformation.
    A neural network module; no independent handle outside the paper.
  • Spatio-temporal probability distribution
    purpose: Softmax-normalized mask providing input to the separated deformation field.
    Derived from the mask, it is an internal representation with no external validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of STDR: Spatio-Temporal Decoupling for Real-Time Dynamic Scene Rendering." pith.science (2026). https://pith.science/paper/IJKFBINR

@misc{pith2026250522400,
  author       = {Pith},
  title        = {Pith review of: STDR: Spatio-Temporal Decoupling for Real-Time Dynamic Scene Rendering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IJKFBINR}},
  note         = {Machine review of arXiv:2505.22400}
}
read the original abstract

Although dynamic scene reconstruction has long been a fundamental challenge in 3D vision, the recent emergence of 3D Gaussian Splatting (3DGS) offers a promising direction by enabling high-quality, real-time rendering through explicit Gaussian primitives. However, existing 3DGS-based methods for dynamic reconstruction often suffer from \textit{spatio-temporal incoherence} during initialization, where canonical Gaussians are constructed by aggregating observations from multiple frames without temporal distinction. This results in spatio-temporally entangled representations, making it difficult to model dynamic motion accurately. To overcome this limitation, we propose \textbf{STDR} (Spatio-Temporal Decoupling for Real-time rendering), a plug-and-play module that learns spatio-temporal probability distributions for each Gaussian. STDR introduces a spatio-temporal mask, a separated deformation field, and a consistency regularization to jointly disentangle spatial and temporal patterns. Extensive experiments demonstrate that incorporating our module into existing 3DGS-based dynamic scene reconstruction frameworks leads to notable improvements in both reconstruction quality and spatio-temporal consistency across synthetic and real-world benchmarks.

Figures

Figures reproduced from arXiv: 2505.22400 by the authors.

Figure 1
Figure 1. Visualization of spatio-temporal incoherence during canonical Gaussian initialization. (a) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed method. (a) During warm-up initialization, we assign each [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visualization of Comparisons on D-NeRF Dataset [ [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization of Comparisons on NeRF-DS Dataset [ [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Visualization of Comparisons on HyperNeRF Dataset [ [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 48 canonical work pages

  1. [1]

    Nerf in detail: Learning to sample for view synthesis, 2021

    Relja Arandjelovi ´c and Andrew Zisserman. Nerf in detail: Learning to sample for view synthesis, 2021

  2. [2]

    Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P

    Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields, 2021

  3. [3]

    Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks

    Yujun Cai, Liuhao Ge, Jun Liu, Jianfei Cai, Tat-Jen Cham, Junsong Yuan, and Nadia Magnenat Thalmann. Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , 2019

  4. [4]

    Learning progressive joint propagation for human motion prediction

    Yujun Cai, Lin Huang, Yiwei Wang, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Xu Yang, Yiheng Zhu, Xiaohui Shen, et al. Learning progressive joint propagation for human motion prediction. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part VII 16, pages 226–242. Springer, 2020

  5. [5]

    Click-gaussian: Interac- tive segmentation to any 3d gaussians, 2024

    Seokhun Choi, Hyeonseop Song, Jaechul Kim, Taehyeong Kim, and Hoseok Do. Click-gaussian: Interac- tive segmentation to any 3d gaussians, 2024

  6. [6]

    Tivne-slam: Dynamic mapping and tracking via time-varying neural radiance fields

    Chengyao Duan and Zhiliu Yang. Tivne-slam: Dynamic mapping and tracking via time-varying neural radiance fields. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , page 12324–12331. IEEE, 2024

  7. [7]

    Momentum-gs: Momentum gaussian self-distillation for high-quality large scene reconstruction, 2024

    Jixuan Fan, Wanhua Li, Yifei Han, and Yansong Tang. Momentum-gs: Momentum gaussian self-distillation for high-quality large scene reconstruction, 2024

  8. [8]

    Fast dynamic radiance fields with time-aware neural voxels

    Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, page 1–9. ACM, 2022

Show all 53 references
  1. [9]

    K- planes: Explicit radiance fields in space, time, and appearance, 2023

    Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K- planes: Explicit radiance fields in space, time, and appearance, 2023

  2. [10]

    Relaygs: Reconstructing dynamic scenes with large-scale and complex motions via relay gaussians, 2024

    Qiankun Gao, Yanmin Wu, Chengxiang Wen, Jiarui Meng, Luyang Tang, Jie Chen, Ronggang Wang, and Jian Zhang. Relaygs: Reconstructing dynamic scenes with large-scale and complex motions via relay gaussians, 2024

  3. [11]

    Speedy-splat: Fast 3d gaussian splatting with sparse pixels and sparse primitives, 2025

    Alex Hanson, Allen Tu, Geng Lin, Vasu Singla, Matthias Zwicker, and Tom Goldstein. Speedy-splat: Fast 3d gaussian splatting with sparse pixels and sparse primitives, 2025

  4. [12]

    Efficientnerf: Efficient neural radiance fields, 2022

    Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. Efficientnerf: Efficient neural radiance fields, 2022

  5. [13]

    Sc-gs: Sparse- controlled gaussian splatting for editable dynamic scenes, 2024

    Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse- controlled gaussian splatting for editable dynamic scenes, 2024

  6. [14]

    3d gaussian splatting for real-time radiance field rendering, 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering, 2023

  7. [15]

    Jackson, and Adrian Hilton

    Hansung Kim, Luca Remaggi, Philip J.B. Jackson, and Adrian Hilton. Immersive spatial audio reproduction for vr/ar using room acoustic modelling from 360° images. pages 120 – 126. IEEE, 2019

  8. [16]

    Vdg: Vision-only dynamic gaussian for driving simulation, 2024

    Hao Li, Jingfeng Li, Dingwen Zhang, Chenming Wu, Jieqi Shi, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, and Junwei Han. Vdg: Vision-only dynamic gaussian for driving simulation, 2024

  9. [17]

    Omnigs: Fast radiance field reconstruction using omnidirectional gaussian splatting, 2024

    Longwei Li, Huajian Huang, Sai-Kit Yeung, and Hui Cheng. Omnigs: Fast radiance field reconstruction using omnidirectional gaussian splatting, 2024

  10. [18]

    Neural 3d video synthesis from multi-view video, 2022

    Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multi-view video, 2022

  11. [19]

    Neural scene flow fields for space-time view synthesis of dynamic scenes, 2021

    Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes, 2021

  12. [20]

    Gradiseg: Gradient-guided gaussian segmentation with enhanced 3d boundary precision, 2024

    Zehao Li, Wenwei Han, Yujun Cai, Hao Jiang, Baolong Bi, Shuqin Gao, Honglong Zhao, and Zhaoqi Wang. Gradiseg: Gradient-guided gaussian segmentation with enhanced 3d boundary precision, 2024. 10

  13. [21]

    Vastgaussian: Vast 3d gaussians for large scene reconstruction, 2024

    Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, and Wenming Yang. Vastgaussian: Vast 3d gaussians for large scene reconstruction, 2024

  14. [22]

    Hybridgs: Decoupling transients and statics with 2d and 3d gaussian splatting, 2025

    Jingyu Lin, Jiaqi Gu, Lubin Fan, Bojian Wu, Yujing Lou, Renjie Chen, Ligang Liu, and Jieping Ye. Hybridgs: Decoupling transients and statics with 2d and 3d gaussian splatting, 2025

  15. [23]

    Citygaussian: Real-time high-quality large-scale scene rendering with gaussians, 2024

    Yang Liu, He Guan, Chuanchen Luo, Lue Fan, Naiyan Wang, Junran Peng, and Zhaoxiang Zhang. Citygaussian: Real-time high-quality large-scale scene rendering with gaussians, 2024

  16. [24]

    Manigaussian: Dynamic gaussian splatting for multi-task robotic manipulation, 2024

    Guanxing Lu, Shiyi Zhang, Ziwei Wang, Changliu Liu, Jiwen Lu, and Yansong Tang. Manigaussian: Dynamic gaussian splatting for multi-task robotic manipulation, 2024

  17. [25]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020

  18. [26]

    Newcombe, Dieter Fox, and Steven M

    Richard A. Newcombe, Dieter Fox, and Steven M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 343–352, 2015

  19. [27]

    Nerfies: Deformable neural radiance fields

    Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision , pages 5865–5874, 2021

  20. [28]

    Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M

    Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields, 2021

  21. [29]

    Desire-gs: 4d street gaussians for static-dynamic decompo- sition and surface reconstruction for urban driving scenes, 2024

    Chensheng Peng, Chengwei Zhang, Yixiao Wang, Chenfeng Xu, Yichen Xie, Wenzhao Zheng, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Desire-gs: 4d street gaussians for static-dynamic decompo- sition and surface reconstruction for urban driving scenes, 2024

  22. [30]

    D-nerf: Neural radiance fields for dynamic scenes, 2020

    Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes, 2020

  23. [31]

    Qi, Hao Su, Kaichun Mo, and Leonidas J

    Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation, 2017

  24. [32]

    Disc-gs: Discontinuity-aware gaussian splatting, 2024

    Haoxuan Qu, Zhuoling Li, Hossein Rahmani, Yujun Cai, and Jun Liu. Disc-gs: Discontinuity-aware gaussian splatting, 2024

  25. [33]

    Stopthepop: Sorted gaussian splatting for view-consistent real-time rendering, 2024

    Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Stein- berger. Stopthepop: Sorted gaussian splatting for view-consistent real-time rendering, 2024

  26. [34]

    Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians, 2024

    Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians, 2024

  27. [35]

    Tensor4d : Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering, 2023

    Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d : Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering, 2023

  28. [36]

    Flashsplat: 2d to 3d gaussian splatting segmentation solved optimally, 2024

    Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Flashsplat: 2d to 3d gaussian splatting segmentation solved optimally, 2024

  29. [37]

    Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields, 2023

    Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields, 2023

  30. [38]

    A vr system for immersive teleoperation and live exploration with a mobile robot

    Patrick Stotko, Stefan Krumpen, Max Schwarz, Christian Lenz, Sven Behnke, Reinhard Klein, and Michael Weinmann. A vr system for immersive teleoperation and live exploration with a mobile robot. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) ...

  31. [39]

    Efficient ray sampling for radiance fields reconstruction, 2023

    Shilei Sun, Ming Liu, Zhongyi Fan, Yuxue Liu, Chengwei Lv, Liquan Dong, and Lingqin Kong. Efficient ray sampling for radiance fields reconstruction, 2023

  32. [40]

    Superpoint gaussian splatting for real-time high-fidelity dynamic scene reconstruction, 2024

    Diwen Wan, Ruijie Lu, and Gang Zeng. Superpoint gaussian splatting for real-time high-fidelity dynamic scene reconstruction, 2024

  33. [41]

    Faster and better 3d splatting via group training, 2025

    Chengbo Wang, Guozheng Ma, Yifei Xue, and Yizhen Lao. Faster and better 3d splatting via group training, 2025. 11

  34. [42]

    Omegas: Object mesh extraction from large scenes guided by gaussian segmentation, 2024

    Lizhi Wang, Feng Zhou, Bo yu, Pu Cao, and Jianqin Yin. Omegas: Object mesh extraction from large scenes guided by gaussian segmentation, 2024

  35. [43]

    Sg-splatting: Accelerating 3d gaussian splatting with spherical gaussians, 2024

    Yiwen Wang, Siyuan Chen, and Ran Yi. Sg-splatting: Accelerating 3d gaussian splatting with spherical gaussians, 2024

  36. [44]

    Bovik, H.R

    Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600–612, 2004

  37. [45]

    4d gaussian splatting for real-time dynamic scene rendering, 2024

    Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering, 2024

  38. [46]

    Street gaussians: Modeling dynamic urban scenes with gaussian splatting, 2024

    Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians: Modeling dynamic urban scenes with gaussian splatting, 2024

  39. [47]

    Nerf-ds: Neural radiance fields for dynamic specular objects, 2023

    Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects, 2023

  40. [48]

    Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction, 2023

    Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction, 2023

  41. [49]

    Gaussian grouping: Segment and edit anything in 3d scenes, 2024

    Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes, 2024

  42. [50]

    Garfield++: Reinforced gaussian radiance fields for large-scale 3d scene reconstruction, 2024

    Hanyue Zhang, Zhiliu Yang, Xinhe Zuo, Yuxin Tong, Ying Long, and Chen Liu. Garfield++: Reinforced gaussian radiance fields for large-scale 3d scene reconstruction, 2024

  43. [51]

    Dynamic 3d gaussian tracking for graph-based neural dynamics modeling, 2024

    Mingtong Zhang, Kaifeng Zhang, and Yunzhu Li. Dynamic 3d gaussian tracking for graph-based neural dynamics modeling, 2024

  44. [52]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018

  45. [53]

    Motiongs: Exploring explicit motion guidance for deformable 3d gaussian splatting, 2024

    Ruijie Zhu, Yanzhe Liang, Hanzhi Chang, Jiacheng Deng, Jiahao Lu, Wenfei Yang, Tianzhu Zhang, and Yongdong Zhang. Motiongs: Exploring explicit motion guidance for deformable 3d gaussian splatting, 2024. 12 A Technical Appendices and Supplementary Material In this supplementary...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.