Pith. sign in

REVIEW 3 major objections 4 minor 51 references

CTRL-GS: Cascaded Temporal Residue Learning for 4D Gaussian Splatting

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that decomposing time into video-constant, segment-constant, and frame-residual layers makes 4D Gaussian splatting more accurate on dynamic scenes with large motion, occlusions, and fine details, while retaining…

desk verdict A clean, modest 4D-GS extension whose gains hinge on flow-chosen temporal windows; the evidence is thinner than the abstract claims, but the idea is coherent and worth refereeing. read the letter →

arxiv 2505.18306 v2 pith:H7SC2RYA submitted 2025-05-23 cs.CV

Karly Hou , Wanhua Li , Hanspeter Pfister This is my paper
classification cs.CV
keywords 4DGaussianSplattingdynamicscenereconstructiontemporalresiduallearningopticalflowwindowingnovelviewsynthesisreal-timerenderingneuralradiancefieldsvideo-segment-framedecomposition
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

CTRL-GS tries to establish that dynamic scenes render more accurately when the deformation of each 3D Gaussian is split into three temporal layers: a video-constant base, a segment-constant offset learned for a short window of frames, and a per-frame residual. The windows themselves are chosen adaptively from optical flow peaks rather than by equal time division. On HyperNeRF and D-NeRF benchmarks the method reports the highest PSNR and SSIM among the compared Gaussian-splatting and NeRF baselines, with the largest gains on scenes that move fast, occlude themselves, or contain fine details. A sympathetic reading is that a residual-in-time formulation gives the model an easier learning target than direct per-frame prediction, because most of a scene's appearance is stable across the video.

What carries the argument

The load-bearing mechanism is the cascaded temporal residual decomposition, formalized in Eq. (8), together with the temporal-window construction that feeds it. The three window schemes (equal division, division at the N frame pairs with highest optical flow, and greedy flow-threshold windows) determine which frames share a segment-level MLP. The quantization coefficient $q$ interpolates between segment-level and frame-level predictions, balancing temporal continuity against flexibility. What this machinery does is give the network a hierarchical target: predict the stable part once, the mid-rate part per window, and only the fast part per frame.

What would settle it

Train CTRL-GS on a dynamic scene captured by multiple cameras where the pretrained optical flow from the window-selection camera disagrees with flow from another camera, then compare reconstruction quality per viewpoint. If windows chosen from one view's flow systematically degrade other views relative to equal windows, the claim that flow-derived windows improve scene-wide learning is falsified. Alternatively, a synthetic scene where flow is large but motion is globally rigid (so high flow does not mark a change in visual setting) should show no benefit from adaptive windows.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is Eq. (8): the deformed Gaussian features are computed as $(X + \Delta_t X + \Delta X, r + \Delta_t r + \Delta r, s + \Delta_t s + \Delta s)$, where $\Delta_t X$, $\Delta_t r$, $\Delta_t s$ come from segment-specific MLPs applied to timestamp-quantized features and $\Delta X$, $\Delta r$, $\Delta s$ come from the original per-frame deformation decoder. This turns time modeling into a sum of video-constant values, segment-constant means, and frame-specific residuals. The paper further claims that greedy threshold-based optical-flow windowing, which creates short windows during high-motion periods and long windows during quiet periods, outperforms equal windows and N-highest-flow windows across the tested datasets. The result is state-of-the-art reconstruction accuracy among compared methods at real-time rendering speeds, with the biggest improvements on scenes where existing 4D-GS degrades most.

Load-bearing premise

The method assumes that optical flow computed from one camera view reliably indicates where the whole scene's motion complexity peaks, so that window boundaries chosen from that flow help all viewpoints; if the flow is view-specific, noisy, or misses occlusions, the windows are misaligned and the residual decomposition loses its benefit.

Editorial extensions

If this is right

  • If correct, CTRL-GS gives real-time 4D reconstruction that handles large movements and occlusions better than the 4D-GS baseline, with the largest gains exactly where prior methods blur or break apart.
  • Because the decomposition starts from video-constant features, the model preserves temporal continuity even when per-frame residuals are noisy.
  • The greedy threshold windowing provides a principled way to spend model capacity: more temporal segments where motion is fast, fewer where it is slow.
  • The equal-window variant already improves over 4D-GS on several scenes, suggesting the residual formulation itself—not just the flow-based windows—carries much of the benefit.
  • Rendering remains interactive (tens of FPS) despite the added segment-level MLPs, and the reported slowdown is attributed to per-frame segment lookup that can be cached.

Reading between the lines

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

  • The video-segment-frame residual idea is a temporal analogue of multi-scale pyramids: it may generalize to any continuously parameterized rendering primitive (e.g., meshes or grids), not just Gaussians, and to any 4D representation that currently predicts deformation directly.
  • The optical-flow windowing could be replaced by a learned motion saliency predictor, making the method end-to-end trainable and removing the dependence on a pretrained flow model and a chosen reference camera.
  • An implied testable extension is to make the number of windows and their boundaries per-scene adaptive, selected by validation performance, rather than fixed by a hyperparameter sweep over N.
  • The per-frame segment lookup slowdown is acknowledged as implementation overhead; if cached during training, the method's FPS would likely match 4D-GS more closely, which the paper states as a planned fix.
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

3 major / 4 minor

Summary. The paper proposes CTRL-GS, an extension of 4D Gaussian Splatting for dynamic scenes. It partitions the temporal domain into windows using either equal divisions, N-highest-flow divisions, or greedy threshold-based divisions derived from RAFT optical flow, and then models per-Gaussian deformation as the sum of video-constant features, segment-specific mean offsets produced by three segment MLPs, and the original point-in-time 4D-GS deformations (Eq. (8)). The method is evaluated on HyperNeRF vrig and D-NeRF datasets, with ablations over the three window construction methods and per-scene tables. The paper claims state-of-the-art visual quality and real-time rendering, with the largest gains on scenes with large motions, occlusions, and fine details.

Significance. If the claims are validated, the cascaded video-segment-frame residual decomposition is a simple and potentially useful design for dynamic Gaussian splatting: it provides temporally local capacity while retaining global stability. The paper is honest enough to include three window-construction variants, per-scene numbers, and a limitations section that acknowledges that gains concentrate on high-motion scenes. The method is clearly specified and the ablation structure lets a reader see exactly where the improvement comes from. However, the evidence as presented does not fully support the state-of-the-art claim: the aggregate gains over 4D-GS are small, the equal-window variant is worse than 4D-GS on D-NeRF, per-scene results are mixed, and important hyperparameters and the optical-flow camera choice are not controlled or reported with uncertainty.

major comments (3)
  1. [Sec. 3.5, 3.6 and Table 2] The state-of-the-art claim is not fully supported by the reported numbers. On D-NeRF (Table 2), the Equal Windows variant achieves 32.87 PSNR, well below the 4D-GS baseline of 34.05; the Dynamic Threshold variant achieves 34.34, only +0.29 dB over 4D-GS. Per-scene (Table 4), CTRL-GS is worse than 4D-GS on T-Rex (33.94 vs 34.23 PSNR), and worse on LPIPS on Lego (0.0611 vs 0.0382) and Jumpingjacks (0.0182 vs 0.0128). The statement in Sec. 4.5 that CTRL-GS shows 'improved accuracy across scenes' should be replaced with a more precise characterization, and the reported aggregate numbers should be accompanied by error bars or multiple-run statistics before claiming state-of-the-art performance.
  2. [Sec. 3.5] The optical-flow-based temporal windows are load-bearing, but the choice of flow camera is neither specified nor ablated. The paper states that a pretrained RAFT model is applied to 'one camera perspective' for HyperNeRF scenes, yet the HyperNeRF vrig sequences are captured with two cameras. If the flow is computed from the validation camera, window boundaries are selected using validation frames; if it is computed from the other camera, the proxy may miss occlusions and parallax visible in the validation view. Since Table 2 shows that the residual decomposition alone (Equal Windows) is worse than 4D-GS on D-NeRF, the claimed gains rest entirely on the flow-derived window placement. Please add experiments that vary the flow source, report whether the resulting window boundaries change, and compare against a view-independent motion signal such as average flow over all available training views.
  3. [Supplementary] The hyperparameter protocol makes the comparison to baselines difficult to interpret. The supplementary material reports a sweep over N ∈ {2,...,9} and q ∈ {0.0,0.1,0.2,0.3,0.5,0.7,0.9}, 'reporting best results,' and Sec. 4.1 says the same for all models. Because baseline numbers are taken from prior papers or code with their own settings, selecting per-scene best N and q for CTRL-GS while comparing against fixed published baselines is not an apples-to-apples comparison. At minimum, the authors should report mean and standard deviation over at least three runs, disclose the chosen N and q per scene or per dataset, and provide a sensitivity analysis showing how PSNR varies over the sweep.
minor comments (4)
  1. [Sec. 2.1] The sentence containing 'dchoose 2 planes' is garbled and should read 'choose 2 planes.'
  2. [Table 1] The FPS column header uses a downward arrow 'FPS↓', which implies lower is better, but higher FPS is better; use 'FPS↑' as in Table 2.
  3. [Sec. 3.4] The role of q is not fully explained. In Eq. (6), q appears to add a constant shift q*I to the quantized timestamp, but the text describes it as 'adjusting the interpolation between segment-level predictions and frame-specific residuals.' Clarify the precise effect of q and why the tested range stops at 0.9.
  4. [References] Reference [45] lists 'Wang Xinggang' in the author list; the name should be formatted as 'Xinggang Wang' for consistency with the other references.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Eq. (8) is a model definition, and the reported gains rest on external benchmarks.

full rationale

The paper's core derivation is the cascaded decomposition in Eq. (8), where deformed Gaussians are defined as X' = X + ΔtX + ΔX (and analogously for r and s). This is an architectural definition, not a derived identity: ΔX is the same point-in-time deformation used by 4D-GS [45], while ΔtX is a new segment-mean output of temporally-local MLPs. The 'frame-specific residual' is therefore not constructed by subtracting the segment mean from a target and then predicting that leftover; it is an independently learned term. Nothing in the paper reduces a prediction to its input by construction. The temporal windows are obtained from a pretrained RAFT model on one camera view (Sec. 3.5), which is an external, fixed source of motion estimates rather than a fitted parameter of the reconstruction. The hyperparameter sweep over N and q in the supplementary is a model-selection procedure; reporting the best validation result may raise benchmarking-fairness questions, but it does not rename a fitted parameter as a prediction. The ablation in Table 2 shows the equal-window variant is worse than 4D-GS on D-NeRF (32.87 vs 34.05 PSNR), demonstrating that the residual decomposition is not tautologically beneficial. Self-citations ([18], [34], [51]) appear only in related work and preliminaries and are not load-bearing for the claimed novelty. The single-camera optical-flow choice is an empirical robustness and potential-leakage concern, not circularity, because the windows are not derived from the validation frames' reconstruction loss and the evaluation is against external held-out data. Overall, the central claim has independent empirical content.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central result rests on two domain assumptions: that dynamic scenes have hierarchical temporal structure, and that single-camera optical flow reliably places temporal windows. The free parameters N and q are chosen by validation sweep, so part of the reported gain is fitted rather than predicted. No invented entities are postulated.

free parameters (2)
  • N (number of temporal windows) = swept over {2,3,4,5,6,7,8,9}, best per scene reported
    Controls temporal segmentation granularity; chosen by validation sweep for each window construction method, so part of the reported gain is fitted.
  • q (quantization coefficient) = swept over {0.0, 0.1, 0.2, 0.3, 0.5, 0.7, 0.9}
    Interpolation weight between segment-level predictions and frame-specific residuals for the equal-window method; selected by validation sweep.
assumptions (3)
  • domain assumption Dynamic scenes exhibit hierarchical temporal structure: video-level stability, segment-level motion, frame-level fluctuations.
    Motivates the video-segment-frame decomposition in Sec. 3.3; if false, the decomposition adds complexity without benefit.
  • domain assumption Optical flow from one camera view reliably identifies motion peaks that divide scenes into visual settings.
    Sec. 3.5 uses RAFT flow from one camera to place window boundaries; if flow is view-specific or noisy, windows are suboptimally placed.
  • domain assumption Residual parameterizations are easier to optimize than direct predictions (He et al. [11]).
    Borrowed from image recognition as an empirical heuristic; the paper assumes it transfers to temporal deformation fields.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CTRL-GS: Cascaded Temporal Residue Learning for 4D Gaussian Splatting." pith.science (2026). https://pith.science/paper/H7SC2RYA

@misc{pith2026250518306,
  author       = {Pith},
  title        = {Pith review of: CTRL-GS: Cascaded Temporal Residue Learning for 4D Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H7SC2RYA}},
  note         = {Machine review of arXiv:2505.18306}
}
read the original abstract

Recently, Gaussian Splatting methods have emerged as a desirable substitute for prior Radiance Field methods for novel-view synthesis of scenes captured with multi-view images or videos. In this work, we propose a novel extension to 4D Gaussian Splatting for dynamic scenes. Drawing on ideas from residual learning, we hierarchically decompose the dynamic scene into a "video-segment-frame" structure, with segments dynamically adjusted by optical flow. Then, instead of directly predicting the time-dependent signals, we model the signal as the sum of video-constant values, segment-constant values, and frame-specific residuals, as inspired by the success of residual learning. This approach allows more flexible models that adapt to highly variable scenes. We demonstrate state-of-the-art visual quality and real-time rendering on several established datasets, with the greatest improvements on complex scenes with large movements, occlusions, and fine details, where current methods degrade most.

Figures

Figures reproduced from arXiv: 2505.18306 by the authors.

Figure 1
Figure 1. Our method achieves improved reconstruction accuracy for dynamic scenes at high image resolutions across settings. Current [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Two key components of our method. a) We divide the frames of the dynamic scene into temporal windows. We assess three [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Comparisons between ground truth (GT), baseline 4D-GS [ [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparisons between ground truth (GT), 4D-GS, and our method with dynamic threshold-based windows on D-NeRF [ [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Magnitude of optical flow fij over each of the HyperNeRF [30] datasets a) BROOM, b) BANANA, c) CHICKEN, 3DPRINTER. Red vertical lines denote optimal temporal window segmentation when constructed with the dynamic thresholds method [PITH_FULL_IMAGE:figures/full_fig_p006…
Figure 6
Figure 6. Figure 6: Additional frame comparisons for the HyperNeRF [ [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 36 canonical work pages

  1. [1]

    Gaussian shell maps for efficient 3d human generation

    Rameen Abdal, Wang Yifan, Zifan Shi, Yinghao Xu, Ryan Po, Zhengfei Kuang, Qifeng Chen, Dit-Yan Yeung, and Gor- don Wetzstein. Gaussian shell maps for efficient 3d human generation. InCVPR, 2024. 3

    work page 2024

  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 neu- ral radiance fields. InConference on Computer Vision and Pattern Recognition (CVPR), 2021. 2

    work page 2021

  3. [3]

    Barron, Ben Mildenhall, Dor Verbin, Pratul P

    Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. InConference on Com- puter Vision and Pattern Recognition (CVPR), 2022

    work page 2022

  4. [4]

    Hexplane: a fast representation for dynamic scenes

    Ang Cao and Justin Johnson. Hexplane: a fast representation for dynamic scenes. InConference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 7

    work page 2023

  5. [5]

    Fast dynamic radiance fields with time-aware neural voxels

    Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xi- aopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. InSIGGRAPH Asia Conference Papers, 2022. 2, 6, 7

    work page 2022

  6. [6]

    Plenoxels: radiance fields without neural networks

    Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: radiance fields without neural networks. InConference on Computer Vision and Pattern Recognition (CVPR), 2022. 2

    work page 2022

  7. [7]

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

    Sara Fridovich-Keil, Giacomo meanti, Frederik Rahbaek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: explicit radiance fields in space, time, and appearance. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 7

    work page 2023

  8. [8]

    Efros, and Xiaolong Wang

    Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splat- ting. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 20796– 20805, 2024. 2

    2024

Show all 51 references
  1. [9]

    V4d: V oxel for 4d novel view synthesis

    Wnahui Gan, Hongbin Xu, Yi Huang, Shifeng Chen, and Naoto Yokoya. V4d: V oxel for 4d novel view synthesis. In IEEE Transactions on Visualization and Computer Graphics,

  2. [10]

    X. Guo, J. Sun, Y . Dai, G. Chen, X. Ye, X. Tan, E. Ding, Y . Zhang, and J. Wang. Forward flow for novel view syn- thesis of dynamic scenes. InInternational Conference on Computer Vision (ICCV), 2023. 2, 6, 7

    2023

  3. [11]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InCVPR,

  4. [12]

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

    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. InCVPR,

  5. [13]

    Splatam: Splat, track & map 3d gaussians for dense rgb-d slam

    Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, 2024. 3

    2024

  6. [14]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering.ACM Transactions on Graphics, 42(4), 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering.ACM Transactions on Graphics, 42(4), 2023. 1, 2, 6, 7

    2023

  7. [15]

    Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splatting

    Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splatting. InECCV, 2024. 2, 7

    2024

  8. [16]

    Laur and P

    D. Laur and P. Hanrahan. Hierarchical splatting: a progres- sive refinement algorithm for volume rendering. InCom- puter Graphics SIGGRAPH, 1991. 1

    1991

  9. [17]

    Hu- man101: Training 100+fps human gaussians in 100s from 1 view, 2023

    Mingwei Li, Jiachen Tao, Zongxin Yang, and Yi Yang. Hu- man101: Training 100+fps human gaussians in 100s from 1 view, 2023. 3

    2023

  10. [18]

    4d langsplat: 4d language gaussian splatting via multimodal large language models.arXiv preprint arXiv:2503.10437, 2025

    Wanhua Li, Renping Zhou, Jiawei Zhou, Yingwei Song, Jo- hannes Herter, Minghan Qin, Gao Huang, and Hanspeter Pfister. 4d langsplat: 4d language gaussian splatting via multimodal large language models.arXiv preprint arXiv:2503.10437, 2025. 3

    2025 arXiv

  11. [19]

    Neural scene flow fields for space-time view synthesis of dy- namic scenes

    Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dy- namic scenes. InConference on Computer Vision and Pat- tern Recognition (CVPR), 2021. 2

    2021

  12. [20]

    Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis

    Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen- Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. InProc. IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025. 2

    2025

  13. [21]

    Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle

    Youtian Lin, Zuozhuo Dai, Siyu Zhu, and Yao Yao. Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. InCVPR, 2024. 7

    2024

  14. [22]

    Lippert and M.H

    L. Lippert and M.H. Gross. Fast wavelet based volume ren- dering by accumulation of transparent texture maps.Com- puter Graphics Forum, 14(3), 1995. 1

    1995

  15. [23]

    Humangaus- sian: Text-driven 3d human generation with gaussian splat- ting.arXiv preprint arXiv:2311.17061, 2023

    Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangaus- sian: Text-driven 3d human generation with gaussian splat- ting.arXiv preprint arXiv:2311.17061, 2023. 3

    2023 arXiv

  16. [24]

    Robust dynamic radiance fields

    Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Jo- hannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. InConference on Computer Vision and Pattern Recog- nition (CVPR), 2023. 2

    2023

  17. [25]

    Neural vol- umes: learning dynamic renderable volumes from images

    Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural vol- umes: learning dynamic renderable volumes from images. ACM Transactions on Graphics, 38(4), 2019. 2

    2019

  18. [26]

    Srinivasan, Matthew Tancik, Monathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Monathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: representing scenes as neural radiance fields for view synthe- sis. InEuropean Conference on Computer Vision (ECCV),

  19. [27]

    Fast perspective volume ren- dering with splatting by utilizing a ray-driven approach

    Klaus Mueller and Roni Yagel. Fast perspective volume ren- dering with splatting by utilizing a ray-driven approach. In IEEE Visualization, 1996. 1

    1996

  20. [28]

    Humansplat: Generalizable single-image human gaus- sian splatting with structure priors

    Panwang Pan, Zhuo Su, Chenguo Lin, Zhen Fan, Yongjie Zhang, Zeming Li, Tingting Shen, Yadong Mu, and Yebin Liu. Humansplat: Generalizable single-image human gaus- sian splatting with structure priors. InAdvances in Neural Information Processing Systems (NeurIPS), 2024. 3

    2024

  21. [29]

    K. Park, U. Sinha, J. T. Barron, S. Bouaziz, D. B. Goldman, S. M. Seitz, and R. Martin-Brualla. Nerfies: deformable neu- ral radiance fields. InInternational Conference on Computer Vision (ICCV), 2021. 1, 6, 7

    2021

  22. [30]

    Hypernerf: A higher- dimensional representation for topologically varying neural radiance fields.ACM Transactions on Graphics, 40(6), 2021

    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.ACM Transactions on Graphics, 40(6), 2021...

    2021

  23. [31]

    Gags: Granularity-aware 3d feature distillation for gaussian splatting.arXiv preprint arXiv:2412.13654, 2024

    Yuning Peng, Haiping Wang, Yuan Liu, Chenglu Wen, Zhen Dong, and Bisheng Yang. Gags: Granularity-aware 3d feature distillation for gaussian splatting.arXiv preprint arXiv:2412.13654, 2024. 3

    2024 arXiv

  24. [32]

    D-nerf: neural radiance fields for dynamic scenes

    Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: neural radiance fields for dynamic scenes. InCVPR, 2021. 6, 7, 1

    2021

  25. [33]

    3dgs-avatar: Animatable avatars via deformable 3d gaussian splatting

    Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs-avatar: Animatable avatars via deformable 3d gaussian splatting. InCVPR, 2024. 3

    2024

  26. [34]

    Langsplat: 3d language gaussian splatting

    Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051–20060, 2024. 3

    2024

  27. [35]

    Dreamgaussian4d: Genera- tive 4d gaussian splatting.arXiv preprint arXiv:2312.17142,

    Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. Dreamgaussian4d: Genera- tive 4d gaussian splatting.arXiv preprint arXiv:2312.17142,

    arXiv

  28. [36]

    Relightable gaussian codec avatars

    Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable gaussian codec avatars. In CVPR, 2024. 3

    2024

  29. [37]

    Structure-from-motion revisited

    Johannes Lutz Sch ¨onberger and Jan-Michael Frahm. Structure-from-motion revisited. InConference on Com- puter Vision and Pattern Recognition (CVPR), 2016. 3, 5

    2016

  30. [38]

    Swings: Sliding windows for dynamic 3d gaussian splatting.arXiv preprint arXiv:2312.13308, 2023

    Richard Shaw, Michal Nazarczuk, Jifei Song, Arthur Moreau, Sibi Catley-Chandar, Helisa Dhamo, and Eduardo Perez-Pellitero. Swings: Sliding windows for dynamic 3d gaussian splatting.arXiv preprint arXiv:2312.13308, 2023. 3

    2023 arXiv

  31. [39]

    Raft: Recurrent all-pairs field transforms for optical flow

    Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. InEuropean Conference on Computer Vision, 2020. 4

    2020

  32. [40]

    Superpoint gaus- sian splatting for real-time high-fidelity dynamic scene re- construction

    Diwen Wan, Ruijie Lu, and Gang Zeng. Superpoint gaus- sian splatting for real-time high-fidelity dynamic scene re- construction. InICML, 2024. 7

    2024

  33. [41]

    Masked space-time hash encoding for efficient dynamic scene reconstruction

    Feng Wang, Zilong Chen, Guokang Wang, Yafei Song, and Huaping Liu. Masked space-time hash encoding for efficient dynamic scene reconstruction. InNeurIPS, 2023. 6, 7

    2023

  34. [42]

    Sg-splatting: accel- erating 3d gaussian splatting with spherical gaussians.arXiv preprint arXiv:2501.00342, 2024

    Yiwen Wang, Siyuan Chen, and Ran Yi. Sg-splatting: accel- erating 3d gaussian splatting with spherical gaussians.arXiv preprint arXiv:2501.00342, 2024. 3

    2024 arXiv

  35. [43]

    Image quality assessment: from error visibil- ity to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibil- ity to structural similarity. InIEEE Transactions on Image Processing, 2004. 6

    2004

  36. [44]

    Interactive volume rendering

    Lee Westover. Interactive volume rendering. InProceedings of the Chapel Hill workshop on Volume Visualization, 1989. 1

    1989

  37. [45]

    4D Gaussian Splatting for Real-Time Dynamic Scene Ren- dering.arXiv preprint arXiv:2310.08528, 2023

    Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Wang Xinggang. 4D Gaussian Splatting for Real-Time Dynamic Scene Ren- dering.arXiv preprint arXiv:2310.08528, 2023. 1, 3, 4, 5, 6, 7

    2023 arXiv

  38. [46]

    Physgaussian: Physics- integrated 3d gaussians for generative dynamics.arXiv preprint arXiv:2311.12198, 2023

    Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics- integrated 3d gaussians for generative dynamics.arXiv preprint arXiv:2311.12198, 2023. 2

    2023 arXiv

  39. [47]

    Reflective gaussian splatting.arXiv preprint arXiv:2412.19282, 2024

    Yuxuan Yao, Zixuan Zeng, Chun Gu, Xiatian Zhu, and Li Zhang. Reflective gaussian splatting.arXiv preprint arXiv:2412.19282, 2024. 3

    2024 arXiv

  40. [48]

    Nerf++: Analyzing and improving neural radiance fields.arXiv preprint arXiv:2010.07492v2, 2020

    Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields.arXiv preprint arXiv:2010.07492v2, 2020. 2

    2010 arXiv

  41. [49]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InConference on Computer Vision and Pattern Recognition (CVPR), 2018. 6

    2018

  42. [50]

    Hugs: Holistic urban 3d scene understanding via gaus- sian splatting

    Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaus- sian splatting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), ...

    2024

  43. [51]

    Ewa volume splatting

    Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. InIEEE Proceedings Visualization, 2001. 3 CTRL-GS: Cascaded Temporal Residue Learning for 4D Gaussian Splatting Supplementary Material Figure 6. Additional frame comparisons for the H...

    2001

Pith tools

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