Incoherent Deformation, Not Capacity: Diagnosing and Mitigating Overfitting in Dynamic Gaussian Splatting

Ahmad Droby

arxiv: 2604.16747 · v2 · submitted 2026-04-17 · 💻 cs.CV

Incoherent Deformation, Not Capacity: Diagnosing and Mitigating Overfitting in Dynamic Gaussian Splatting

Ahmad Droby This is my paper

Pith reviewed 2026-05-10 08:01 UTC · model grok-4.3

classification 💻 cs.CV

keywords dynamic 3D Gaussian Splattingoverfittingdeformation regularizationneural rendering3D reconstructionmonocular videoPSNR gap

0 comments

The pith

Overfitting in dynamic 3D Gaussian Splatting stems from incoherent per-Gaussian deformations rather than excess capacity.

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

The paper examines the large train-test PSNR gap in dynamic 3D Gaussian Splatting on monocular video benchmarks. Ablations show that Gaussian splitting accounts for most of the gap and correlates with higher primitive counts, which initially points to a capacity problem. Introducing Elastic Energy Regularization, a local smoothness penalty on the deformation field, cuts the gap by 40.8 percent while increasing the Gaussian count by 85 percent, and direct strain measurements confirm sharply lower per-Gaussian incoherence. Further regularizers and a soft growth cap push the gap reduction to 57 percent. The same coherence penalty transfers to another deformation architecture and to real monocular video scenes with little quality loss.

Core claim

The overfitting in dynamic 3DGS is driven by incoherent deformation, not parameter count. Splitting explains over 80 percent of the gap and produces a near-perfect log-linear relation between Gaussian count and gap size, yet a local-smoothness penalty on the per-Gaussian deformation field reduces the gap by 40.8 percent while growing the cloud by 85 percent; measured strain drops by 99.72 percent on average, and the approach generalizes across architectures and real data.

What carries the argument

Elastic Energy Regularization (EER), a penalty that measures and reduces local strain in the per-Gaussian deformation field to enforce coherence.

If this is right

Disabling splitting collapses both the gap and the Gaussian count, confirming the capacity correlation.
EER plus GAD and PTDrop together close 57 percent of the gap.
The coherence benefit appears in a second deformation architecture and on real monocular video.
Per-Gaussian strain can be read directly from checkpoints to diagnose the source of overfitting.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Coherence penalties may be useful in other per-point or per-primitive dynamic reconstruction methods.
The finding suggests that deformation incoherence, not raw parameter count, could be the dominant overfitting driver in broader classes of dynamic scene models.
Enforcing coherence while allowing growth might let future dynamic splatting systems scale to more complex motion without retraining on denser views.

Load-bearing premise

That the measured PSNR gaps on D-NeRF and HyperNeRF directly reflect overfitting from incoherent deformation rather than differences in view sampling or lighting.

What would settle it

Train a high-capacity dynamic 3DGS model whose deformation field is forced to remain locally coherent and check whether the train-test PSNR gap stays small despite the large primitive count.

Figures

Figures reproduced from arXiv: 2604.16747 by Ahmad Droby.

**Figure 2.** Figure 2: Densification is front-loaded: 84–89% of cloud growth happens before iter 7,500. The red dashed line is [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Count–gap relation. Ablations (gray) are log-linear, which would suggest a capacity story. EER (green) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: Deformation field on Lego. Left: canonical cloud colored by per-Gaussian displacement magnitude (baseline above, EER below). Middle: subsampled quiver of u(x, t= 0.5). Right: distribution of per-Gaussian k-NN strain. Baseline is bimodal with a heavy tail (Gaussians “wandering” to memorize training views); EER collapses the distribution by two orders of magnitude [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 5.** Figure 5: GAD λ sweep. Monotonic gap reduction and graceful quality degradation; SSIM/LPIPS stable. cloud is large, each new Gaussian’s marginal contribution to the reconstruction is small, and is more likely to encode training-view-specific residuals than scene structure—so the threshold should rise with K. (ii) When training loss has plateaued (small ∆ℓema), the residual gradients that trigger ADC reflect noise ra… view at source ↗

**Figure 6.** Figure 6: Primary result: GAD+EER. GAD alone reduces the gap by 11.3%, EER alone by 40.8%. Combined, GAD+EER reaches 48.2% gap reduction at 38K Gaussians with a −0.86 dB test-PSNR cost. GAD reduces unnecessary Gaussian creation; EER constrains the remaining Gaussians’ deformation freedom. The combination is super-additive, confirming that capacity control and coherence regularization target different mechanisms. W… view at source ↗

**Figure 6.** Figure 6: Quality–gap Pareto frontier: baseline, GAD, EER, PTDrop, and their combinations. GAD+EER [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗

read the original abstract

Dynamic 3D Gaussian Splatting methods achieve strong training-view PSNR on monocular video but generalize poorly: on the D-NeRF benchmark we measure an average train-test PSNR gap of 6.18 dB, rising to 11 dB on individual scenes. We report two findings that together account for most of that gap. Finding 1 (the role of splitting). A systematic ablation of the Adaptive Density Control pipeline (split, clone, prune, frequency, threshold, schedule) shows that splitting is responsible for over 80% of the gap: disabling split collapses the cloud from 44K to 3K Gaussians and the gap from 6.18 dB to 1.15 dB. Across all threshold-varying ablations, gap is log-linear in count (r = 0.995, bootstrap 95% CI [0.99, 1.00]), which suggests a capacity-based explanation. Finding 2 (the role of deformation coherence). We show that the capacity explanation is incomplete. A local-smoothness penalty on the per-Gaussian deformation field -- Elastic Energy Regularization (EER) -- reduces the gap by 40.8% while growing the cloud by 85%. Measuring per-Gaussian strain directly on trained checkpoints, EER reduces mean strain by 99.72% (median 99.80%) across all 8 scenes; on 8/8 scenes the median Gaussian under EER is less strained than the 1st-percentile (best-behaved) Gaussian under baseline. Alongside EER, we evaluate two further regularizers: GAD, a loss-rate-aware densification threshold, and PTDrop, a jitter-weighted Gaussian dropout. GAD+EER reduces the gap by 48%; adding PTDrop and a soft growth cap reaches 57%. We confirm that coherence generalizes to (a) a different deformation architecture (Deformable-3DGS, +40.6% gap reduction at re-tuned lambda), and (b) real monocular video (4 HyperNeRF scenes, reducing the mean PSNR gap by 14.9% at the same lambda as D-NeRF, with near-zero quality cost). The overfitting in dynamic 3DGS is driven by incoherent deformation, not parameter count.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The main claim holds: the train-test gap in dynamic 3DGS is mostly from incoherent per-Gaussian deformation, not raw capacity, and the strain measurements plus EER ablation make that case cleanly.

read the letter

The paper's core finding is that splitting in the density control step drives most of the 6 dB average gap on D-NeRF, but the real problem is that the extra Gaussians end up deforming incoherently. They back this with two pieces: first, an ablation that turns off splitting and drops both Gaussian count and the gap sharply, with a tight log-linear correlation across thresholds. Second, they add Elastic Energy Regularization, which penalizes local strain in the deformation field. This lets the model grow to 85% more Gaussians while cutting the gap by 40%, and direct strain checks show a 99.7% drop in mean strain. The same pattern appears on Deformable-3DGS and on real HyperNeRF scenes at the same lambda. They also test two lighter regularizers (GAD and PTDrop) that push the total gap reduction to 57% when combined with a soft cap. That combination of controlled ablations, direct strain metric, and cross-architecture check is the new part; prior dynamic 3DGS work mostly treated the gap as a capacity or optimization issue without measuring deformation coherence this way. The experiments look solid on the numbers given, with consistent patterns across eight scenes. The main soft spot is that the benchmarks are still synthetic or limited real video, so some of the remaining gap could trace to view sampling or lighting ambiguities rather than deformation alone. The regularizers also introduce a couple of new hyperparameters that need scene-specific tuning, and the paper does not yet show whether the coherence fix holds on longer sequences or more complex motion. Overall this is useful for anyone building or debugging dynamic Gaussian methods. It deserves a serious referee because the diagnosis is falsifiable, the measurements are direct, and the proposed fixes are cheap to add. I would bring it to a reading group and would cite the strain analysis if I were working on similar regularization.

Referee Report

1 major / 2 minor

Summary. The manuscript claims that overfitting in dynamic 3D Gaussian Splatting on monocular video (average 6.18 dB train-test PSNR gap on D-NeRF, up to 11 dB per scene) is driven by incoherent per-Gaussian deformations rather than model capacity. Systematic ablations of Adaptive Density Control show splitting accounts for >80% of the gap, with log-linear correlation (r=0.995) between Gaussian count and gap across thresholds. Elastic Energy Regularization (EER) reduces the gap 40.8% while increasing count 85% and mean strain 99.72%; combined with GAD (loss-rate-aware densification) and PTDrop (jitter-weighted dropout) the reduction reaches 57%. Results generalize to Deformable-3DGS (+40.6%) and 4 HyperNeRF scenes (14.9% gap reduction at fixed lambda).

Significance. If the empirical patterns hold, the work offers a useful diagnosis and practical mitigation for generalization failures in dynamic 3DGS by shifting attention from parameter count to deformation coherence. Strengths include consistent ablation results across 8 scenes and two architectures, direct per-Gaussian strain quantification, bootstrap confidence intervals on the correlation, and zero-cost generalization to real video. The regularizers (EER, GAD, PTDrop) provide immediately usable tools that increase capacity while improving test performance, which could influence future dynamic reconstruction pipelines.

major comments (1)

[Finding 2] Finding 2: the claim that EER enforces coherence and thereby reduces the gap while growing the cloud by 85% is load-bearing. The manuscript must supply the exact mathematical definition of the local-smoothness penalty (including neighborhood selection and strain tensor computation) and state whether the regularizer is active only at training time or also at inference; without this the 99.72% strain reduction cannot be reproduced or verified independently.

minor comments (2)

[Finding 1] The abstract and Finding 1 report the 80% attribution to splitting from a single ablation (disabling split). A supplementary table showing the incremental effect of disabling each ADC component individually (split, clone, prune, frequency, threshold) would make the dominance claim easier to assess.
Implementation details for the hyper-parameter search of lambda (EER weight) and the GAD loss-rate threshold are referenced but not fully specified (grid ranges, whether tuning was global or per-scene, number of trials). Adding these in the supplementary material would address reproducibility concerns.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the positive evaluation and the recommendation of minor revision. We address the single major comment below and will update the manuscript accordingly to improve clarity and reproducibility.

read point-by-point responses

Referee: [Finding 2] Finding 2: the claim that EER enforces coherence and thereby reduces the gap while growing the cloud by 85% is load-bearing. The manuscript must supply the exact mathematical definition of the local-smoothness penalty (including neighborhood selection and strain tensor computation) and state whether the regularizer is active only at training time or also at inference; without this the 99.72% strain reduction cannot be reproduced or verified independently.

Authors: We agree that the precise formulation of Elastic Energy Regularization (EER) must be stated explicitly for reproducibility. The current manuscript describes EER at a high level but does not provide the full mathematical definition, neighborhood selection details, or strain tensor computation in a self-contained way. In the revised version we will add a dedicated paragraph (or short subsection) giving the exact expression for the local-smoothness penalty, the criterion used to select neighboring Gaussians, and the formula for the per-Gaussian strain tensor. We will also state explicitly that the regularizer is active only during training and is disabled at inference. These additions will allow independent verification of the reported 99.72% mean strain reduction and the associated generalization improvements. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper presents empirical findings from controlled ablations on D-NeRF and HyperNeRF benchmarks, measuring PSNR gaps, Gaussian counts, and strain values directly from trained models. No mathematical derivation chain, self-definitional equations, or fitted parameters renamed as predictions exist. The central claim (incoherent deformation drives overfitting) is supported by observed correlations (e.g., r=0.995) and intervention effects (EER reducing gap while increasing count), all externally verifiable on public benchmarks without looping back to inputs by construction. Hyperparameter choices for regularizers are standard tuning and do not define the result.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 3 invented entities

The work relies on standard 3D Gaussian Splatting assumptions and introduces new regularization terms whose weights are tuned; no new physical entities are postulated.

free parameters (2)

EER regularization weight (lambda)
Tuned to balance gap reduction against training-view quality.
GAD loss-rate threshold
Adjusted per-scene to control densification timing.

axioms (2)

domain assumption PSNR difference between train and test views measures overfitting in novel-view synthesis
Standard evaluation practice in the field invoked throughout the findings.
domain assumption Gaussian deformation fields can be regularized independently per point without breaking scene consistency
Underlying premise for Elastic Energy Regularization.

invented entities (3)

Elastic Energy Regularization (EER) no independent evidence
purpose: Penalize incoherent local deformations in the per-Gaussian deformation field
New loss term introduced to enforce coherence.
GAD (loss-rate-aware densification) no independent evidence
purpose: Dynamically adjust splitting threshold based on loss rate
New densification heuristic.
PTDrop (jitter-weighted Gaussian dropout) no independent evidence
purpose: Drop Gaussians with high deformation jitter
New dropout mechanism.

pith-pipeline@v0.9.0 · 5739 in / 1485 out tokens · 48748 ms · 2026-05-10T08:01:50.804612+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages

[1]

Per-Gaussian embedding-based deformation for deformable 3D Gaussian splatting

Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh. Per-Gaussian embedding-based deformation for deformable 3D Gaussian splatting. InEuropean Conference on Com- puter Vision (ECCV), 2024

work page 2024
[2]

Efficient density control for 3d gaussian splat- ting.arXiv preprint arXiv:2411.10133, 2024

Xiaobin Deng, Changyu Diao, Min Li, Ruohan Yu, and Duanqing Xu. Efficient density control for 3D Gaussian splatting.arXiv preprint arXiv:2411.10133, 2024

work page arXiv 2024
[3]

Mini-Splatting: Representing scenes with a constrained number of 11 Gaussians

Guangchi Fang and Bing Wang. Mini-Splatting: Representing scenes with a constrained number of 11 Gaussians. InEuropean Conference on Computer Vision (ECCV), 2024

work page 2024
[4]

Improving adaptive density control for 3D Gaussian splatting

Glenn Grubert, Florian Barthel, Anna Hilsmann, and Peter Eisert. Improving adaptive density control for 3D Gaussian splatting. InInternational Joint Conference on Computer Vision, Imaging and Com- puter Graphics Theory and Applications (VISAPP), pages 610–621. SCITEPRESS, 2025

work page 2025
[5]

PUP 3D-GS: Principled uncertainty pruning for 3D Gaussian splatting

Alex Hanson, Allen Tu, Vasu Singla, Mayuka Jayawardhana, Matthias Zwicker, and Tom Gold- stein. PUP 3D-GS: Principled uncertainty pruning for 3D Gaussian splatting. InIEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2025

work page 2025
[6]

SC-GS: Sparse-controlled Gaussian splatting for editable dy- namic scenes

Yi-Hua Huang, Yang-Tian Sun, Zilong Yang, Xi- aoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. SC-GS: Sparse-controlled Gaussian splatting for editable dy- namic scenes. InIEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2024

work page 2024
[7]

3D Gaussian splatting for real-time radiance field rendering

Bernhard Kerbl, Georgios Kopanas, Thomas Leimk¨ uhler, and George Drettakis. 3D Gaussian splatting for real-time radiance field rendering. In ACM Transactions on Graphics (SIGGRAPH), 2023

work page 2023
[8]

Yiqing Liang, Mikhail Okunev, Mikaela Angelina Uy, Runfeng Li, Leonidas Guibas, James Tompkin, and Adam W. Harley. Monocular dynamic Gaus- sian splatting: Fast, brittle, and scene complexity rules.Transactions on Machine Learning Research (TMLR), 2025

work page 2025
[9]

Taming 3DGS: High-quality radiance fields with limited resources

Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Car- rasco, and Fernando De La Torre. Taming 3DGS: High-quality radiance fields with limited resources. InSIGGRAPH Asia Conference Papers, 2024

work page 2024
[10]

Drop- Gaussian: Structural regularization for sparse-view Gaussian splatting

Hyunwoo Park, Gun Ryu, and Wonjun Kim. Drop- Gaussian: Structural regularization for sparse-view Gaussian splatting. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

work page 2025
[11]

Barron, Sofien Bouaziz, Dan B

Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B. Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. InIEEE/CVF International Conference on Computer Vision (ICCV), 2021

work page 2021
[12]

Barron, Sofien Bouaziz, Dan B

Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B. Gold- man, 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

work page 2021
[13]

D-NeRF: Neural ra- diance fields for dynamic scenes

Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural ra- diance fields for dynamic scenes. InIEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR), 2021

work page 2021
[14]

Revising densification in Gaussian splatting

Samuel Rota Bul` o, Lorenzo Porzi, and Peter Kontschieder. Revising densification in Gaussian splatting. InEuropean Conference on Computer Vision (ECCV), 2024

work page 2024
[15]

Steepest descent density control for compact 3D Gaussian splatting

Peihao Wang, Yuehao Wang, Dilin Wang, Sreyas Mohan, Zhiwen Fan, Lemeng Wu, Ruisi Cai, Yu- Ying Yeh, Zhangyang Wang, Qiang Liu, and Rakesh Ranjan. Steepest descent density control for compact 3D Gaussian splatting. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

work page 2025
[16]

4D Gaussian splatting for real-time dynamic scene rendering

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. InIEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2024

work page 2024
[17]

DropoutGS: Dropping out gaussians for better sparse-view rendering

Yexing Xu, Longguang Wang, Minglin Chen, Sheng Ao, Li Li, and Yulan Guo. DropoutGS: Dropping out gaussians for better sparse-view rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

work page 2025
[18]

Deformable 3D Gaussians for high-fidelity monocular dynamic scene reconstruction

Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3D Gaussians for high-fidelity monocular dynamic scene reconstruction. InIEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2024

work page 2024
[19]

AbsGS: Recovering fine details in 3D Gaussian splatting

Zongxin Ye, Wenyu Li, Sidun Liu, Peng Qiao, and Yong Dou. AbsGS: Recovering fine details in 3D Gaussian splatting. InProceedings of the 32nd ACM International Conference on Multimedia (MM), 2024

work page 2024
[20]

Pixel-GS: Density control with pixel-aware gradient for 3D Gaussian splatting

Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-GS: Density control with pixel-aware gradient for 3D Gaussian splatting. In European Conference on Computer Vision (ECCV), 2024

work page 2024
[21]

Gradient- direction-aware density control for 3D Gaussian splat- ting

Zheng Zhou, Yu-Jie Xiong, Jia-Chen Zhang, Chun- Ming Xia, Xihe Qiu, and Hongjian Zhan. Gradient- direction-aware density control for 3D Gaussian splat- ting. InInternational Conference on Learning Rep- resentations (ICLR), 2026. 12

work page 2026

[1] [1]

Per-Gaussian embedding-based deformation for deformable 3D Gaussian splatting

Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh. Per-Gaussian embedding-based deformation for deformable 3D Gaussian splatting. InEuropean Conference on Com- puter Vision (ECCV), 2024

work page 2024

[2] [2]

Efficient density control for 3d gaussian splat- ting.arXiv preprint arXiv:2411.10133, 2024

Xiaobin Deng, Changyu Diao, Min Li, Ruohan Yu, and Duanqing Xu. Efficient density control for 3D Gaussian splatting.arXiv preprint arXiv:2411.10133, 2024

work page arXiv 2024

[3] [3]

Mini-Splatting: Representing scenes with a constrained number of 11 Gaussians

Guangchi Fang and Bing Wang. Mini-Splatting: Representing scenes with a constrained number of 11 Gaussians. InEuropean Conference on Computer Vision (ECCV), 2024

work page 2024

[4] [4]

Improving adaptive density control for 3D Gaussian splatting

Glenn Grubert, Florian Barthel, Anna Hilsmann, and Peter Eisert. Improving adaptive density control for 3D Gaussian splatting. InInternational Joint Conference on Computer Vision, Imaging and Com- puter Graphics Theory and Applications (VISAPP), pages 610–621. SCITEPRESS, 2025

work page 2025

[5] [5]

PUP 3D-GS: Principled uncertainty pruning for 3D Gaussian splatting

Alex Hanson, Allen Tu, Vasu Singla, Mayuka Jayawardhana, Matthias Zwicker, and Tom Gold- stein. PUP 3D-GS: Principled uncertainty pruning for 3D Gaussian splatting. InIEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2025

work page 2025

[6] [6]

SC-GS: Sparse-controlled Gaussian splatting for editable dy- namic scenes

Yi-Hua Huang, Yang-Tian Sun, Zilong Yang, Xi- aoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. SC-GS: Sparse-controlled Gaussian splatting for editable dy- namic scenes. InIEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2024

work page 2024

[7] [7]

3D Gaussian splatting for real-time radiance field rendering

Bernhard Kerbl, Georgios Kopanas, Thomas Leimk¨ uhler, and George Drettakis. 3D Gaussian splatting for real-time radiance field rendering. In ACM Transactions on Graphics (SIGGRAPH), 2023

work page 2023

[8] [8]

Yiqing Liang, Mikhail Okunev, Mikaela Angelina Uy, Runfeng Li, Leonidas Guibas, James Tompkin, and Adam W. Harley. Monocular dynamic Gaus- sian splatting: Fast, brittle, and scene complexity rules.Transactions on Machine Learning Research (TMLR), 2025

work page 2025

[9] [9]

Taming 3DGS: High-quality radiance fields with limited resources

Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Car- rasco, and Fernando De La Torre. Taming 3DGS: High-quality radiance fields with limited resources. InSIGGRAPH Asia Conference Papers, 2024

work page 2024

[10] [10]

Drop- Gaussian: Structural regularization for sparse-view Gaussian splatting

Hyunwoo Park, Gun Ryu, and Wonjun Kim. Drop- Gaussian: Structural regularization for sparse-view Gaussian splatting. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

work page 2025

[11] [11]

Barron, Sofien Bouaziz, Dan B

Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B. Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. InIEEE/CVF International Conference on Computer Vision (ICCV), 2021

work page 2021

[12] [12]

Barron, Sofien Bouaziz, Dan B

Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B. Gold- man, 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

work page 2021

[13] [13]

D-NeRF: Neural ra- diance fields for dynamic scenes

Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural ra- diance fields for dynamic scenes. InIEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR), 2021

work page 2021

[14] [14]

Revising densification in Gaussian splatting

Samuel Rota Bul` o, Lorenzo Porzi, and Peter Kontschieder. Revising densification in Gaussian splatting. InEuropean Conference on Computer Vision (ECCV), 2024

work page 2024

[15] [15]

Steepest descent density control for compact 3D Gaussian splatting

Peihao Wang, Yuehao Wang, Dilin Wang, Sreyas Mohan, Zhiwen Fan, Lemeng Wu, Ruisi Cai, Yu- Ying Yeh, Zhangyang Wang, Qiang Liu, and Rakesh Ranjan. Steepest descent density control for compact 3D Gaussian splatting. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

work page 2025

[16] [16]

4D Gaussian splatting for real-time dynamic scene rendering

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. InIEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2024

work page 2024

[17] [17]

DropoutGS: Dropping out gaussians for better sparse-view rendering

Yexing Xu, Longguang Wang, Minglin Chen, Sheng Ao, Li Li, and Yulan Guo. DropoutGS: Dropping out gaussians for better sparse-view rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

work page 2025

[18] [18]

Deformable 3D Gaussians for high-fidelity monocular dynamic scene reconstruction

Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3D Gaussians for high-fidelity monocular dynamic scene reconstruction. InIEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2024

work page 2024

[19] [19]

AbsGS: Recovering fine details in 3D Gaussian splatting

Zongxin Ye, Wenyu Li, Sidun Liu, Peng Qiao, and Yong Dou. AbsGS: Recovering fine details in 3D Gaussian splatting. InProceedings of the 32nd ACM International Conference on Multimedia (MM), 2024

work page 2024

[20] [20]

Pixel-GS: Density control with pixel-aware gradient for 3D Gaussian splatting

Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-GS: Density control with pixel-aware gradient for 3D Gaussian splatting. In European Conference on Computer Vision (ECCV), 2024

work page 2024

[21] [21]

Gradient- direction-aware density control for 3D Gaussian splat- ting

Zheng Zhou, Yu-Jie Xiong, Jia-Chen Zhang, Chun- Ming Xia, Xihe Qiu, and Hongjian Zhan. Gradient- direction-aware density control for 3D Gaussian splat- ting. InInternational Conference on Learning Rep- resentations (ICLR), 2026. 12

work page 2026