Pith. sign in

REVIEW 4 major objections 6 minor 46 references

ACA-GS: Adaptive-Capacity Anchored Gaussian Splatting for Compact Dynamic Radiance Fields

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that adaptive per-anchor capacity allocation renders dynamic scenes at comparable quality with roughly 1.5× smaller models than existing anchor-based compression.

desk verdict A credible and practically useful 4DGS compression advance, but the headline claim slightly overstates quality and the opacity-accumulation importance metric needs a robustness check against nonuniform camera coverage. read the letter →

arxiv 2608.04581 v1 pith:BEUJMMYI submitted 2026-08-05 cs.CV

Seunghyeon Song , Joo Chan Lee , Chanung Park , Jun Young Jeong , Minseo Lee , Eunbyung Park , Jong Hwan Ko This is my paper
classification cs.CV
keywords Gaussiansplattingdynamicradiancefieldsanchor-basedrepresentationadaptivecardinalityfeaturemaskingnovelviewsynthesiscompression4Drendering
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 3D scenes rendered with Gaussian splatting usually trade storage against fidelity: to capture fast motion and fine detail, models add anchors and Gaussians everywhere, including plain regions where they are wasted. This paper argues that the fix is to let each anchor's capacity vary over space and time instead of giving every anchor the same fixed number of Gaussians and the same feature width. It proposes two coordinated mechanisms: relocating low-importance Gaussians to anchors that need more primitives, and masking out unneeded feature channels per anchor. On MPEG, Panoptic Sports, and N3DV the resulting models are 35–42% smaller than the prior anchor-based method at comparable visual quality, with MPEG storage at 3.5 MB versus 5.4 MB. If correct, the central insight is that representational budget should follow local spatiotemporal complexity, not a uniform template.

What carries the argument

The load-bearing object is the accumulated view-dependent opacity importance score of Eq. (1), $I_{i,k}$, summed over a relocation interval. It stands in for per-primitive importance in anchor-based models that do not store explicit standalone opacity. A cardinality gating mask keeps a fixed upper bound of Gaussians per anchor so shared MLPs decode a fixed-size tensor while only active Gaussians render, and relocation preserves the population by transferring dead Gaussians to importance-sampled parent anchors. For features, a base/adaptive factorization plus straight-through-learned masks and the conditional regularization loss $L_{CR}$ selectively disable channels while protecting motion-critical anchors.

What would settle it

Train the same method on a dynamic scene with deliberately nonuniform training camera coverage, or sweep the relocation interval from very short to very long, and compare storage and PSNR against a uniform-view model: if quality drops or size rises as view bias grows, the accumulated-opacity importance assumption is the weak link. Alternatively, replace importance-based relocation with random relocation under the same budget; if random relocation reproduces the 1.5× compression win, the opacity score is not doing the work.

Watch

Extended reading notes

Core claim

The paper's claim is that the rigid uniform parametrization of anchor-based dynamic Gaussian splatting is a source of memory waste, and that two adaptive mechanisms remove it. Adaptive Anchor Cardinality computes an accumulated view-dependent opacity score $I_{i,k}$ for every Neural Gaussian over a relocation interval; Gaussians below a threshold are deactivated and relocated to anchors sampled by importance, keeping the total Gaussian population fixed. Adaptive Anchor Feature Masking factorizes each anchor feature into a fixed base and a maskable adaptive part, with binary masks learned through a straight-through estimator and a conditional regularization loss that stops temporal features from absorbing spatial detail in moving regions. Together these mechanisms move capacity into high-complexity regions while shrinking model size. The paper reports up to 1.5× higher compression than the state-of-the-art anchor-based method, for example 3.5 MB at 29.92 PSNR versus 5.4 MB at 29.89 PSNR on MPEG.

Load-bearing premise

The load-bearing premise is that opacity accumulated over one relocation interval, using only anchors that happen to be spawned in the current training views, is a faithful global importance signal for every Neural Gaussian; if training views are unevenly distributed or the interval is short, capacity would migrate toward frequently seen views and the reported compression gains would shrink.

Editorial extensions

If this is right

  • On MPEG, the compressed model fits in 3.5 MB at 29.92 PSNR, about 35–42% smaller than GIFStream at comparable quality on MPEG, Panoptic Sports, and N3DV.
  • Because fewer anchors and Gaussians are rendered, per-frame cost also drops: the MPEG ablation reports FPS rising from 77 to 84 with the compression pipeline.
  • Relocation and feature masking are orthogonal to attribute entropy coding; combining them with the existing compression pipeline produces the full storage win.
  • The fixed upper-bound gating design means variable per-anchor cardinality does not force cardinality-specific MLP decoders or break batch-parallel rendering.
  • Simple static regions keep only a few Gaussians and few active feature channels, while fine or fast-moving regions receive the extra capacity, matching where the visual error actually lives.

Reading between the lines

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

  • The importance score is only evaluated on anchors inside the current training view frustum, so on capture setups with uneven camera coverage, capacity could drift toward frequently observed viewpoints over successive relocation cycles.
  • Relocation is driven by opacity contribution rather than by reconstruction error or bitrate; a rate-distortion-aware importance target could plausibly push compression further.
  • The same base/adaptive feature-masking idea could transfer to static scenes or hierarchical anchor structures, where each level could modulate capacity independently.
  • A direct test would sweep the relocation interval and the view distribution; if either strongly changes the compression gain, the importance metric is the sensitive component.
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 / 6 minor

Summary. The paper proposes ACA-GS, an anchor-based 4D Gaussian Splatting method that adaptively allocates representational capacity. It introduces Adaptive Anchor Cardinality (AAC), which relocates low-importance Neural Gaussians to anchors in more demanding regions based on an accumulated view-dependent opacity score, and Adaptive Anchor Feature Masking (AAFM), which learns per-anchor binary masks over a subset of feature channels to reduce redundancy. A conditional regularization loss is added to keep temporal features from absorbing spatial detail in dynamic regions. Experiments on MPEG, Panoptic Sports, and N3DV report storage reductions of roughly 35–42% relative to the GIFStream baseline, e.g., 3.5 MB vs. 5.4 MB on MPEG at comparable PSNR. The central claim is that adaptive capacity preserves visual quality while improving compression.

Significance. If the adaptive-capacity mechanism is validated, it addresses a genuine limitation of anchor-based 4DGS: the rigid uniform allocation of Gaussian counts and feature channels per anchor. The paper is clearly written and the ablation structure is sensible, with Tables 3–7 individually supporting the contribution of AAC, AAFM, and the regularization terms. The main MPEG result (3.5 MB at 29.92 PSNR versus GIFStream's 5.4 MB at 29.89 PSNR) is an encouraging demonstration of the compression benefit. However, the paper does not yet convincingly establish the view-independence of the importance score in Eq. (1), which is load-bearing for the whole relocation mechanism, and the reported SSIM/LPIPS values are consistently slightly worse than GIFStream. The lack of multiple runs and error bars also makes it hard to judge whether the observed differences are meaningful. The ideas are novel and well suited to the multimedia community, but the current evidence is not yet sufficient to support the strong 'without degrading visual quality' claim.

major comments (4)
  1. [Section 3.1, Eq. (1)] The paper asserts that accumulating view-dependent opacity over a relocation interval T_R makes the importance score 'independent of momentary view frustum and the corresponding temporal state.' This is not demonstrated. Since opacity is only decoded when the parent anchor is spawned inside the current training camera's frustum, anchors that appear in frequently sampled views will systematically accumulate higher scores than anchors visible only in rarely sampled views. The manuscript provides no ablation that varies the camera sampling distribution, no analysis of held-out view quality, and no ablation of the relocation interval T_R itself. This assumption is load-bearing because the reported compression gains come from pruning tens of thousands of anchors and Gaussians. Please add an experiment that (i) trains with a deliberately nonuniform camera distribution and reports per-view PSNR for under-sampled views, and (ii) compares Eq. (1) against a frustum-free or per-view importance proxy to verify that the accumulation actually removes view bias.
  2. [Tables 1 and 2] The abstract and Section 4.1 claim that the method preserves comparable quality or 'matching visual performance' while compressing more. The reported metrics do not fully support this. On MPEG, SSIM is 0.898 vs. 0.900 and LPIPS is 0.228 vs. 0.223 relative to GIFStream; on Panoptic Sports, PSNR, SSIM, and LPIPS are all worse (29.30/0.922/0.128 vs. 29.50/0.931/0.114); on N3DV, LPIPS is worse (0.058 vs. 0.051). The compression advantage is real, but the quality comparison is not strictly 'without degradation.' The authors should either report confidence intervals or paired significance tests for all three metrics, or soften the quality claim to 'comparable PSNR with small perceptual metric differences.'
  3. [Section 4, Tables 3–7] All quantitative results appear to be single runs with no error bars. The relocation procedure involves importance-proportional stochastic sampling and the masking mechanism uses a straight-through estimator, so run-to-run variance could plausibly be on the same order as the reported differences (e.g., 0.1–0.2 dB PSNR in the ablations). Hyperparameters such as lambda_M, gamma, tau, and rho are tuned on the same benchmark scenes, which further increases the risk of overfitting to those datasets. Please report mean and standard deviation over at least three seeds for the main comparison tables and for the key ablations. At minimum, the main MPEG comparison should include multiple runs.
  4. [Figure 6 and Table 1] The rate-distortion comparison in Figure 6 is described as controlling model size 'exclusively via anchor densification' for both GIFStream and the proposed method. However, ACA-GS has two additional capacity controls (Gaussian cardinality per anchor and feature masking), so controlling size only through anchor density may not span the method's natural operating range and could bias the RD comparison in either direction. The caption should state which variable is swept, over what range, and whether all other hyperparameters are fixed. Without this information, the RD advantage is difficult to interpret.
minor comments (6)
  1. [Table 3] The formatting of Table 3 is garbled: the checkmark columns for Comp., AAC, and AAFM are misaligned, and the rows are not clearly separated. For example, the row labeled 'GIFStream' appears after an incomplete row, and the sizes 248.3, 132.2, 117.6, 5.3, 4.0, 3.5 are not visually mapped to the correct method variants. Please reformat the table so each configuration is a distinct, labeled row.
  2. [Figure 6 caption] The caption refers to 'MPEG datasets' but Table 1 reports a single aggregate MPEG row. Please specify which MPEG sequences are used, how the rate-distortion curve is aggregated, and whether the curve is averaged over sequences or shown for a representative one.
  3. [Table 5 and Section 4.2] The claim that 'Our 48-channel configuration attains an optimal balance' is based on a single scene (N3DV Flame Steak). The text later says the setting is used 'across various scene scales,' but no cross-scene validation of the channel split is shown. Please clarify the generalization claim.
  4. [Section 3.2, Eq. (3)] The masking threshold gamma and the 'masking ratio' reported in Table 6 are not precisely defined. Please state explicitly whether the masking ratio is the fraction of masked adaptive channels after training, and how the straight-through estimator interacts with the threshold during inference.
  5. [Section 4] The text says 'AAC is activated every 500 iterations starting from 3,000 iteration,' but no ablation or sensitivity analysis is provided for the relocation interval. This is tied to the main concern about Eq. (1); at minimum, please state the total number of iterations and the number of relocation cycles.
  6. [General] The paper does not mention code or model availability. Given the number of hyperparameters and the stochastic relocation procedure, releasing code would substantially improve reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: compression results are measured against external baselines, and self-citations are not load-bearing.

full rationale

The paper's central claims—adaptive anchor cardinality and adaptive anchor feature masking leading to 35-42% storage reduction at comparable quality—are evaluated by direct measurement of model size and rendering quality against external baselines (GIFStream, STG, CSTG, etc.) on standard datasets. The compression mechanism is not derived by construction from the losses; instead, the method trains with rendering and entropy losses, prunes based on learned masks and an accumulated-opacity importance score, and then reports the resulting size and PSNR. No equation reduces to its own input, and no fitted parameter is renamed as a prediction. The importance score I_{i,k} (Eq. 1) is a heuristic proxy, not a self-defined target; the concern that its view-independence is unverified with nonuniform camera distributions is a correctness risk, not a circularity. Self-citations in the related work (Refs. [15]-[18]) are contextual and are not invoked to justify any load-bearing premise of the present method. The implementation builds on GIFStream [19], which is external. Therefore, no specific circular step can be quoted and exhibited.

Assumptions & free parameters 9 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a collection of hand-set hyperparameters and on assumptions about the anchor representation inherited from GIFStream and Scaffold-GS. The most fragile premise is the opacity-accumulation importance metric. No new physical entities are introduced; all capacities are algorithmic. The method's success depends on several datasets and on the compression pipeline of the baseline, so the transfer to other dynamic scenes is not guaranteed.

free parameters (9)
  • importance threshold tau = 0.005
    Gaussians with accumulated opacity below tau are flagged dead and eligible for relocation; value chosen via Table 7 ablation.
  • relocation cap rho = 20%
    At most 20 percent of Neural Gaussians are deactivated per relocation step; chosen via Table 7 ablation.
  • perturbation scale delta = 0.05
    Controls offset noise for relocated Gaussians in Eq. 2; set by default with no reported ablation.
  • maximum anchor cardinality K = 10
    Fixed upper bound on Neural Gaussians per anchor in gating design; selected as optimum in Table 4.
  • anchor feature channel split (F_base | F_adapt) = 24 | 24 on 48-channel anchor features
    Selected as the best capacity trade-off in Table 5.
  • masking threshold gamma = 0.01
    Sigmoid threshold for binary feature masks in Eq. 3; ablated in Table 7.
  • mask regularization weight lambda_M = tuned to reach ~60% masking ratio
    Chosen to reach the target final masking ratio; Table 6 shows the quality-size trade-off.
  • conditional regularization weight lambda_CR = 0.1
    Balances feature sparsity versus temporal capacity; ablated in Table 7.
  • relocation interval = every 500 iterations from iteration 3,000
    AAC activation schedule reported in Section 4; no ablation is shown for this schedule.
assumptions (4)
  • domain assumption Anchor-based 4DGS with anchor-level rigid deformation can represent dynamic scenes faithfully if anchors are dense enough in motion regions.
    The whole capacity allocation story assumes motion is captured by anchor-level deformation; the paper states this in Section 4.2 when explaining why dynamic regions need more anchors.
  • ad hoc to paper Accumulated view-dependent opacity over training views (Eq. 1) is a valid proxy for global primitive importance independent of view frustum and temporal state.
    This assumption justifies the relocation mechanism; the paper notes single-view opacities are biased, but gives no proof or experiment that accumulation removes the bias.
  • domain assumption A fixed-size shared MLP decoder can predict attributes for a fixed maximum cardinality while gating inactive primitives without changing optimization behavior.
    The gating mask design in Section 3.1 relies on this to keep batch-parallel execution; no formal or empirical validation is provided beyond the main results.
  • domain assumption The GIFStream compression pipeline (entropy coding, pruning, feature-stream prediction) remains optimal or near-optimal when combined with AAC and AAFM.
    The final objective inherits L_base from GIFStream and the storage numbers depend on that compressor; compatibility is assumed rather than analyzed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ACA-GS: Adaptive-Capacity Anchored Gaussian Splatting for Compact Dynamic Radiance Fields." pith.science (2026). https://pith.science/paper/BEUJMMYI

@misc{pith2026260804581,
  author       = {Pith},
  title        = {Pith review of: ACA-GS: Adaptive-Capacity Anchored Gaussian Splatting for Compact Dynamic Radiance Fields},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BEUJMMYI}},
  note         = {Machine review of arXiv:2608.04581}
}
read the original abstract

Recent advances in 4D Gaussian Splatting (4DGS) enable high-fidelity, real-time spatiotemporal rendering, but expose a fundamental trade-off between motion expressiveness and storage efficiency. While anchor-based designs achieve compactness through anchor-level parameter sharing, their rigid uniform parametrization enforces fixed Neural Gaussian counts and feature budgets per anchor. Consequently, insufficient fidelity is addressed by excessive anchor density, rather than lightweight, targeted increases in Neural Gaussian count or feature capacity, resulting in memory waste. To overcome this rigidity, we introduce an adaptive-capacity anchor-based framework that dynamically allocates the representational capacity based on local spatiotemporal demands. Adaptive Anchor Cardinality varies the number of Neural Gaussians per anchor, concentrating primitives in regions of high geometric or motion complexity while suppressing redundancy. In parallel, Adaptive Anchor Feature Masking modulates anchor-level feature channels, assigning rich features to complex regions and lightweight representations to simpler ones. Experiments on MPEG, Panoptic Sports, and N3DV datasets demonstrate substantial storage reduction without degrading visual quality. Notably, on challenging MPEG sequences with complex motion, our method achieves up to 1.5x higher compression than state-of-the-art anchor-based methods while preserving comparable quality.

Figures

Figures reproduced from arXiv: 2608.04581 by the authors.

Figure 1
Figure 1. We propose adaptive-capacity anchor-based representation that dynamically adjusts both per-anchor cardinality and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview. (a) Fixed Anchor Capacity enforces uni [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Distribution of Anchor and Neural Gaussian within [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Overall structure of our method. Given a set of 3D anchors, MLPs decode per-anchor and time-dependent features to [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Per-view visibility of anchors. Since only anchors [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Rate-distortion (PSNR) curve on MPEG datasets. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Visualization. In this figure, we visualize the pixel-wise density of anchor components using a heatmap spectrum [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Qualitative comparison on N3DV Cook Spinach [ [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 29 canonical work pages

  1. [1]

    Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh. 2024. Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting. InEuropean Conference on Computer Vision (ECCV)

    work page 2024

  2. [2]

    Yoshua Bengio, Nicholas Léonard, and Aaron Courville. 2013. Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation. arXiv:1308.3432 [cs.LG] https://arxiv.org/abs/1308.3432

    arXiv 2013

  3. [3]

    Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. 2024. HAC: Hash-grid Assisted Context for 3D Gaussian Splatting Compression. In European Conference on Computer Vision

    work page 2024

  4. [4]

    Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. 2025. HAC++: Towards 100X Compression of 3D Gaussian Splatting.IEEE Transactions on Pattern Analysis and Machine Intelligence47, 11 (nov 2025)

    work page 2025

  5. [5]

    Zicong Chen, Zhenghao Chen, Wei Jiang, Wei Wang, Lei Liu, and Dong Xu. 2025. 4DGS-CC: A Contextual Coding Framework for 4D Gaussian Splatting Data Compression. arXiv:2504.18925 [cs.CE] https://arxiv.org/abs/2504.18925

    arXiv 2025

  6. [6]

    Woong Oh Cho, In Cho, Seoha Kim, Jeongmin Bae, Youngjung Uh, and Seon Joo Kim. 2026. 4D scaffold Gaussian splatting with dynamic-aware anchor growing for efficient and high-fidelity dynamic scene reconstruction(AAAI’26). AAAI Press, Article 375, 9 pages. doi:10.1609/aaai.v40i5.37332

    work page doi:10.1609/aaai.v40i5.37332 2026

  7. [7]

    Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. 2024. LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and 200+ FPS. InAdvances in Neural Information Processing Systems, Vol. 37. Curran Associates, Inc., 140138–140158. doi:10.52202/079017-4447

    doi:10.52202/079017-4447 2024

  8. [8]

    Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. 2024. EAGLES: Efficient Accelerated 3D Gaussians with Lightweight EncodingS. InEuropean Conference on Computer Vision (ECCV). 54–71

    work page 2024

Show all 46 references
  1. [9]

    He Huang, Qi Yang, Mufan Liu, Yiling Xu, and Zhu Li. 2025. ADC-GS: Anchor- Driven Deformable and Compressed Gaussian Splatting for Dynamic Scene Reconstruction. InProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, IJCAI-25. 1179–1187

    2025

  2. [10]

    Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiao- juan Qi. 2024. SC-GS: Sparse-Controlled Gaussian Splatting for Editable Dynamic Scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 4220–4230

    2024

  3. [11]

    Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. 2015. Panoptic Studio: A Massively Multi- view System for Social Motion Capture. InProceedings of the IEEE International Conference on Computer Vision (ICCV)

    2015

  4. [12]

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis

  5. [13]

    Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi

  6. [14]

    Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. 2024. DynMF: Neural Motion Factorization for Real-time Dynamic View Synthesis with 3D Gaussian Splatting.ECCV(2024)

    2024

  7. [15]

    Joo Chan Lee, Jong Hwan Ko, and Eunbyung Park. 2025. Optimized Minimal 3D Gaussian Splatting. InAdvances in Neural Information Processing Systems

    2025

  8. [17]

    Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park

  9. [18]

    Minseo Lee, Byeonghyeon Lee, Lucas Yunkyu Lee, Eunsoo Lee, Sangmin Kim, Seunghyeon Song, Joo Chan Lee, Jong Hwan Ko, Jaesik Park, and Eunbyung Park. 2025. Optimized Minimal 4D Gaussian Splatting. arXiv:2510.03857 [cs.CV] https://arxiv.org/abs/2510.03857

    2025

  10. [19]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Compact 3D Gaussian Representation for Radiance Field. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

  11. [20]

    Tianye Li, Mira Slavcheva, Michael Zollhöfer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard New- combe, and Zhaoyang Lv. 2022. Neural 3D Video Synthesis From Multi-View Video. InProceedings of the IEEE/CVF Conference on...

    2022

  12. [21]

    arXiv preprint arXiv:2408.03822(2024)

    Compact 3D Gaussian Splatting for Static and Dynamic Radiance Fields. arXiv preprint arXiv:2408.03822(2024)

    2024 arXiv

  13. [22]

    Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. 2024. Spacetime Gaussian Feature Splatting for Real-Time Dynamic View Synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 8508–8520

    2024

  14. [23]

    Hao Li, Sicheng Li, Xiang Gao, Abudouaihati Batuer, Lu Yu, and Yiyi Liao. 2025. GIFStream: 4D Gaussian-based Immersive Video with Feature Stream. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2025

  15. [24]

    Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. 2024. Scaffold-gs: Structured 3d gaussians for view-adaptive render- ing. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2024

  16. [25]

    Yilong Li, Bo Pang, Yisong Chen, and Guoping Wang. 2025. Anchored 4D Gaussian Splatting for Dynamic Novel View Synthesis. InProceedings of the SIGGRAPH Asia 2025 Conference Papers. Article 73, 11 pages

    2025

  17. [26]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. 2020. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. InECCV

    2020

  18. [27]

    Mufan Liu, Qi Yang, He Huang, Wenjie Huang, Zhenlong Yuan, Zhu Li, Yiling Xu, and Yunfeng Guan. 2026. Light4GS: Lightweight Compact 4D Gaussian Splatting Generation via Context Model.IEEE Trans. Circuits Syst. Video Technol.36, 6 (2026), 7992–8005. doi:10.1109/TCSVT.2026.3656896

    2026

  19. [28]

    Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai

  20. [29]

    Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. 2024. Dynamic 3D Gaussians: Tracking by Persistent Dynamic View Synthesis. In3DV

    2024

  21. [30]

    Henan Wang, Hanxin Zhu, Tianyu He, Runsen Feng, Jiajun Deng, Jiang Bian, and Zhibo Chen. 2024. End-to-End Rate-Distortion Optimized 3D Gaussian Representation. InEuropean Conference on Computer Vision

    2024

  22. [31]

    Wieland Morgenstern, Florian Barthel, Anna Hilsmann, and Peter Eisert. 2025. Compact 3D Scene Representation via Self-Organizing Gaussian Grids. InEuro- pean Conference on Computer Vision (ECCV). 18–34

    2025

  23. [32]

    Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 2024. 4D Gaussian Splatting for Real- Time Dynamic Scene Rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 20310–20320

    2024

  24. [33]

    Zhen Xu, Yinghao Xu, Zhiyuan Yu, Sida Peng, Jiaming Sun, Hujun Bao, and Xi- aowei Zhou. 2024. Representing Long Volumetric Video with Temporal Gaussian Hierarchy.ACM Transactions on Graphics43, 6 (November 2024)

    2024

  25. [34]

    Schönberger and Jan-Michael Frahm

    Johannes L. Schönberger and Jan-Michael Frahm. 2016. Structure-from-Motion Revisited. InCVPR. doi:10.1109/CVPR.2016.445

    2016 doi

  26. [35]

    Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin

  27. [36]

    Bovik, Hamid R

    Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. 2004. Image Quality Assessment: From Error Visibility to Structural Similarity.IEEE Transactions on Image Processing13, 4 (2004), 600–612

    2004

  28. [37]

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. 2024. Mip-Splatting: Alias-free 3D Gaussian Splatting.Conference on Computer Vision and Pattern Recognition (CVPR)(2024)

    2024

  29. [38]

    Yuheng Yuan, Qiuhong Shen, Xingyi Yang, and Xinchao Wang. 2025. 1000+ FPS 4D Gaussian Splatting for Dynamic Scene Rendering. InAdvances in Neural Information Processing Systems. 74665–74683

    2025

  30. [39]

    Jinbo Yan, Rui Peng, Luyang Tang, and Ronggang Wang. 2024. 4D Gaussian Splatting with Scale-aware Residual Field and Adaptive Optimization for Real- time rendering of temporally complex dynamic scenes. InACM Multimedia 2024

    2024

  31. [40]

    Xinjie Zhang, Zhening Liu, Yifan Zhang, Xingtong Ge, Dailan He, Tongda Xu, Yan Wang, Zehong Lin, Shuicheng Yan, and Jun Zhang. 2025. MEGA: Memory-Efficient 4D Gaussian Splatting for Dynamic Scenes. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV...

    2025

  32. [41]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 20331–20341

  33. [42]

    Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. 2024. Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting. In International Conference on Learning Representations (ICLR)

    2024

  34. [45]

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang

  35. [2018]

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. InCVPR

  36. [2023]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering.ACM Transactions on Graphics42, 4 (July 2023)

    2023

  37. [2024]

    InAdvances in Neural Information Processing Systems (NeurIPS)

    3D Gaussian Splatting as Markov Chain Monte Carlo. InAdvances in Neural Information Processing Systems (NeurIPS)

  38. [2025]

    Octree-GS: Towards Consistent Real-time Rendering with LOD-Structured 3D Gaussians.IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.