Pith. sign in

REVIEW 6 major objections 6 minor 1 cited by

LongSplat: Online Generalizable 3D Gaussian Splatting from Long Sequence Images

T0 review · 6 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 LongSplat claims that online 3D Gaussian reconstruction from long image sequences can run in real time by storing accumulated Gaussians in an image-like 2D map and updating them per frame, instead of re-optimizing the scene.

desk verdict A novel streaming Gaussian fusion/compression idea whose headline numbers rest on an unfair baseline comparison and no timing measurements. read the letter →

arxiv 2507.16144 v1 pith:ZYQGLNY6 submitted 2025-07-22 cs.CV

Guichen Huang , Ruoyu Wang , Xiangjun Gao , Che Sun , Yuwei Wu , Shenghua Gao , Yunde Jia This is my paper
classification cs.CV
keywords 3DGaussianSplattingnovelviewsynthesisonlinereconstructionlong-sequencevideoredundancycompressionGaussian-ImageRepresentationfeed-forwardreal-timerendering
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

LongSplat claims that online 3D Gaussian reconstruction from long image sequences can run in real time by storing accumulated Gaussians in an image-like 2D map and updating them per frame, instead of re-optimizing the scene. The paper introduces the Gaussian-Image Representation (GIR), which encodes Gaussian parameters into a per-pixel 2D format, enabling fusion of current-view observations with historical context and identity-aware pruning of redundant historical Gaussians. On the DL3DV benchmark, the compressed model removes about 44% of Gaussians while still beating the per-pixel baseline by several decibels, suggesting that per-frame incremental updates with selective compression solve the degradation that dense per-pixel predictors suffer on long sequences. If true, this makes real-time 3D perception practical for embodied agents and robotics, where scenes must be rebuilt continuously as new frames arrive.

What carries the argument

The central object is the Gaussian-Image Representation (GIR), a per-pixel 2D encoding of 3D Gaussian parameters (projected position, covariance, opacity, and unique ID). For each pixel the representation keeps only the dominant Gaussian along the ray, selected either by a nearest or most-contributive criterion; the accompanying ID map gives a deterministic link between pixels and 3D Gaussians. GIR carries the argument by enabling three operations in 2D: fusing historical Gaussians into the current view, predicting a soft visibility mask that compresses redundant Gaussians, and applying geometry and mask losses against ground-truth compressed Gaussians without full 3D computation.

What would settle it

Render a controlled scene containing a semi-transparent object, such as tinted glass, in front of a textured background using a trained LongSplat model and compare with a full alpha-composited 3DGS render; if the single-Gaussian projection omits the background contribution, the synthesized view will miss or blur the background pattern visible through the glass.

Watch

Extended reading notes

Core claim

The central claim is that an online, generalizable 3D Gaussian Splatting system can maintain a global scene representation across arbitrarily long image sequences by treating Gaussian parameters as 2D image-like maps. The paper's GIR projects each view's Gaussians into a structured per-pixel format carrying position, covariance, opacity, and a unique Gaussian ID; rendering uses a sparse strategy that keeps only the single dominant Gaussian per ray. This representation lets the model fuse historical Gaussians into the current frame, predict a confidence mask that deletes redundant or outdated splats, and supervise geometry directly against per-scene optimized Gaussians compressed with LightGaussian. On DL3DV, the full model reaches 23.71 dB PSNR at 50 views compared to DepthSplat's 21.39 dB, and the compressed variant keeps 23.54 dB while dropping 43.77% of Gaussians; at 120 views the compressed model stays at 21.34 dB where DepthSplat falls to 17.77 dB. The authors conclude that per-frame incremental updates with selective compression solve the degradation that dense per-pixel predictors suffer on long sequences.

Load-bearing premise

The method assumes that each pixel can be represented by a single dominant Gaussian, so scenes with transparency or multiple overlapping surfaces may lose important information when Gaussians are compressed or when the single-Gaussian projection is used for supervision.

Editorial extensions

If this is right

  • Long-sequence novel view synthesis becomes memory-bounded because the global scene does not grow linearly with frames when redundant Gaussians are removed each step.
  • Real-time online reconstruction becomes feasible for embodied agents, since each frame only requires a small update rather than a full re-optimization.
  • The compression mask doubles as an editing handle, allowing outdated or dynamic-content Gaussians to be removed incrementally without reprocessing the sequence.
  • Because GIR grounds supervision in 2D, training signals from per-scene optimized Gaussian datasets transfer cheaply to feed-forward models.
  • The same feature extraction pipeline that LongSplat builds on gains scalability without changing its learned parameters, since the baseline is kept fixed during training.

Reading between the lines

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

  • GIR's sparse single-Gaussian-per-pixel assumption could be extended to multi-Gaussian channels, preserving transparency while retaining the 2D structure; that would likely recover quality on glossy or translucent surfaces.
  • The identity-aware compression might serve as a lightweight dynamic-scene mechanism: if Gaussian IDs persist, deletions and additions per frame correspond to scene changes, which could be read as motion or novelty signals.
  • The 44% compression number is measured on DL3DV; on scenes with more repeated textures or more transparent structures the compression ratio would plausibly vary, and the mask threshold provides a direct trade-off knob for different applications.
  • A testable prediction is that on scenes with strong multi-view appearance changes (specular highlights), the most-contributive rendering rule should be preferred over nearest rendering, and the fusion should reduce artifacts; a benchmark isolating such scenes would separate the contribution of GIR from the baseline's depth cues.
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

6 major / 6 minor

Summary. The paper proposes LongSplat, an online generalizable 3D Gaussian Splatting framework for long image sequences. The central idea is a Gaussian-Image Representation (GIR) that projects 3D Gaussian parameters, together with a Gaussian ID map, into a structured 2D image-like format. GIR is used for three functions: fusing historical global Gaussians with current-view features via a transformer module, predicting a soft update mask that selectively compresses redundant Gaussians, and providing 2D supervision targets derived from per-scene LightGaussian-optimized teacher scenes. Training combines photometric, geometric alignment, and mask losses, and experiments on DL3DV-10K compare LongSplat with DepthSplat and MVSplat-360 under 12, 50, and 120 context views, reporting PSNR up to 23.71 dB and Gaussian compression up to 44.37%.

Significance. If the results were substantiated, the GIR representation and identity-aware compression would be a meaningful step toward streaming 3DGS with bounded storage and rendering cost. The paper also ships a falsifiable experimental protocol (fixed target sets, three view counts, compression curves), which is a strength. However, the current evidence does not support the headline claims: the only strong baseline comparison (DepthSplat) is confounded by training-distribution mismatch, no timing or memory numbers back the 'real-time' claim, the 44% compression is an operating point of a tunable threshold, and the mask-supervision loss as written has a sign inconsistency that would train the opposite of the intended behavior. These are load-bearing issues that require new experiments and corrections.

major comments (6)
  1. [§4, Table 1] The DepthSplat baseline is used with all parameters fixed (Section 4, Implementation Details) and is not fine-tuned on DL3DV-10K, whereas LongSplat is trained on DL3DV-10K and an auxiliary LightGaussian-compressed set. The PSNR gains in Table 1, such as +2.32 dB at 50 views and the 3.6 dB advantage claimed in the Abstract for the compressed 120-view variant, conflate method improvement with adaptation to the evaluation distribution. To support the state-of-the-art claim, the authors should fine-tune DepthSplat on DL3DV-10K under the same view-count and resolution settings, or else provide a zero-shot evaluation in which both models are evaluated on a distribution neither has seen during training.
  2. [Abstract/§4] No timing or memory figures appear anywhere in Section 4. The only efficiency quantity is the c-ratio (percentage of compressed Gaussians), which is a model-output count, not a runtime or memory measurement. The 'real-time reconstruction' claim in the Abstract and Introduction is therefore unsupported. The paper should report per-frame update time, rendering FPS, and GPU memory for Ours, Ours-c, and DepthSplat on the same hardware.
  3. [§3.4, Eqs. (9)–(12)] There is a sign inconsistency in the mask supervision. The text states 'we treat Gaussians with high overlap as redundant and assign ground truth M^gt_t = 1', but Eq. (9) defines alpha^mod = M_t · alpha_uv, so M=1 preserves opacity and M=0 removes it. Under the stated goal of pruning redundant Gaussians, high-overlap pixels should receive M^gt_t = 0, not 1. As written, the BCE loss would encourage retaining redundant splats. The authors must correct either the equations or the text and clarify the intended semantics of the mask.
  4. [§3.3, §4.1, Table 3] The reported 44% Gaussian reduction is a direct consequence of the tunable confidence threshold tau. Table 3 shows the c-ratio climbing from 23.32% at tau=0.1 to 43.77% at tau=0.5, so the Abstract's '44%' is an operating point rather than an emergent property of the method. The paper should present the full quality-versus-compression curve for a range of tau values and should de-emphasize the single-threshold number in the headline claims.
  5. [§3.3] The sparse GIR rendering assumption, in which each pixel is associated with only a single dominant Gaussian via Eq. (3) or (4), is never validated for real scenes containing translucency, overlapping surfaces, or multi-modal distributions along a ray. Since the history fusion, mask prediction, and compression all operate on these single-Gaussian maps, a failure of this assumption could directly produce artifacts in the rendered output. The authors should provide an analysis of the reconstruction error induced by this sparsification, for example by comparing nearest and most-contributive rendering against full alpha-compositing in scenes with known semi-transparent or layered geometry.
  6. [§4.1] The comparison set is too narrow for the claimed state-of-the-art status. Table 1 includes only DepthSplat and MVSplat-360 (which runs out of memory at 50 views), while Section 2 names FreeSplat, Long-LRM, Zpressor, and StreamGS as related sequential generalizable methods. At least one or two of these should be evaluated on the same protocol so that the 'state-of-the-art' claim in the Abstract has empirical support.
minor comments (6)
  1. [Figure 2 caption] The word 'Gussian' should be 'Gaussian' in the caption of Figure 2.
  2. [References] Reference [1] is a placeholder entry ('Our Proposed AdaptiveGaussian') and should be replaced with a proper citation or removed.
  3. [Eqs. (2) and (6)] The symbol mu_uv in Eq. (2) is described as a projected 2D position, but Eq. (6) uses mu_pred and mu_gt in a 3D position alignment loss; please clarify the coordinate space and how the 2D projections are lifted to 3D for the geometric loss.
  4. [Table 3] The paper reports DepthSplat results under varying thresholds, but DepthSplat has no learned mask; the post-hoc procedure for obtaining 'DepthSplat-c' (shown in Figure 4) is not described anywhere and should be specified.
  5. [§3.2, §4.2] The number N of temporally adjacent neighbors used for multi-view feature extraction is not given in Section 3.2; the implementation section should state the exact value used in the experiments.
  6. [§4, Training Datasets] The auxiliary training set is filtered to scenes with 'compression rate > 30% and PSNR > 28.0' after LightGaussian optimization; this selection biases the training distribution toward scenes in which the teacher is highly compressible, and the sensitivity of the reported c-ratio and PSNR to this filter should be analyzed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation chain is self-contained; the DepthSplat comparison confound is an experimental-validity issue, not a circular one.

full rationale

I walked the claimed derivation chain in Sections 3 and 4 and found no equation-level or citation-level reduction. GIR (Eq. 2) encodes Gaussian parameters into a 2D map, and the sparse rendering rules (Eqs. 3–4) define how teacher Gaussians are projected; these are not defined in terms of the quantities they are later said to predict. The LightGaussian-derived GIRs serve as teacher-student distillation targets: the model learns to predict Gaussian parameters from images, and at inference the teacher is not reused, so the prediction is not constructed from its own label. The OBB overlap metric (Eq. 10) defines a ground-truth mask that supervises the learned mask through BCE (Eq. 11); the learned mask is a neural network output, not the overlap metric itself. The compression ratios are operating points set by a tunable threshold tau and are reported as empirical outcomes, not as fitted quantities renamed as predictions. The unusual self-citation [1], 'Our Proposed AdaptiveGaussian', is anomalous and incomplete, but it appears only in related work and is not load-bearing for any derivation. The main empirical weakness is that DepthSplat is evaluated with fixed parameters while LongSplat is trained on DL3DV-10K, so Table 1 may conflate method gains with training-distribution adaptation; likewise, the 'real-time reconstruction' claim is asserted without wall-clock timings. These are correctness and evidence concerns, not circularity. No equation reduces to its own input, and no central claim is forced by a self-citation chain.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The central claims rest on several hand-set hyperparameters, most importantly the confidence threshold tau that directly produces the reported 44% Gaussian reduction, and on domain assumptions about the reliability of the DepthSplat features, the single-Gaussian sparse rendering, the OBB overlap metric as a redundancy proxy, and the LightGaussian teacher. The GIR is a new internal representation without independent external evidence.

free parameters (3)
  • confidence threshold tau = 0.5 for the reported 44% compression; also tested at 0.1 and 0.3
    The binary compression mask is obtained by thresholding predicted soft weights. Table 3 shows the c-ratio varies with tau, so the headline 44% reduction is a direct consequence of choosing tau=0.5.
  • lambda_pos and lambda_neg = not reported
    Weights in the weighted BCE mask loss (Eq. 12) are set by hand to counteract the photometric loss, but their values are not given in the paper.
  • lambda_Sigma = 0.5
    Weight of the covariance consistency loss in Eq. 8, chosen by hand to balance position and shape alignment.
assumptions (5)
  • domain assumption DepthSplat's fixed features and per-pixel Gaussian predictions are a sufficient and reliable basis for the streaming reconstruction.
    The method freezes DepthSplat parameters and builds on its outputs. If these features are not robust under pose noise or unusual scenes, the GIR fusion and compression inherit the errors. Stated in Implementation Details, Section 4.
  • domain assumption A single dominant Gaussian per pixel (sparse rendering) faithfully represents the scene for fusion and compression.
    GIR adopts a sparse rendering strategy where each pixel is associated with only one Gaussian (Eqs. 3-4). This discards all other contributions along the ray, which may be invalid for translucent, overlapping, or multi-surface regions. Section 3.3.
  • domain assumption The OBB-based asymmetric overlap metric IoUp (Eq. 10) is an accurate proxy for redundancy and compression targets.
    The mask supervision assumes that a historical Gaussian covered by a current-view Gaussian's OBB is redundant and should be masked out. This geometric heuristic may not align with perceptual redundancy or with correct handling of dynamic scenes. Section 3.4.
  • domain assumption LightGaussian per-scene optimized Gaussians provide a valid ground-truth teacher for training GIR.
    The training targets are constructed by running per-scene optimization with LightGaussian and rendering Most-Contributive GIRs as supervision. Any systematic bias or failure mode of LightGaussian is baked into the teacher. Section 3.3.
  • standard math Standard perspective projection and alpha compositing mathematics.
    Equations 1-4 and the projection operator Pi rely on standard computer graphics and camera model mathematics.
invented entities (1)
  • Gaussian-Image Representation (GIR) with Gaussian ID map
    purpose: To encode 3D Gaussian parameters into a view-aligned 2D image-like format for fusion, compression, and supervision.
    GIR is a new internal representation introduced by the paper; it has no external falsifiable handle outside the paper's own pipeline. The ID map is a bookkeeping device. No independent evidence is provided beyond the paper's experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LongSplat: Online Generalizable 3D Gaussian Splatting from Long Sequence Images." pith.science (2026). https://pith.science/paper/ZYQGLNY6

@misc{pith2026250716144,
  author       = {Pith},
  title        = {Pith review of: LongSplat: Online Generalizable 3D Gaussian Splatting from Long Sequence Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZYQGLNY6}},
  note         = {Machine review of arXiv:2507.16144}
}
read the original abstract

3D Gaussian Splatting achieves high-fidelity novel view synthesis, but its application to online long-sequence scenarios is still limited. Existing methods either rely on slow per-scene optimization or fail to provide efficient incremental updates, hindering continuous performance. In this paper, we propose LongSplat, an online real-time 3D Gaussian reconstruction framework designed for long-sequence image input. The core idea is a streaming update mechanism that incrementally integrates current-view observations while selectively compressing redundant historical Gaussians. Crucial to this mechanism is our Gaussian-Image Representation (GIR), a representation that encodes 3D Gaussian parameters into a structured, image-like 2D format. GIR simultaneously enables efficient fusion of current-view and historical Gaussians and identity-aware redundancy compression. These functions enable online reconstruction and adapt the model to long sequences without overwhelming memory or computational costs. Furthermore, we leverage an existing image compression method to guide the generation of more compact and higher-quality 3D Gaussians. Extensive evaluations demonstrate that LongSplat achieves state-of-the-art efficiency-quality trade-offs in real-time novel view synthesis, delivering real-time reconstruction while reducing Gaussian counts by 44\% compared to existing per-pixel Gaussian prediction methods.

Figures

Figures reproduced from arXiv: 2507.16144 by the authors.

Figure 1
Figure 1. Overview of the Longsplat framework. Given an input image sequence {It} T t=1, our model incrementally constructs a global 3D Gaussian scene representation G g through iterative frame-wise updates. At each timestep t, we extract two complementary feature streams: (1) a multi-view spatial feature map Fc from the current frame and its temporally adjacent neighbors using the DepthSplat pipeline, providing local geometr… view at source ↗
Figure 2
Figure 2. Overview of the proposed Gaussian-Image Represen￾tation (GIR) and its four core capabilities. GIR encodes per-pixel Gaussian parameters into a structured 2D image space, enabling efficient and flexible 3D reasoning. (a) History Fusion: GIR al￾lows temporally consistent fusion of 3D Gaussians across multiple frames by leveraging shared Gaussian IDs. (b) Global Update: GIR supports gradient-based updates to the global… view at source ↗
Figure 3
Figure 3. Novel view synthesis on 12 context views. Ref, GT are the input image and the ground truth. Ours-c delete the mask regions. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Novel view synthesis on 50 context views. Ref. rgb and Ref. mask are the input image and compression mask. GT is the ground [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. OF$^3$GS: On-the-Fly Feed-Forward 3D Gaussian Splatting from Unposed Images

    cs.CV 2026-06 unverdicted novelty 6.0 of 10

    FreeStreamGS achieves online NVS from unposed streaming inputs competitive with offline 3DGS methods via decoupled intrinsic recovery and dynamic point refinement.

Reference graph

Works this paper leans on

56 extracted references · 34 canonical work pages · cited by 1 Pith paper

  1. [1]

    Adaptivegaussian: Gener- alizable 3d gaus-sian reconstruction from arbitrary views

    Our Proposed AdaptiveGaussian. Adaptivegaussian: Gener- alizable 3d gaus-sian reconstruction from arbitrary views

    work page

  2. [2]

    Distractor- free generalizable 3d gaussian splatting.arXiv preprint arXiv:2411.17605, 2024

    Yanqi Bao, Jing Liao, Jing Huo, and Yang Gao. Distractor- free generalizable 3d gaussian splatting.arXiv preprint arXiv:2411.17605, 2024

    arXiv 2024

  3. [3]

    Mip-nerf: A multiscale representation for anti-aliasing neu- ral radiance fields

    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. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 5855–5864, 2021

    work page 2021

  4. [4]

    Zip-nerf: Anti-aliased grid-based neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697–19705, 2023

    2023

  5. [5]

    Hexplane: A fast representa- tion for dynamic scenes

    Ang Cao and Justin Johnson. Hexplane: A fast representa- tion for dynamic scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130–141, 2023

    2023

  6. [6]

    pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction

    David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19457–19467, 2024

    work page 2024

  7. [7]

    Mvsnerf: Fast general- izable radiance field reconstruction from multi-view stereo

    Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast general- izable radiance field reconstruction from multi-view stereo. InProceedings of the IEEE/CVF international conference on computer vision, pages 14124–14133, 2021

    work page 2021

  8. [8]

    Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images

    Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. InEuropean Conference on Computer Vision, pages 370–386. Springer, 2024

    2024

Show all 56 references
  1. [9]

    Mvsplat360: Feed-forward 360 scene synthesis from sparse views.arXiv preprint arXiv:2411.04924, 2024

    Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views.arXiv preprint arXiv:2411.04924, 2024

    2024 arXiv

  2. [10]

    Splatter-360: Generalizable 360 gaussian splatting for wide- baseline panoramic images

    Zheng Chen, Chenming Wu, Zhelun Shen, Chen Zhao, We- icai Ye, Haocheng Feng, Errui Ding, and Song-Hai Zhang. Splatter-360: Generalizable 360 gaussian splatting for wide- baseline panoramic images. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 2...

    2025

  3. [11]

    Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps.Advances in neural information processing systems, 37:140138–140158, 2024

    Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, Zhangyang Wang, et al. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps.Advances in neural information processing systems, 37:140138–140158, 2024

    2024

  4. [12]

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

    Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 12479–12488, 2023

    2023

  5. [13]

    Mani-gs: Gaussian splatting manipulation with triangular mesh

    Xiangjun Gao, Xiaoyu Li, Yiyu Zhuang, Qi Zhang, Wenbo Hu, Chaopeng Zhang, Yao Yao, Ying Shan, and Long Quan. Mani-gs: Gaussian splatting manipulation with triangular mesh. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 21392–21402, 2025

    2025

  6. [14]

    Mps-nerf: Generalizable 3d hu- man rendering from multiview images.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022

    Xiangjun Gao, Jiaolong Yang, Jongyoo Kim, Sida Peng, Zicheng Liu, and Xin Tong. Mps-nerf: Generalizable 3d hu- man rendering from multiview images.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022

    2022

  7. [15]

    Rgbd gs-icp slam

    Seongbo Ha, Jiung Yeon, and Hyeonwoo Yu. Rgbd gs-icp slam. InEuropean Conference on Computer Vision, pages 180–197. Springer, 2024

    2024

  8. [16]

    2d gaussian splatting for geometrically ac- curate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically ac- curate radiance fields. InACM SIGGRAPH 2024 conference papers, pages 1–11, 2024

    2024

  9. [17]

    Neds-slam: A neural explicit dense semantic slam framework using 3d gaussian splatting.IEEE Robotics and Automation Letters, 2024

    Yiming Ji, Yang Liu, Guanghu Xie, Boyu Ma, Zongwu Xie, and Hong Liu. Neds-slam: A neural explicit dense semantic slam framework using 3d gaussian splatting.IEEE Robotics and Automation Letters, 2024

    2024

  10. [18]

    Selfsplat: Pose-free and 3d prior-free generalizable 3d gaussian splatting

    Gyeongjin Kang, Jisang Yoo, Jihyeon Park, Seungtae Nam, Hyeonsoo Im, Sangheon Shin, Sangpil Kim, and Eunbyung Park. Selfsplat: Pose-free and 3d prior-free generalizable 3d gaussian splatting. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 22012–...

    2025

  11. [19]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139–1, 2023

    2023

  12. [20]

    Streamgs: Online general- izable gaussian splatting reconstruction for unposed image streams.arXiv preprint arXiv:2503.06235, 2025

    Yang Li, Jinglu Wang, Lei Chu, Xiao Li, Shiu-hong Kao, Ying-Cong Chen, and Yan Lu. Streamgs: Online general- izable gaussian splatting reconstruction for unposed image streams.arXiv preprint arXiv:2503.06235, 2025

    2025 arXiv

  13. [21]

    Efficient neural radiance fields for interactive free-viewpoint video

    Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video. InSIGGRAPH Asia 2022 Conference Papers, pages 1–9, 2022

    2022

  14. [22]

    Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision

    Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pag...

    2024

  15. [23]

    Neural rays for occlusion-aware image-based render- ing

    Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based render- ing. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 7824–7833, 2022

    2022

  16. [24]

    Scaffold-gs: Structured 3d gaussians for view-adaptive rendering

    Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654–20664, 2024

    2024

  17. [25]

    Gaussian splatting slam

    Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and An- drew J Davison. Gaussian splatting slam. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 18039–18048, 2024

    2024

  18. [26]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021

    2021

  19. [27]

    Generative densifi- cation: Learning to densify gaussians for high-fidelity gener- alizable 3d reconstruction

    Seungtae Nam, Xiangyu Sun, Gyeongjin Kang, Younggeun Lee, Seungjun Oh, and Eunbyung Park. Generative densifi- cation: Learning to densify gaussians for high-fidelity gener- alizable 3d reconstruction. InProceedings of the Computer Vision and Pattern Recognition Conference, pag...

    2025

  20. [28]

    Vigs slam: Imu-based large-scale 3d gaussian splatting slam.arXiv preprint arXiv:2501.13402, 2025

    Gyuhyeon Pak and Euntai Kim. Vigs slam: Imu-based large-scale 3d gaussian splatting slam.arXiv preprint arXiv:2501.13402, 2025

    2025 arXiv

  21. [29]

    Convolutional occupancy networks

    Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceed- ings, Part III 16, pages 523–540. Springer, 2020

    2020

  22. [30]

    Catsplat: Context-aware transformer with spatial guidance for generalizable 3d gaus- sian splatting from a single-view image.arXiv preprint arXiv:2412.12906, 2024

    Wonseok Roh, Hwanhee Jung, Jong Wook Kim, Seungg- wan Lee, Innfarn Yoo, Andreas Lugmayr, Seunggeun Chi, Karthik Ramani, and Sangpil Kim. Catsplat: Context-aware transformer with spatial guidance for generalizable 3d gaus- sian splatting from a single-view image.arXiv preprint ...

    2024 arXiv

  23. [31]

    Nerf- slam: Real-time dense monocular slam with neural radiance fields

    Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf- slam: Real-time dense monocular slam with neural radiance fields. In2023 IEEE/RSJ International Conference on Intel- ligent Robots and Systems (IROS), pages 3437–3444. IEEE, 2023

    2023

  24. [32]

    Simplere- con: 3d reconstruction without 3d convolutions

    Mohamed Sayed, John Gibson, Jamie Watson, Victor Prisacariu, Michael Firman, and Cl´ement Godard. Simplere- con: 3d reconstruction without 3d convolutions. InEuropean Conference on Computer Vision, pages 1–19. Springer, 2022

    2022

  25. [33]

    Spatialsplat: Efficient semantic 3d from sparse unposed images.arXiv preprint arXiv:2505.23044, 2025

    Yu Sheng, Jiajun Deng, Xinran Zhang, Yu Zhang, Bei Hua, Yanyong Zhang, and Jianmin Ji. Spatialsplat: Efficient semantic 3d from sparse unposed images.arXiv preprint arXiv:2505.23044, 2025

    2025

  26. [34]

    Neuralrecon: Real-time coherent 3d re- construction from monocular video

    Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d re- construction from monocular video. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15598–15607, 2021

    2021

  27. [35]

    Mm3dgs slam: Multi-modal 3d gaussian splatting for slam using vision, depth, and inertial measurements

    Lisong C Sun, Neel P Bhatt, Jonathan C Liu, Zhiwen Fan, Zhangyang Wang, Todd E Humphreys, and Ufuk Topcu. Mm3dgs slam: Multi-modal 3d gaussian splatting for slam using vision, depth, and inertial measurements. In2024 IEEE/RSJ International Conference on Intelligent Robots and ...

    2024

  28. [36]

    Hisplat: Hierarchical 3d gaus- sian splatting for generalizable sparse-view reconstruction

    Shengji Tang, Weicai Ye, Peng Ye, Weihao Lin, Yang Zhou, Tao Chen, and Wanli Ouyang. Hisplat: Hierarchical 3d gaus- sian splatting for generalizable sparse-view reconstruction. arXiv preprint arXiv:2410.06245, 2024

    2024 arXiv

  29. [37]

    Vggt: Vi- sual geometry grounded transformer

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Vi- sual geometry grounded transformer. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 5294–5306, 2025

    2025

  30. [38]

    Ibr- net: Learning multi-view image-based rendering

    Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibr- net: Learning multi-view image-based rendering. InPro- ceedings of the IEEE/CVF conference on computer vision and p...

    2021

  31. [39]

    Continuous 3d perception model with persistent state.arXiv preprint arXiv:2501.12387, 2025

    Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state.arXiv preprint arXiv:2501.12387, 2025

    2025 arXiv

  32. [40]

    Dust3r: Geometric 3d vi- sion made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vi- sion made easy. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697– 20709, 2024

    2024

  33. [41]

    Zpressor: Bottleneck-aware compression for scalable feed-forward 3dgs.arXiv preprint arXiv:2505.23734, 2025

    Weijie Wang, Donny Y Chen, Zeyu Zhang, Duochao Shi, Akide Liu, and Bohan Zhuang. Zpressor: Bottleneck-aware compression for scalable feed-forward 3dgs.arXiv preprint arXiv:2505.23734, 2025

    2025

  34. [42]

    Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes.Advances in Neural Information Processing Systems, 37:107326–107349, 2024

    Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes.Advances in Neural Information Processing Systems, 37:107326–107349, 2024

    2024

  35. [43]

    Freesplat++: Generalizable 3d gaussian splatting for efficient indoor scene reconstruction.arXiv preprint arXiv:2503.22986, 2025

    Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat++: Generalizable 3d gaussian splatting for efficient indoor scene reconstruction.arXiv preprint arXiv:2503.22986, 2025

    2025 arXiv

  36. [44]

    latentsplat: Autoencoding varia- tional gaussians for fast generalizable 3d reconstruction

    Christopher Wewer, Kevin Raj, Eddy Ilg, Bernt Schiele, and Jan Eric Lenssen. latentsplat: Autoencoding varia- tional gaussians for fast generalizable 3d reconstruction. In European Conference on Computer Vision, pages 456–473. Springer, 2024

    2024

  37. [45]

    Large-scale gaussian splatting slam.arXiv preprint arXiv:2505.09915, 2025

    Zhe Xin, Chenyang Wu, Penghui Huang, Yanyong Zhang, Yinian Mao, and Guoquan Huang. Large-scale gaussian splatting slam.arXiv preprint arXiv:2505.09915, 2025

    2025 arXiv

  38. [46]

    Depthsplat: Connecting gaussian splatting and depth.arXiv preprint arXiv:2410.13862, 2024

    Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth.arXiv preprint arXiv:2410.13862, 2024

    2024 arXiv

  39. [47]

    Instant gaussian stream: Fast and generalizable streaming of dy- namic scene reconstruction via gaussian splatting

    Jinbo Yan, Rui Peng, Zhiyan Wang, Luyang Tang, Jiayu Yang, Jie Liang, Jiahao Wu, and Ronggang Wang. Instant gaussian stream: Fast and generalizable streaming of dy- namic scene reconstruction via gaussian splatting. InPro- ceedings of the Computer Vision and Pattern Recognitio...

    2025

  40. [48]

    Slam meets nerf: A survey of implicit slam methods.World Electric Vehicle Journal, 15(3):85, 2024

    Kaiyun Yang, Yunqi Cheng, Zonghai Chen, and Jikai Wang. Slam meets nerf: A survey of implicit slam methods.World Electric Vehicle Journal, 15(3):85, 2024

    2024

  41. [49]

    Featuren- erf: Learning generalizable nerfs by distilling foundation models

    Jianglong Ye, Naiyan Wang, and Xiaolong Wang. Featuren- erf: Learning generalizable nerfs by distilling foundation models. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 8962–8973, 2023

    2023

  42. [50]

    pixelnerf: Neural radiance fields from one or few images

    Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4578–4587, 2021

    2021

  43. [51]

    Mip-splatting: Alias-free 3d gaussian splat- ting

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splat- ting. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 19447–19456, 2024

    2024

  44. [52]

    Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers

    Chuanrui Zhang, Yingshuang Zou, Zhuoling Li, Minmin Yi, and Haoqian Wang. Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers. InProceedings of the AAAI Conference on Artificial Intelli- gence, volume 39, pages 9869–9877, 2025

    2025

  45. [53]

    Gps- gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis

    Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps- gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. InProceedings of the IEEE/CVF conference on computer vision and pattern rec...

    2024

  46. [54]

    Nicer-slam: Neural implicit scene encoding for rgb slam

    Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R Oswald, Andreas Geiger, and Marc Pollefeys. Nicer-slam: Neural implicit scene encoding for rgb slam. In 2024 International Conference on 3D Vision (3DV), pages 42–52. IEEE, 2024

    2024

  47. [55]

    Nice-slam: Neural implicit scalable encoding for slam

    Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hu- jun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Polle- feys. Nice-slam: Neural implicit scalable encoding for slam. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 12786–12796, 2022

    2022

  48. [56]

    Long-lrm: Long- sequence large reconstruction model for wide-coverage gaussian splats.arXiv preprint arXiv:2410.12781, 2024

    Chen Ziwen, Hao Tan, Kai Zhang, Sai Bi, Fujun Luan, Yi- cong Hong, Li Fuxin, and Zexiang Xu. Long-lrm: Long- sequence large reconstruction model for wide-coverage gaussian splats.arXiv preprint arXiv:2410.12781, 2024

    2024 arXiv

Pith tools

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