Pith. sign in

REVIEW 4 minor 71 references

Decoupling geometry from appearance into asymmetric branches removes the redundant cost of long-sequence 3D Gaussian reconstruction while matching optimization-based quality at nearly 800 imes speed.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

An asymmetric geometry-appearance architecture for generalizable 3DGS reallocates computation so smaller models match optimization-based NVS quality at ~800× speedup on 32-view 960P inputs while improving zero-shot results.

T0 review reviewed 2026-07-14 challenge →

load-bearing objection Solid efficiency architecture for long-sequence 3DGS: task-aware asymmetry works, numbers hold, soft spots are minor.

arxiv 2607.10995 v1 pith:735EYUPL submitted 2026-07-13 cs.CV

AsySplat: Efficient Asymmetric 3D Gaussian Splatting for Long-Sequence Scene Modeling

Yingji Zhong , Dave Zhenyu Chen , Fuzhao Ou , Youyu Chen , Zhihao Li , Lanqing Hong , Dan Xu This is my paper
classification cs.CV
keywords 3D Gaussian Splattingnovel view synthesisgeneralizable reconstructionasymmetric architecturelong-sequence modelingparameter efficiencyMambasparse attention
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Generalizable 3D Gaussian Splatting models can already reconstruct long, high-resolution image sequences into novel views without per-scene optimization, but they waste computation by treating every token the same way. The paper argues that this waste is unnecessary: high-fidelity rendering does not require millimeter-precise geometry, and once geometry is known, appearance attributes are comparatively easy to predict from pixel-aligned color. AsySplat therefore splits the network into two unequal branches. The geometry branch works on coarse tokens and receives the bulk of the parameters so it can still perform hard multi-view matching; the appearance branch works on fine tokens with only about 10 % of the parameters and never performs expensive inter-frame interactions. Bilateral connections let the branches exchange depth cues and high-frequency detail. The resulting compact model (98 M parameters) matches classic optimization-based 3DGS quality on 32-view 960P inputs while running almost 800 times faster, and it outperforms the previous best feed-forward model on zero-shot benchmarks with lower training time, fewer FLOPs, half the Gaussians, and faster post-refinement.

Core claim

Computational redundancy in long-sequence Gaussian reconstruction models can be removed by an asymmetric two-branch design that allocates most parameters and inter-frame computation to coarse-grained geometry recovery while confining fine-grained appearance modeling to a lightweight intra-frame pathway; the two branches communicate bilaterally, so smaller models still achieve strong novel-view quality and superior zero-shot transfer.

What carries the argument

AsySplat’s asymmetric architecture: a three-stage geometry branch on large-patch tokens (majority of parameters, bidirectional Mamba + sparse attention) paired with a slim appearance branch on small-patch tokens (≈10 % parameters, intra-frame MambaVision only), linked by bilateral cross-attention that supplies mutual guidance before the final Gaussian head.

Load-bearing premise

The claim rests on the premise that coarse geometry is already good enough for high-quality novel-view synthesis, so the geometry branch can safely operate on large patches and still leave most capacity free for multi-view matching.

What would settle it

Measure rendering quality (PSNR/SSIM/LPIPS) on a scene whose fine surface detail is essential for correct alpha compositing (e.g., thin foliage or specular edges) when the geometry branch is forced to use only coarse tokens versus when it is given fine tokens; a clear quality collapse under the coarse setting would falsify the central premise.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Feed-forward models for long high-resolution sequences can be made substantially smaller and faster without sacrificing quality by deliberately under-allocating capacity to appearance.
  • Training and inference budgets for generalizable 3DGS drop by tens of percent simply by keeping fine tokens out of inter-frame attention.
  • Post-optimization becomes cheaper because half as many Gaussians need to be refined, allowing more iterations inside the same wall-clock budget.
  • The same asymmetry can be reused for sparse-view settings, where the model already matches or exceeds larger specialized baselines with far fewer parameters.

Where Pith is reading between the lines

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

  • If coarse geometry really is sufficient, future reconstruction models may deliberately discard high-resolution depth supervision once a coarse multi-view scaffold is obtained, further cutting training cost.
  • The bilateral-connection pattern is a general recipe for any task that mixes a hard global matching problem with an easy local regression problem; the same split could apply to multi-view stereo plus material estimation or to video depth plus optical-flow refinement.
  • Because the appearance branch never sees inter-frame tokens, scaling to hundreds of views becomes almost free in that branch, suggesting a natural path to city-scale feed-forward reconstruction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

0 major / 4 minor

Summary. The paper introduces AsySplat, a dual-branch generalizable 3D Gaussian Splatting architecture for long-sequence, high-resolution novel view synthesis. Motivated by the observations that high-precision geometry is unnecessary for high-quality NVS (due to alpha compositing) and that appearance attributes are easier to model once geometry is available, it decouples the tasks asymmetrically: a geometry branch processes coarse tokens (patch size 16) with the majority of parameters and full inter-frame interactions (bidirectional Mamba + sparse grid attention across three hierarchical stages), while a lightweight appearance branch processes fine tokens (patch size 8) with only intra-frame MambaVision layers (~10% of parameters). Bilateral cross-attention connections enable mutual guidance, followed by a merge (Eq. 4) for pixel-aligned Gaussian prediction. On 32-view 540 imes960 inputs the 98 M model matches optimization-based 3DGS quality at nearly 800 imes speedup and surpasses LongLRM zero-shot across DL3DV-140, Tanks & Temples, Mip-NeRF 360, Deep Blending and Replica while cutting parameters by ~30 %, training time by ~30 %, inference FLOPs by ~20 %, Gaussians by 50 % and post-optimization time by ~60 %.

Significance. If the reported efficiency-quality trade-off holds, AsySplat supplies a concrete, task-aware recipe for removing computational redundancy in generalizable 3DGS under realistic long-sequence high-resolution regimes, enabling smaller models to remain competitive. The work is strengthened by extensive external-benchmark comparisons (Tables 2–5) against both optimization-based baselines and the released LongLRM checkpoint, by progressive-resolution training measurements (Table 1), and by targeted ablations (Table 6) that isolate prediction source, bilateral links, Mamba/attention ratio, sparse-attention convolutions and parameter allocation. Geometry metrics and qualitative depth maps (Fig. 5, Table 6c) directly support the key observation that coarse geometry suffices for competitive rendering. These elements make the efficiency claims falsifiable and the design reusable for future long-sequence reconstruction models.

minor comments (4)
  1. [Figure 1] Figure 1 caption contains several typos (“Redudant”, “Param(M)”, inconsistent arrow notation) that should be cleaned for camera-ready.
  2. [Sec. 3.2] In Sec. 3.2 the notation for token shapes mixes V·H_pc·W_pc with (V HW / p_c^{2}); a single consistent convention would improve readability.
  3. [Table 6c] Table 6c reports GFLOPs for the single-branch baseline that appear roughly 3 imes higher than the asymmetric design; a short note clarifying whether the measurement includes only the network or also the subsequent Gaussian rasterization would help.
  4. [Sec. 4.2 / Supp. A] The supplementary material mentions monocular depth supervision (Depth Anything v2) with λ_d = 0.01; stating the exact down-sampling factor relative to p_c/p_f in the main text would make the geometry-supervision claim self-contained.

Circularity Check

0 steps flagged

No significant circularity: empirical architecture paper whose claims rest on measured external-benchmark results rather than any self-referential derivation.

full rationale

AsySplat is a systems/architecture paper. Its two motivating observations (coarse geometry suffices for NVS; appearance is easier once geometry is known) are stated as empirical priors drawn from prior surface-reconstruction literature and then directly stress-tested in the authors’ own ablations (Table 6c, Fig. 5). The subsequent design choices (patch-size asymmetry, parameter split, bilateral fusion, sparse attention) are engineering decisions whose consequences are quantified by standard metrics (PSNR/SSIM/LPIPS, FLOPs, wall-clock time, parameter count) on held-out public datasets (DL3DV-140, Tanks & Temples, Mip-NeRF 360, Deep Blending, Replica, RealEstate10K) and against independently released LongLRM checkpoints. No equation equates a claimed “prediction” to a fitted constant by construction, no uniqueness theorem is imported from the authors’ prior work, and self-citations appear only in ordinary related-work context. The central efficiency-quality claim is therefore falsifiable by external measurement and does not reduce to its own inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central efficiency claim rests on two domain observations treated as design axioms, a set of hand-chosen architectural hyper-parameters, and the invented dual-branch + sparse-attention machinery. No free parameters are fitted to the final test metrics; all numbers are chosen a priori or by progressive training schedule.

free parameters (4)
  • geometry/appearance patch sizes (pc, pf) = 16 / 8
    Set to 16 and 8 by design; control token granularity and final Gaussian count.
  • parameter allocation ratio (geometry vs appearance) = ~90/10
    Approximately 90 % / 10 % chosen to match the claimed difficulty asymmetry; ablated but not optimized on test sets.
  • stage widths and layer counts (ds, Nb, Na, Nv, pg) = {192,512,896}; 4+4; 2; 2
    Hierarchical dimensions {192,512,896}, 4 Mamba + 4 sparse-attn per geometry stage, 2 MambaVision per appearance stage, grid size 2; selected for capacity/efficiency trade-off.
  • loss weights λp, λd and progressive learning rates = 0.5 / 0.01; 4e-4 → 4e-5
    Perceptual and monocular-depth weights plus stage-wise peak LRs; standard schedule choices.
axioms (4)
  • domain assumption High-precision geometry is not strictly required for high-quality 3DGS novel-view synthesis because alpha compositing can compensate for modest depth error.
    Stated as observation (i) in Sec. 3.1 and used to justify coarse tokens for the geometry branch; supported by surface-reconstruction literature but not proved for all scenes.
  • domain assumption Once multi-view depth is known, appearance attributes are easier to predict and need only intra-frame interactions.
    Observation (ii) in Sec. 3.1; motivates the slim appearance branch and the restriction to per-frame MambaVision.
  • ad hoc to paper Bidirectional Mamba plus sparse grid attention can recover sufficient multi-view correspondences for coarse depth when given most of the parameters.
    Architectural premise of the three-stage geometry pathway (Eq. 1 and Sec. 3.3); validated only by ablation.
  • domain assumption Standard 3D Gaussian Splatting representation and pixel-aligned unprojection are adequate scene models.
    Inherited from Kerbl et al. and subsequent generalizable 3DGS literature; used throughout.
invented entities (2)
  • AsySplat dual-branch architecture with bilateral connection module no independent evidence
    purpose: Decouple and reallocate computation between geometry and appearance for long-sequence inputs.
    Core contribution; no independent existence outside this paper.
  • Sparse multi-view grid attention module no independent evidence
    purpose: Reduce quadratic cost of attention on coarse tokens while preserving context via preceding convolutions.
    Extension of MaxViT-style grid attention to multi-view sequences; introduced and ablated here.

reviewed 2026-07-14 · how reviews work

0 comments
Cite this review

Pith. "Pith review of AsySplat: Efficient Asymmetric 3D Gaussian Splatting for Long-Sequence Scene Modeling." pith.science (2026). https://pith.science/paper/735EYUPL

@misc{pith2026260710995,
  author       = {Pith},
  title        = {Pith review of: AsySplat: Efficient Asymmetric 3D Gaussian Splatting for Long-Sequence Scene Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/735EYUPL}},
  note         = {Machine review of arXiv:2607.10995}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recent generalizable 3D Gaussian Splatting models have advanced long-sequence novel view synthesis (NVS), but at the cost of substantial redundant computation. We identify that the redundancy can be mitigated based on two observations: (i) high-precision geometry is not strictly required for high-quality NVS; (ii) appearance learning is generally easier than geometry recovery. Motivated by these insights, we propose an asymmetric architecture that decouples geometry and appearance modeling. The geometry branch processes coarse-grained tokens with most of the parameters for multi-view reconstruction, while the appearance branch operates on fine-grained tokens to capture details using significantly fewer parameters. The two branches interact through bilateral connections, enabling mutual guidance for their respective tasks. This task-aware asymmetry reduces the computational redundancy and allocates the computation more judiciously, thereby increasing parameter efficiency and enabling smaller models to achieve strong performance. On 32-view 960P inputs, our model matches optimization-based methods while delivering nearly 800x speedup, and surpasses the zero-shot performance of state-of-the-art generalizable models with markedly fewer parameters and reduced training/inference overhead, achieving an overall efficiency improvement.

Figures

Figures reproduced from arXiv: 2607.10995 by Dan Xu, Dave Zhenyu Chen, Fuzhao Ou, Lanqing Hong, Yingji Zhong, Youyu Chen, Zhihao Li.

Figure 1
Figure 1. Figure 1: Our AsySplat reduces computational redundancy in the long-sequence Gaussian reconstruc [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Framework overview of AsySplat. For Gaussian reconstruction from long-sequence inputs, AsySplat reallocates computation by decoupling the modeling of geometry and appearance into two branches, aligning each with its precision needs and learning difficulty. The geometry branch processes coarse-grained tokens through three hierarchical stages with progressively increasing network capacity, leveraging bidirec… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the sparse attention module. The input token sequence is reshaped into 2D [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison with LongLRM. Given 32 input views at a resolution of 540 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Rendering and depth prediction results of test views from AsySplat. The model demonstrates [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

71 extracted references · 12 linked inside Pith

  1. [1]

    Neural rgb-d surface reconstruction

    Dejan Azinovi´c, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. InCVPR, 2022

    2022

  2. [2]

    Mip-nerf: A multiscale representation for anti-aliasing neural 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 neural radiance fields. InICCV, 2021

    2021

  3. [3]

    Mip-nerf 360: Unbounded anti-aliased neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. InCVPR, 2022

    2022

  4. [4]

    Zip-nerf: Anti-aliased grid-based neural radiance fields.ICCV, 2023

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields.ICCV, 2023

    2023

  5. [5]

    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. InCVPR, 2024

    2024

  6. [6]

    Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo

    Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. InICCV, 2021

    2021

  7. [7]

    Tensorf: Tensorial radiance fields

    Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. InECCV, 2022

    2022

  8. [8]

    Lara: Efficient large-baseline radiance fields

    Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields. InECCV, 2024

    2024

  9. [9]

    Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction.IEEE Transactions on Visualization and Computer Graphics, 2024

    Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction.IEEE Transactions on Visualization and Computer Graphics, 2024

    2024

  10. [10]

    Quantifying and alleviating co-adaptation in sparse-view 3d gaussian splatting.arXiv preprint arXiv:2508.12720, 2025

    Kangjie Chen, Yingji Zhong, Zhihao Li, Jiaqi Lin, Youyu Chen, Minghan Qin, and Haoqian Wang. Quantifying and alleviating co-adaptation in sparse-view 3d gaussian splatting.arXiv preprint arXiv:2508.12720, 2025

    arXiv 2025

  11. [11]

    Photographic image synthesis with cascaded refinement networks

    Qifeng Chen and Vladlen Koltun. Photographic image synthesis with cascaded refinement networks. InICCV, 2017

    2017

  12. [12]

    Training deep nets with sublinear memory cost.arXiv preprint arXiv:1604.06174, 2016

    Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost.arXiv preprint arXiv:1604.06174, 2016

    Pith/arXiv arXiv 2016

  13. [13]

    Dashgaussian: Optimizing 3d gaussian splatting in 200 seconds

    Youyu Chen, Junjun Jiang, Kui Jiang, Xiao Tang, Zhihao Li, Xianming Liu, and Yinyu Nie. Dashgaussian: Optimizing 3d gaussian splatting in 200 seconds. InCVPR, 2025

    2025

  14. [14]

    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. InECCV, 2024

    2024

  15. [15]

    Mvsplat360: Feed-forward 360 scene synthesis from sparse views

    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. InNeurIPS, 2024

    2024

  16. [16]

    Flashattention-2: Faster attention with better parallelism and work partitioning.arXiv preprint arXiv:2307.08691, 2023

    Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning.arXiv preprint arXiv:2307.08691, 2023

    Pith/arXiv arXiv 2023

  17. [17]

    Slowfast networks for video recognition

    Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. InICCV, 2019

    2019

  18. [18]

    Colmap- free 3d gaussian splatting

    Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap- free 3d gaussian splatting. InCVPR, 2024

    2024

  19. [19]

    Gaussianflow: Splatting gaussian dynamics for 4d content creation.arXiv preprint arXiv:2403.12365, 2024

    Quankai Gao, Qiangeng Xu, Zhe Cao, Ben Mildenhall, Wenchao Ma, Le Chen, Danhang Tang, and Ulrich Neumann. Gaussianflow: Splatting gaussian dynamics for 4d content creation.arXiv preprint arXiv:2403.12365, 2024. 11

    Pith/arXiv arXiv 2024

  20. [20]

    Mamba: Linear-time sequence modeling with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. In First conference on language modeling, 2024

    2024

  21. [21]

    Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering

    Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. InCVPR, 2024

    2024

  22. [22]

    Mambavision: A hybrid mamba-transformer vision backbone

    Ali Hatamizadeh and Jan Kautz. Mambavision: A hybrid mamba-transformer vision backbone. InCVPR, 2025

    2025

  23. [23]

    Deep blending for free-viewpoint image-based rendering.ACM Transactions on Graphics (ToG), 37(6):1–15, 2018

    Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering.ACM Transactions on Graphics (ToG), 37(6):1–15, 2018

    2018

  24. [24]

    Lrm: Large reconstruction model for single image to 3d

    Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023

    Pith/arXiv arXiv 2023

  25. [25]

    2d gaussian splatting for geometrically accurate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. InACM SIGGRAPH 2024 conference papers, 2024

    2024

  26. [26]

    Anysplat: Feed-forward 3d gaussian splatting from unconstrained views.arXiv preprint arXiv:2505.23716, 2025

    Lihan Jiang, Yucheng Mao, Linning Xu, Tao Lu, Kerui Ren, Yichen Jin, Xudong Xu, Mulin Yu, Jiangmiao Pang, Feng Zhao, et al. Anysplat: Feed-forward 3d gaussian splatting from unconstrained views.arXiv preprint arXiv:2505.23716, 2025

    arXiv 2025

  27. [27]

    Perceptual losses for real-time style transfer and super-resolution

    Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. InECCV, 2016

    2016

  28. [28]

    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. InCVPR, 2025

    2025

  29. [29]

    ilrm: An iterative large 3d reconstruction model

    Gyeongjin Kang, Seungtae Nam, Seungkwon Yang, Xiangyu Sun, Sameh Khamis, Abdelrahman Mohamed, and Eunbyung Park. ilrm: An iterative large 3d reconstruction model. InCVPR, 2026

    2026

  30. [30]

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

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

    2023

  31. [31]

    A hierarchical 3d gaussian representation for real-time rendering of very large datasets.ACM Transactions on Graphics (TOG), 43(4):1–15, 2024

    Bernhard Kerbl, Andreas Meuleman, Georgios Kopanas, Michael Wimmer, Alexandre Lanvin, and George Drettakis. A hierarchical 3d gaussian representation for real-time rendering of very large datasets.ACM Transactions on Graphics (TOG), 43(4):1–15, 2024

    2024

  32. [32]

    Tanks and temples: Bench- marking large-scale scene reconstruction.ACM Transactions on Graphics (ToG), 36(4):1–13, 2017

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Bench- marking large-scale scene reconstruction.ACM Transactions on Graphics (ToG), 36(4):1–13, 2017

    2017

  33. [33]

    Vastgaussian: Vast 3d gaussians for large scene reconstruction

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

    2024

  34. [34]

    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. InCVPR, 2024

    2024

  35. [35]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    Pith/arXiv arXiv 2017

  36. [36]

    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. InCVPR, 2024

    2024

  37. [37]

    Taming 3dgs: High-quality radiance fields with limited resources.arXiv preprint arXiv:2406.15643, 2024

    Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Francisco Vicente Carrasco, Markus Steinberger, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources.arXiv preprint arXiv:2406.15643, 2024. 12

    Pith/arXiv arXiv 2024

  38. [38]

    Nerf: Representing scenes as neural radiance fields for view synthesis

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

    2020

  39. [39]

    Instant neural graphics primitives with a multiresolution hash encoding.ACM Transactions on Graphics (ToG), 41(4): 1–15, 2022

    Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding.ACM Transactions on Graphics (ToG), 41(4): 1–15, 2022

    2022

  40. [40]

    D-nerf: Neural radiance fields for dynamic scenes

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

    2021

  41. [41]

    The replica dataset: A digital replica of indoor spaces.arXiv preprint arXiv:1906.05797, 2019

    Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces.arXiv preprint arXiv:1906.05797, 2019

    Pith/arXiv arXiv 1906

  42. [42]

    Splatter image: Ultra-fast single-view 3d reconstruction

    Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. InCVPR, 2024

    2024

  43. [43]

    Lgm: Large multi-view gaussian model for high-resolution 3d content creation

    Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. InECCV, 2024

    2024

  44. [44]

    Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timo- thée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023

    Pith/arXiv arXiv 2023

  45. [45]

    Maxvit: Multi-axis vision transformer

    Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. InECCV, 2022

    2022

  46. [46]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. InNeurIPS, 2017

    2017

  47. [47]

    Vggt: Visual geometry grounded transformer

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. InCVPR, 2025

    2025

  48. [48]

    Ibrnet: 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. Ibrnet: Learning multi-view image-based rendering. InCVPR, 2021

    2021

  49. [49]

    Hyrf: Hybrid radiance fields for memory-efficient and high-quality novel view synthesis.arXiv preprint arXiv:2509.17083, 2025

    Zipeng Wang and Dan Xu. Hyrf: Hybrid radiance fields for memory-efficient and high-quality novel view synthesis.arXiv preprint arXiv:2509.17083, 2025

    arXiv 2025

  50. [50]

    Gs2mesh: Surface reconstruction from gaussian splatting via novel stereo views

    Yaniv Wolf, Amit Bracha, and Ron Kimmel. Gs2mesh: Surface reconstruction from gaussian splatting via novel stereo views. InECCV, 2024

    2024

  51. [51]

    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. In CVPR, 2024

    2024

  52. [52]

    Depthsplat: Connecting gaussian splatting and depth

    Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. InCVPR, 2025

    2025

  53. [53]

    Freesplatter: Pose-free gaussian splatting for sparse- view 3d reconstruction

    Jiale Xu, Shenghua Gao, and Ying Shan. Freesplatter: Pose-free gaussian splatting for sparse- view 3d reconstruction. InICCV, 2025

    2025

  54. [54]

    Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation

    Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. InECCV, 2024

    2024

  55. [55]

    Depth anything v2

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. InNeurIPS, 2024

    2024

  56. [56]

    No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images

    Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207, 2024. 13

    Pith/arXiv arXiv 2024

  57. [57]

    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. InCVPR, 2021

    2021

  58. [58]

    Wonder- world: Interactive 3d scene generation from a single image

    Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T Freeman, and Jiajun Wu. Wonder- world: Interactive 3d scene generation from a single image. InCVPR, 2025

    2025

  59. [59]

    Mip-splatting: Alias-free 3d gaussian splatting

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. InCVPR, 2024

    2024

  60. [60]

    Root mean square layer normalization

    Biao Zhang and Rico Sennrich. Root mean square layer normalization. InNeurIPS, 2019

    2019

  61. [61]

    Cor-gs: sparse-view 3d gaussian splatting via co-regularization

    Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. Cor-gs: sparse-view 3d gaussian splatting via co-regularization. InECCV, 2024

    2024

  62. [62]

    Arf: Artistic radiance fields

    Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. InECCV, 2022

    2022

  63. [63]

    Gs-lrm: Large reconstruction model for 3d gaussian splatting

    Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. InECCV, 2024

    2024

  64. [64]

    Switch-nerf: Learning scene decomposition with mixture of experts for large-scale neural radiance fields

    MI Zhenxing and Dan Xu. Switch-nerf: Learning scene decomposition with mixture of experts for large-scale neural radiance fields. InICLR, 2022

    2022

  65. [65]

    Cvt-xrf: Contrastive in-voxel trans- former for 3d consistent radiance fields from sparse inputs

    Yingji Zhong, Lanqing Hong, Zhenguo Li, and Dan Xu. Cvt-xrf: Contrastive in-voxel trans- former for 3d consistent radiance fields from sparse inputs. InCVPR, 2024

    2024

  66. [66]

    Taming video diffusion prior with scene-grounding guidance for 3d gaussian splatting from sparse inputs

    Yingji Zhong, Zhihao Li, Dave Zhenyu Chen, Lanqing Hong, and Dan Xu. Taming video diffusion prior with scene-grounding guidance for 3d gaussian splatting from sparse inputs. In CVPR, 2025

    2025

  67. [67]

    Empowering sparse-input neural radiance fields with dual-level semantic guidance from dense novel views

    Yingji Zhong, Kaichen Zhou, Zhihao Li, Lanqing Hong, Zhenguo Li, and Dan Xu. Empowering sparse-input neural radiance fields with dual-level semantic guidance from dense novel views. arXiv preprint arXiv:2503.02230, 2025

    Pith/arXiv arXiv 2025

  68. [68]

    Stereo magnifi- cation: Learning view synthesis using multiplane images.arXiv preprint arXiv:1805.09817, 2018

    Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnifi- cation: Learning view synthesis using multiplane images.arXiv preprint arXiv:1805.09817, 2018

    Pith/arXiv arXiv 2018

  69. [69]

    Vision mamba: Efficient visual representation learning with bidirectional state space model

    Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024

    Pith/arXiv arXiv 2024

  70. [70]

    Fsgs: Real-time few-shot view synthesis using gaussian splatting

    Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. InECCV, 2024

    2024

  71. [71]

    Long-lrm: Long-sequence large reconstruction model for wide-coverage gaussian splats

    Chen Ziwen, Hao Tan, Kai Zhang, Sai Bi, Fujun Luan, Yicong Hong, Li Fuxin, and Zexiang Xu. Long-lrm: Long-sequence large reconstruction model for wide-coverage gaussian splats. InICCV, 2025. 14 AsySplat: Efficient Asymmetric 3D Gaussian Splatting for Long-Sequence Scene Modeling Supplementary Material A Implementation Details Architecture.All transformer ...

    2025

This paper was first reviewed by grok-4.5 on July 14, 2026.