G$^2$ARD-GS: Geometry-Guided Anchor-Regularized Gaussian Splatting Distillation

Jianming Huang; Puyuan Zhang; Wei Dong; Wenkai Ye

REVIEW 3 major objections 6 minor 40 references

G$^2$ARD-GS: Geometry-Guided Anchor-Regularized Gaussian Splatting Distillation

T0 review · 3 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Gaussian compression that keeps surfaces at 30x

desk verdict Honest, internally consistent compression paper whose headline margins are plausibly real but confounded by unmatched optimization compute; worth a serious referee, with a compute-matched comparison as the key ask. read the letter →

arxiv 2608.05704 v1 pith:I2JBT766 submitted 2026-08-06 cs.CV

Puyuan Zhang , Jianming Huang , Wenkai Ye , Wei Dong This is my paper · ORCID

classification cs.CV

keywords 3DGaussianSplattingmapcompressionLiDARpriordistillationanchorregularizationsurface-awaresimplificationviewselectioncameraregistration

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

This paper tries to show that aggressively compressed 3D Gaussian maps can keep both their rendering quality and their geometric usefulness when compression is guided by surface structure rather than image-space fidelity alone. The proposed method, G2ARD-GS, progressively simplifies a dense Gaussian prior, freezes the surviving primitives' surface-aligned frames as fixed anchors, and then recovers appearance without adding or removing primitives. On MatrixCity block A it reports the best PSNR, SSIM, and LPIPS across 5x to 30x compression budgets, reaching 21.21 dB at 30x versus 17.99 dB for the PUP baseline. The same compact geometry, with only appearance refit, is said to support out-of-distribution novel-view adaptation and to preserve camera-registration accuracy, indicating that compression need not sacrifice reusability.

What carries the argument

The load-bearing mechanism is the construction-time anchor: for each retained Gaussian, the mean, local orthonormal frame obtained from PCA over the dense prior, and scales are frozen after simplification, and recovery is penalized only when a primitive leaves an anisotropic trust region defined by those anchors. This couples appearance optimization to the surface support established during compression, preventing the off-surface drift and needle-shaped degeneration identified as failure modes of photometric-only recovery.

What would settle it

Run G2ARD-GS at 30x compression on a LiDAR-assisted urban model that includes vegetation, cars, or glass facades, then evaluate frozen-geometry out-of-distribution adaptation and camera registration; if the compact model no longer beats image-driven pruning baselines such as PUP on those scenes, the load-bearing assumption of trustworthy PCA anchors is refuted outside planar surface regions.

Watch

Extended reading notes

Core claim

The central claim is that a compact 3D Gaussian map can be both visually faithful and geometrically reusable if simplification respects local surface support and recovery stays attached to that support. G2ARD-GS consolidates a dense prior into surface-aware representatives through progressive rounds, protecting persistent high-frequency texture, then freezes each retained primitive's construction-time mean, local frame, and scale as an anchor. Appearance recovery minimizes a photometric loss under an anisotropic anchor trust region that allows tangential sliding but penalizes off-surface drift, plus a shape regularizer that suppresses needle-like covariance degeneration. On MatrixCity block A this yields the best held-out PSNR, SSIM, and LPIPS at matched 5x to 30x budgets, with a 3.2 to 4.9 dB advantage over PUP in frozen-geometry out-of-distribution adaptation, and a 30.3x compact model preserves GS-CPR registration accuracy on Cambridge KingsCollege while lowering median translation error.

Load-bearing premise

The method assumes that local PCA over the dense prior yields reliable surface anchors, so freezing those anchors preserves the geometry that later appearance recovery should build on; if the prior contains non-planar or noisy structure, the trust region can lock in misplaced support that appearance optimization cannot fix.

Editorial extensions

If this is right

At matched 5x to 30x primitive budgets on MatrixCity block A, G2ARD-GS claims the best PSNR, SSIM, and LPIPS across all baselines, with a 3.22 dB PSNR lead over PUP at 30x.
Frozen-geometry appearance adaptation to a disjoint out-of-distribution trajectory improves by 3.68 to 4.93 dB over PUP, and degrading from 10x to 30x costs only about 1.2 dB in OOD PSNR.
A 30.3x compact model preserves or slightly exceeds teacher-level GS-CPR camera-registration accuracy on KingsCollege, also reducing median translation error from 27.0 cm to 22.4 cm.
The method works from a training-free point-cloud lift as well as from a trained Gaussian model, with distilled students from the two priors agreeing to within 0.01 dB.
The fixed recipe transfers to bounded 3DGS scenes such as Mip-NeRF 360 room, giving the smallest teacher-relative PSNR drop among compared methods.

Reading between the lines

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

The anchor trust region effectively trades geometric fidelity for appearance flexibility; scenes whose dense prior is non-planar or noisy (vegetation, cars, glass) may lock in misplaced surface support that appearance optimization cannot correct, a regime the paper does not test.
Because the view-selection module shows no advantage over random sampling at matched budgets in the supplement, its practical role is likely cost reduction rather than higher quality, and combining it with attribute-level compression such as vector quantization could compound storage savings.
The frozen-anchor design suggests the compact model could serve as a stable substrate for continual appearance updates, since repeated refitting of appearance alone cannot erode the underlying geometry.

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.

Desk Editor's Note

read the letter

You should know two things about this one. The positive result is real on its own terms: on MatrixCity block A, with a common 5.99M-Gaussian teacher, matched primitive budgets, a unified gsplat evaluator, and held-out views, the method beats PUP by 3.2–6.8 dB across 5×–30× and holds up under frozen-geometry adaptation on a disjoint OOD trajectory. The supplement is unusually honest: it reports a matched-budget negative result for its own view-selection module, discloses the compute mismatch, and gives per-run configurations. This is not a paper that hides its warts.

What is actually new is the combination: progressive multi-round simplify-and-distill with persistent high-frequency protection, moment-matched merging, anchor-regularized recovery on fixed topology, and geometry-aware supervision. Every ingredient exists in the cited literature, but the multi-round schedule plus frozen construction-time anchors as a reuse mechanism is new. The point-cloud-lift result is a genuinely nice finding: a training-free 3M-primitive lift distills to a student within 0.01 dB of one distilled from a trained teacher, which suggests the method mainly consumes the per-primitive surface frames.

The big soft spot is optimization compute. G²ARD-GS runs roughly 126k gradient steps per configuration (8k per round plus 30k final) while baselines run their released schedules. The paper discloses this in Sec. 4.1 and Supp. Table S5, and the limitation section names it, but the abstract's 'leads all baselines' framing does not carry the caveat. The 3.2–6.8 dB margins are exactly the kind of gap that extra distillation rounds can produce. The one-shot ablation in Table 3b matches optimizer steps and still shows a 0.92 dB gain for the multi-round schedule, which partially answers the concern, but that control is on the PCLift prior at 30× with K50 supervision, not the headline trained-teacher sweep. A compute-matched comparison, or at least giving PUP and LightGaussian comparable step budgets, is the main thing I would want before taking the headline margins at face value.

Second, the primary evidence is one urban scene. Mip-NeRF 360 transfer is mixed, and the anchor trust region's assumption — that local PCA frames from the dense prior are reliable surface anchors — is untested on vegetation, cars, and glass, where non-planar structure would get locked in. Legitimate generalization question, not an internal flaw. Third, minor: no code is released. The config tables are detailed enough that reproduction is plausible, but a multi-round schedule with hand-set per-run regularizer weights is what code would settle.

Also worth noting: Table S8 shows their coverage-driven view selection is not better than random at matched budget. They say so plainly and reposition the module as a budget-reduction tool. That is good practice, but it does mean the 'three coupled requirements' framing in the introduction overstates the role of view selection.

Bottom line: solid, honest methods paper with a real evaluation-fairness question attached. I would send it to review; the referees should push the compute-matched comparison and the non-planar scene question, not hunt for an internal contradiction. I would bring it to a reading group — the supplement is a good case study in reporting what did not work.

Referee Report

3 major / 6 minor

Summary. The manuscript introduces G2ARD-GS, a multi-round pipeline for compressing dense 3D Gaussian maps into compact, geometrically reusable representations. Starting from either a trained Gaussian model or a training-free point-cloud lift, it progressively simplifies the prior through moment-matched merging with high-frequency protection, selects supervision views via a normal-aware directional-coverage objective, and then recovers appearance on a fixed topology under shape and anisotropic anchor regularizers. On MatrixCity block A, the method reports the best PSNR, SSIM, and LPIPS across matched 5x-30x primitive budgets, improved frozen-geometry appearance adaptation on a disjoint OOD trajectory, and preserved GS-CPR registration accuracy on KingsCollege. The supplement provides detailed configurations, additional ablations, and honest null results for the view-selection module.

Significance. The problem addressed is relevant and timely: producing compact, geometrically stable 3DGS assets from dense LiDAR maps is useful for storage, transmission, and downstream geometric tasks. The paper has notable strengths: all MatrixCity methods start from the same 5.99M-Gaussian teacher, use matched primitive budgets, a unified gsplat evaluator, and held-out views; the supplement is unusually transparent, including a matched-budget control that shows the proposed view-selection policy does not outperform random sampling. If the reported gains survive a compute-matched comparison, the method would be a solid contribution. However, the central head-to-head claim currently rests on an unequal-compute comparison, which the paper itself acknowledges in the Limitations section, and this undermines the strength of the 'leads all baselines' framing.

major comments (3)

[Sec. 4.1, Table 1, Limitations]
[Sec. 3.3, Table S8]
[Sec. 3.2, Sec. 4.3, Sec. 5]

minor comments (6)

[Table 2]
[Sec. 4.4, Table 3(b)]
[Sec. 4.4]
[Sec. 3.2, Eq. (1)]
[Fig. 1, Table 1]
[Table S7]

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central compression and reuse claims are measured against held-out views and independent downstream tasks; the unmatched-compute caveat is a fairness issue, not a circularity.

full rationale

The paper's derivation chain is self-contained rather than circular. A dense Gaussian prior is progressively simplified (Eqs. 1–2), views are selected via a coverage objective (Eq. 3), and appearance is recovered under anchor and shape regularizers (Eqs. 4–6); however, the headline results are evaluated on 510 held-out views with a unified gsplat evaluator, and the reuse claims are tested by independent downstream tasks: frozen-geometry OOD appearance adaptation on a disjoint 495-view set and GS-CPR camera registration on KingsCollege. Teacher renders are used only as distillation targets during recovery, not as the reported metrics. The paper also candidly reports in Supp. Table S8 that geometry-aware view selection does not outperform random or pose-diverse selection at matched budget, so the central results do not rest on a self-fulfilling selection mechanism. No load-bearing argument reduces to a self-citation, a fitted parameter renamed as a prediction, or an imported uniqueness theorem. The acknowledged unmatched optimization budgets (Sec. 4.1 and the Limitations paragraph) weaken the fairness of the head-to-head comparison, but they do not make any reported quantity equal to an input by construction; this is an external-validity concern, not a circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical entities or mathematical axioms beyond standard linear algebra and 3DGS rasterization. Its load-bearing assumptions are about the data (LiDAR surface normals are reliable) and about optimization (photometric loss plus regularizers can recover appearance on fixed topology). The listed free parameters are hyperparameters chosen for the reported runs, not fitted to the evaluation set; still, several of them directly control the strength of the geometry preservation that the claims rest on.

free parameters (6)

anchor loss weight lambda_a = 5e-2 (trained teachers), 1e-3 (PCLift)
Weights Lanchor in Eq. (6); chosen per prior type in Supp. Table S4.
shape loss weight lambda_s = 0 or 1e-4/1e-3
Weights Lshape in Eq. (6); active only for less regularized teachers and PCLift runs.
per-round retention ratio = 0.75
Budget schedule reduces to 75% of current primitives per round; controls how much simplification is absorbed per round.
high-frequency threshold epsilon_h = 0.01
Sets the affine-color residual threshold in Eq. (1) that flags protected texture cells.
protection weight lambda_h = 10
Boosts trimming priority of protected cells in Eq. (2).
normal compression factor for point-cloud lift = 0.1 (anisotropy clipped at 4.0)
Shapes lifted primitives as surface-aligned discs; used in Supp. Sec. 1.2.

assumptions (4)

domain assumption Local PCA over 10 nearest neighbors yields a reliable surface-aligned normal per primitive
Used in Supp. Sec. 1.2 for the point-cloud lift; anchors inherit this frame.
domain assumption Moment-matched merging of compatible Gaussians preserves local radiance behavior
Assumed in Sec. 3.2 for consolidation under affine-energy and surface-coverage guards.
domain assumption The photometric loss plus geometric regularizers is sufficient for appearance recovery on fixed topology
Implicit in Sec. 3.4; recovery only optimizes per-primitive attributes, never topology.
standard math Standard 3DGS rasterization (Kerbl et al. 2023) is a correct forward rendering model
All evaluations use the 3DGS renderer; its correctness is not re-derived.

how reviews work

0 comments

Cite this review

Pith. "Pith review of G$^2$ARD-GS: Geometry-Guided Anchor-Regularized Gaussian Splatting Distillation." pith.science (2026). https://pith.science/paper/I2JBT766

@misc{pith2026260805704,
  author       = {Pith},
  title        = {Pith review of: G$^2$ARD-GS: Geometry-Guided Anchor-Regularized Gaussian Splatting Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I2JBT766}},
  note         = {Machine review of arXiv:2608.05704}
}

abstract

Dense colored LiDAR maps provide accurate city-scale geometry, but lifting them into 3D Gaussian Splatting (3DGS) retains millions of primitives, making the resulting models costly to store, transmit, render, and adapt. Aggressive primitive reduction alleviates this burden, but can remove the local surface support needed for stable novel-view synthesis and downstream geometric use. We introduce G$^2$ARD-GS, a geometry-guided distillation method that converts a dense Gaussian prior instantiated either as a training-free point-cloud lift or a trained GS model into a compact, reusable representation. G$^2$ARD-GS progressively consolidates the prior into surface-aware representatives, then recovers appearance on the resulting fixed topology under construction-time anchor constraints, with no primitives added or removed during recovery. Under limited supervision, geometry-aware view selection allocates the available view budget. On MatrixCity, G$^2$ARD-GS achieves the best PSNR, SSIM, and LPIPS across matched $5\times$--$30\times$ compression budgets, outperforming PUP by $3.2$--$6.8$,dB in PSNR. When reused as frozen geometry, the compact model improves off-trajectory appearance adaptation by $3.7$--$4.9$,dB over PUP 3D-GS and preserves image-to-model registration accuracy on Cambridge KingsCollege at $30\times$ compression. Project page: https://patrick1159.github.io/gardGS-page/.

Figures

Figures reproduced from arXiv: 2608.05704 by the authors.

**Figure 1.** G 2ARD-GS leads all baselines at every compression ratio (MatrixCity block A, 510 held-out views). derable representation with learnable spatial footprints and appearance. This makes 3DGS a promising interface between metric maps and image-based applications such as sensor simulation, appearance adaptation, and image-to-map localization [9, 14, 23]. LiDAR-assisted Gaussian methods further demonstrate that dense poi… view at source ↗

**Figure 2.** G 2ARD-GS overview. Given a dense Gaussian prior GT , each round simplifies the current representation toward the next primitive budget while protecting high-frequency structures and freezing construction-time anchors (1). Geometry-aware informative-view selection allocates the supervision budget K to complementary observations (2). Anchor-regularized recovery transfers appearance from selected real or teacher-rende… view at source ↗

**Figure 3.** Geometry-aware mechanisms. (a) Local PCA gives each primitive an orthonormal frame (t1 ⊥ t2 ⊥ n) and a compressed normal axis, yielding a surface-aligned disc. (b) The percell affine residual Ej separates locally predictable content from texture that a smooth model cannot explain (real MatrixCity crops; grid = the scoring lattice, boxes = flagged cells); the flag softly boosts optical mass through trim-and-recover… view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Shape and position regularization. (a) The effectiverank barrier prevents needle degeneration; the thickness cap prevents the normal axis from exceeding its anchor value. (b) The anisotropic trust region permits in-plane sliding while penalizing off-surface drift. Sh…

**Figure 5.** Figure 5: Qualitative comparisons at 10×. (a–c) MatrixCity held-out and training views; (d–e) Mip-NeRF 360 garden and room. Columns show GT, Ours, PUP, NanoGS, and LightGaussian; blue boxes mark magnified regions. (−0.31 dB versus −0.48/−0.86 dB for PUP/LightGaussian) and falls …

**Figure 7.** Figure 7: Qualitative ablation at 30×. Columns show GT, oneshot, −Lshape, −Lanchor, and the full model on two MatrixCity OOD views. 4.4. Ablation and Analysis High-frequency protection. Removing the persistent protection boost and ranking by bare optical mass at 10× reduces O…

**Figure 6.** Figure 6: Frozen-geometry appearance adaptation at 30×. Only spherical harmonics and opacity are optimized. Columns show GT, Ours, PUP, NanoGS, and LightGaussian. Two OOD views are shown with per-view PSNR. Prior K50 K75 K100 Trained GS 20.92 21.70 21.92 PCLift GS 19.00 19.54 20…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 30 canonical work pages

[1]

Barron, Ben Mildenhall, Dor Verbin, Pratul P

Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5460–5469. IEEE, 2022. 5, 7

work page 2022
[2]

Accelerated coordinate encoding: Learning to relocalize in minutes using rgb and poses

Eric Brachmann, Tommaso Cavallari, and Victor Adrian Prisacariu. Accelerated coordinate encoding: Learning to relocalize in minutes using rgb and poses. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5044–5053, 2023. 8

work page 2023
[3]

HAC: Hash-grid assisted context for 3d gaussian splatting compression

Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. HAC: Hash-grid assisted context for 3d gaussian splatting compression. InEuropean Conference on Computer Vision, 2024. 2

work page 2024
[4]

LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and 200+ FPS,

Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and 200+ FPS,

work page
[5]

Mini-Splatting: Represent- ing Scenes with a Constrained Number of Gaussians

Guangchi Fang and Bing Wang. Mini-Splatting: Represent- ing Scenes with a Constrained Number of Gaussians. In Computer Vision – ECCV 2024, pages 165–181. Springer Nature Switzerland, Cham, 2024. Series Title: Lecture Notes in Computer Science. 1, 2, 3

work page 2024
[6]

EA- GLES: Efficient accelerated 3d gaussians with lightweight encodings

Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. EA- GLES: Efficient accelerated 3d gaussians with lightweight encodings. InEuropean Conference on Computer Vision,

work page
[7]

SuGaR: Surface- aligned gaussian splatting for efficient 3d mesh reconstruc- tion and high-quality mesh rendering

Antoine Gu ´edon and Vincent Lepetit. SuGaR: Surface- aligned gaussian splatting for efficient 3d mesh reconstruc- tion and high-quality mesh rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354–5363, 2024. 2

work page 2024
[8]

PUP 3D-GS: Princi- pled Uncertainty Pruning for 3D Gaussian Splatting, 2025

Alex Hanson, Allen Tu, Vasu Singla, Mayuka Jayawardhana, Matthias Zwicker, and Tom Goldstein. PUP 3D-GS: Princi- pled Uncertainty Pruning for 3D Gaussian Splatting, 2025. arXiv:2406.10219 [cs]; CVPR 2025. 1, 2, 3, 5, 7

arXiv 2025

Show all 40 references

[9]

SplatAD: Real-time Lidar and camera rendering with 3D gaussian splatting for autonomous driving

Georg Hess, Carl Lindstr¨om, Maryam Fatemi, Christoffer Pe- tersson, and Lennart Svensson. SplatAD: Real-time Lidar and camera rendering with 3D gaussian splatting for autonomous driving. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pa...

2025
[10]

2D gaussian splatting for geometrically accu- rate radiance fields

Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D gaussian splatting for geometrically accu- rate radiance fields. InACM SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 2

2024
[11]

VEGS: View Extrapolation of Urban Scenes in 3D Gaussian Splatting using Learned Priors, 2024

Sungwon Hwang, Min-Jung Kim, Taewoong Kang, Jayeon Kang, and Jaegul Choo. VEGS: View Extrapolation of Urban Scenes in 3D Gaussian Splatting using Learned Priors, 2024. arXiv:2407.02945 [cs.CV]. 1, 2

2024 arXiv
[12]

Effective Rank Analysis and Regularization for Enhanced 3D Gaussian Splatting, 2024

Junha Hyung, Susung Hong, Sungwon Hwang, Jaeseong Lee, Jaegul Choo, and Jin-Hwa Kim. Effective Rank Analysis and Regularization for Enhanced 3D Gaussian Splatting, 2024. arXiv:2406.11672 [cs.CV]. 2, 4

2024 arXiv
[13]

LI-GS: Gaussian Splatting with LiDAR Incorporated for Accurate Large-Scale Reconstruction, 2024

Changjian Jiang, Ruilan Gao, Kele Shao, Yue Wang, Rong Xiong, and Yu Zhang. LI-GS: Gaussian Splatting with LiDAR Incorporated for Accurate Large-Scale Reconstruction, 2024. arXiv:2409.12899 [cs.RO]. 1, 2

2024 arXiv
[14]

3DGS- ReLoc: 3D gaussian splatting for map representation and visual relocalization, 2024

Peng Jiang, Gaurav Pandey, and Srikanth Saripalli. 3DGS- ReLoc: 3D gaussian splatting for map representation and visual relocalization, 2024. arXiv:2403.11367 [cs.CV , cs.GR, cs.RO]. 1

2024 arXiv
[15]

Fish- erRF: Active View Selection and Uncertainty Quantifica- tion for Radiance Fields using Fisher Information, 2024

Wen Jiang, Boshu Lei, and Kostas Daniilidis. Fish- erRF: Active View Selection and Uncertainty Quantifica- tion for Radiance Fields using Fisher Information, 2024. arXiv:2311.17874 [cs.CV]. 2, 4

2024 arXiv
[16]

PoseNet: A convolutional network for real-time 6-dof cam- era relocalization

Alex Kendall, Matthew Grimes, and Roberto Cipolla. PoseNet: A convolutional network for real-time 6-dof cam- era relocalization. InProceedings of the IEEE International Conference on Computer Vision, pages 2938–2946, 2015. 2, 5

2015
[17]

3D Gaussian Splatting for Real-Time Ra- diance Field Rendering, 2023

Bernhard Kerbl, Georgios Kopanas, Thomas Leimk¨uhler, and George Drettakis. 3D Gaussian Splatting for Real-Time Ra- diance Field Rendering, 2023. arXiv:2308.04079 [cs]. 1, 2, 4

2023 arXiv
[18]

Compact 3d gaussian representation for radiance field

Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21719– 21728, 2024. 1, 2

2024
[19]

Ground- ing image matching in 3d with MASt3R

Vincent Leroy, Yohann Cabon, and J´erˆome Revaud. Ground- ing image matching in 3d with MASt3R. InEuropean Con- ference on Computer Vision, pages 72–91, 2024. 8

2024
[20]

MAGICIAN: Efficient Long-Term Planning with Imag- ined Gaussians for Active Mapping, 2026

Shiyao Li, Antoine Gu´edon, Shizhe Chen, and Vincent Lep- etit. MAGICIAN: Efficient Long-Term Planning with Imag- ined Gaussians for Active Mapping, 2026. Version Number: 2; Accepted at CVPR 2026 (Oral). 2, 4

2026
[21]

MatrixCity: A large-scale city dataset for city-scale neural rendering and beyond

Yixuan Li, Lihan Jiang, Linning Xu, Yuanbo Xiangli, Zhenzhi Wang, Dahua Lin, and Bo Dai. MatrixCity: A large-scale city dataset for city-scale neural rendering and beyond. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3205–3215, 2023. 2, 5

2023
[22]

GS-CPR: Efficient Camera Pose Refinement via 3D Gaussian Splatting, 2025

Changkun Liu, Shuai Chen, Yash Bhalgat, Siyan Hu, Ming Cheng, Zirui Wang, Victor Adrian Prisacariu, and Tristan Braud. GS-CPR: Efficient Camera Pose Refinement via 3D Gaussian Splatting, 2025. arXiv:2408.11085 [cs.CV]; Ac- cepted to ICLR 2025. 8, 3

2025 arXiv
[23]

GS-CPR: Efficient camera pose refinement via 3D gaussian splatting

Changkun Liu, Shuai Chen, Yash Bhalgat, Siyan Hu, Ming Cheng, Zirui Wang, Victor Adrian Prisacariu, and Tristan Braud. GS-CPR: Efficient camera pose refinement via 3D gaussian splatting. InThe Thirteenth International Confer- ence on Learning Representations. OpenReview.net, 2...

2025
[24]

ECLAIR: A high- fidelity aerial LiDAR dataset for semantic segmentation

Iaroslav Melekhov, Anand Umashankar, Hyeong-Jin Kim, Vladislav Serkov, and Dusty Argyle. ECLAIR: A high- fidelity aerial LiDAR dataset for semantic segmentation. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 7627–7637. IEEE,

2024
[25]

Srinivasan, Matthew Tancik, Jonathan T

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. InEuropean Conference on Computer Vision, pages 405–421, 2020. 1, 2

2020
[26]

K. L. Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. CompGS: Smaller and faster gaussian splatting with vector quantization. In European Conference on Computer Vision, 2024. 2

2024
[27]

Compressed 3D Gaussian Splatting for Accelerated Novel View Synthesis, 2024

Simon Niedermayr, Josef Stumpfegger, and R¨udiger Wester- mann. Compressed 3D Gaussian Splatting for Accelerated Novel View Synthesis, 2024. arXiv:2401.02436 [cs.CV]. 1, 2

2024 arXiv
[28]

ActiveN- eRF: Learning where to see with uncertainty estimation

Xuran Pan, Zihang Lai, Shiji Song, and Gao Huang. ActiveN- eRF: Learning where to see with uncertainty estimation. In European Conference on Computer Vision, pages 230–246,
[29]

Paris-Lille-3d: A point cloud dataset for urban scene segmentation and classification

Xavier Roynard, Jean-Emmanuel Deschaud, and Fran c ¸ois Goulette. Paris-Lille-3d: A point cloud dataset for urban scene segmentation and classification. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2027–2030. IEEE, 2018. 1

2027
[30]

Toronto-3d: A large- scale mobile LiDAR dataset for semantic segmentation of urban roadways

Weikai Tan, Nannan Qin, Lingfei Ma, Ying Li, Jing Du, Guorong Cai, Ke Yang, and Jonathan Li. Toronto-3d: A large- scale mobile LiDAR dataset for semantic segmentation of urban roadways. In2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pa...

2020
[31]

NanoGS: Training-Free Gaussian Splat Simplification, 2026

Butian Xiong, Rong Liu, Tiantian Zhou, Meida Chen, Zhiwen Fan, and Andrew Feng. NanoGS: Training-Free Gaussian Splat Simplification, 2026. arXiv:2603.16103 [cs]. 1, 2, 3, 5

2026 arXiv
[32]

RSGaussian:3D Gaussian Splat- 9 ting with LiDAR for Aerial Remote Sensing Novel View Synthesis, 2024

Yiling Yao, Wenjuan Zhang, Bing Zhang, Bocheng Li, Yaning Wang, and Bowen Wang. RSGaussian:3D Gaussian Splat- 9 ting with LiDAR for Aerial Remote Sensing Novel View Synthesis, 2024. arXiv:2412.18380 [cs.CV], version 1. 1, 2

2024
[33]

FewViewGS: Gaussian splatting with few view matching and multi-stage training

Ruihong Yin, Vladimir Yugay, Yue Li, Sezer Karaoglu, and Theo Gevers. FewViewGS: Gaussian splatting with few view matching and multi-stage training. InAdvances in Neural In- formation Processing Systems. Curran Associates, Inc., 2024. NeurIPS 2024. 1, 4

2024
[34]

Drivinggaussian: Compos- ite gaussian splatting for surrounding dynamic autonomous driving scenes

Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, De- qing Sun, and Ming-Hsuan Yang. Drivinggaussian: Compos- ite gaussian splatting for surrounding dynamic autonomous driving scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages...

2024
[36]

Datasets and Evaluation Splits Table S1 lists every split used in the paper and this supple- ment

Additional Implementation and Evaluation Details 1.1. Datasets and Evaluation Splits Table S1 lists every split used in the paper and this supple- ment. All subsequent sections refer back to these names rather than restating them. 1.2. Dense Prior Construction G2ARD-GS accepts...
[37]

Rendering Quality under Aggressive Compres- sion G2ARD-GS preserves coherent surface and texture support as the primitive budget is reduced from5×to30×

Compact and Reusable Gaussian Maps 2.1. Rendering Quality under Aggressive Compres- sion G2ARD-GS preserves coherent surface and texture support as the primitive budget is reduced from5×to30×. Figure S1 shows the three operating points not promoted to the main paper. All metho...
[38]

Why the Geometry Remains Stable 3.1. Progressive Recovery and Geometry Con- straints Multi-round recovery and construction-time geometric con- straints keep the compact topology from drifting into photo- metrically convenient but geometrically unstable configura- tions. Table ...
[39]

Mip-NeRF 360 The fixed G 2ARD-GS recipe transfers to conventional bounded 3DGS scenes without scene-specific redesign

Cross-Domain Transfer 4.1. Mip-NeRF 360 The fixed G 2ARD-GS recipe transfers to conventional bounded 3DGS scenes without scene-specific redesign. Established GS-compression methods are benchmarked on Mip-NeRF 360 [ 1], where a fully converged dense 3D-GS model is always availa...
[40]

S1 es- tablishes that quality survives aggressive reduction, Fig

Extended Qualitative Results The figures collected above are ordered to follow the evi- dence chain rather than the experiment history: Fig. S1 es- tablishes that quality survives aggressive reduction, Fig. S2 that the surviving geometry supports appearance adaptation on a new...
[2023]

1, 2, 3, 5

arXiv:2311.17245 [cs.CV]; NeurIPS 2024 Spotlight. 1, 2, 3, 5

2024 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Barron, Ben Mildenhall, Dor Verbin, Pratul P

Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5460–5469. IEEE, 2022. 5, 7

work page 2022

[2] [2]

Accelerated coordinate encoding: Learning to relocalize in minutes using rgb and poses

Eric Brachmann, Tommaso Cavallari, and Victor Adrian Prisacariu. Accelerated coordinate encoding: Learning to relocalize in minutes using rgb and poses. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5044–5053, 2023. 8

work page 2023

[3] [3]

HAC: Hash-grid assisted context for 3d gaussian splatting compression

Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. HAC: Hash-grid assisted context for 3d gaussian splatting compression. InEuropean Conference on Computer Vision, 2024. 2

work page 2024

[4] [4]

LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and 200+ FPS,

Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and 200+ FPS,

work page

[5] [5]

Mini-Splatting: Represent- ing Scenes with a Constrained Number of Gaussians

Guangchi Fang and Bing Wang. Mini-Splatting: Represent- ing Scenes with a Constrained Number of Gaussians. In Computer Vision – ECCV 2024, pages 165–181. Springer Nature Switzerland, Cham, 2024. Series Title: Lecture Notes in Computer Science. 1, 2, 3

work page 2024

[6] [6]

EA- GLES: Efficient accelerated 3d gaussians with lightweight encodings

Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. EA- GLES: Efficient accelerated 3d gaussians with lightweight encodings. InEuropean Conference on Computer Vision,

work page

[7] [7]

SuGaR: Surface- aligned gaussian splatting for efficient 3d mesh reconstruc- tion and high-quality mesh rendering

Antoine Gu ´edon and Vincent Lepetit. SuGaR: Surface- aligned gaussian splatting for efficient 3d mesh reconstruc- tion and high-quality mesh rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354–5363, 2024. 2

work page 2024

[8] [8]

PUP 3D-GS: Princi- pled Uncertainty Pruning for 3D Gaussian Splatting, 2025

Alex Hanson, Allen Tu, Vasu Singla, Mayuka Jayawardhana, Matthias Zwicker, and Tom Goldstein. PUP 3D-GS: Princi- pled Uncertainty Pruning for 3D Gaussian Splatting, 2025. arXiv:2406.10219 [cs]; CVPR 2025. 1, 2, 3, 5, 7

arXiv 2025