Creating Virtual Environments with 3D Gaussian Splatting: A Comparative Study

Binzhu Xie; Pheng-Ann Heng; Qixuan Liu; Shi Qiu

REVIEW 4 major objections 4 minor 14 references

Creating Virtual Environments with 3D Gaussian Splatting: A Comparative Study

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that among three 3D Gaussian Splatting pipelines for building virtual environments, direct novel view synthesis gives the strongest user presence while all three sustain roughly 70 frames per second on a headset.

desk verdict A legitimate but statistically thin comparison of three 3DGS-based VE creation methods; the claim that novel view synthesis gives the best presence outruns the data. read the letter →

arxiv 2501.09302 v1 pith:23QO34DO submitted 2025-01-16 cs.CV cs.GRcs.HC

Shi Qiu , Binzhu Xie , Qixuan Liu , Pheng-Ann Heng This is my paper · ORCID

classification cs.CVcs.GRcs.HC

keywords 3DGaussianSplattingvirtualenvironmentsextendedrealitynovelviewsynthesismeshreconstructionpanoramicsceneuserpresenceSUSquestionnaire

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 asks whether 3D Gaussian Splatting (3DGS) can serve as a practical basis for building virtual environments for extended reality. It builds a virtual environment by three routes: direct novel-view synthesis from multiview photographs, surface-aligned Gaussian meshing followed by Poisson reconstruction and texturing, and a 360-degree panoramic Gaussian scene generated from a fixed viewpoint. In a user study with eight participants, the direct novel-view approach scored highest on the Slater-Usoh-Steed presence questionnaire, and all three approaches ran at roughly 70 frames per second on a desktop GPU driving a Meta Quest 3. The paper concludes that 3DGS is feasible for real-time XR virtual environments, while flagging the need for better object-level interaction and the fixed-viewpoint weakness of the panoramic route.

What carries the argument

The object that carries the argument is the 3D Gaussian Splatting representation: a scene stored as an explicit cloud of 3D Gaussian points that can be rendered in real time. The paper compares three pipelines built on that representation: direct view synthesis (COLMAP poses plus standard 3DGS training, rendered in Unity with an open-source shader), mesh reconstruction (SuGaR's surface-aligned Gaussian regularization, Poisson surface reconstruction, and texture cues from trained Gaussians), and panoramic scene synthesis (multiview images fused into a 360-degree Gaussian panorama using monocular depth estimation and an optimizable geometric field). The user-presence measurement is the SUS questionnaire, which turns the comparison into numbers the authors use to rank the approaches.

What would settle it

Re-run the three pipelines on a fixed set of scenes with matched training budgets and hyperparameter tuning, collect SUS ratings from a larger blinded participant pool, and measure per-frame view-consistency error along a scripted head path; if novel view synthesis no longer scores highest or no longer has the lowest consistency error, the paper's central ordering fails.

Watch

Extended reading notes

Core claim

The central discovery is that the way 3DGS is used to create a virtual environment changes how present users feel in it. The novel-view-synthesis approach, which keeps the scene as a trained cloud of 3D Gaussians and renders them directly, produced the highest SUS Count (2.8 ± 1.8) and SUS Mean (4.9 ± 1.2) and won five of the six individual presence questions; participants said visual quality stayed consistent as they moved. The mesh-reconstruction route scored lower (SUS Mean 4.5 ± 1.1) and showed reconstruction incompleteness and difficulty with distant objects, while the panoramic route scored far lower (SUS Mean 1.9 ± 0.9) and produced artifacts when users shifted their standing position. The paper treats the first route as the most suitable immediate choice for immersive free-movement XR, and the other two as receding or constrained options.

Load-bearing premise

The comparison assumes the three pipelines were implemented and tuned to a comparable level, so the large difference in presence scores reflects the approaches themselves rather than the amount of engineering effort each one received in this study.

Editorial extensions

If this is right

If the ordering holds, teams building XR virtual environments should favor direct Gaussian view synthesis over converting the scene to a textured mesh or a fixed-viewpoint panorama when users need to move freely.
3DGS-based mesh reconstruction still has to catch up with traditional TSDF methods, especially for distant objects, before it can serve as a drop-in mesh source for XR.
Panoramic Gaussian scenes remain useful for stationary virtual tours but not for room-scale or larger XR experiences, since even slight changes in standing position degrade the display.
Real-time performance around 70 FPS is reachable with 3DGS on current consumer GPU and head-mounted-display hardware, which removes a major practical obstacle to XR use.
Object-level interaction in direct Gaussian scenes requires a mechanism to identify and prioritize Gaussians belonging to specific objects, rather than treating all Gaussians uniformly.

Reading between the lines

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

Editorial inference: with eight participants and no statistical significance test, the reported SUS ordering should be read as a pilot result; a larger pre-registered study would be needed to confirm that the advantage is intrinsic to the approach.
Editorial inference: a natural testable extension is to measure view-consistency error along a scripted head path for the three pipelines; the paper's explanation predicts that direct view synthesis will also win on objective consistency metrics.
Editorial inference: the fixed-viewpoint failure of the panoramic route suggests that stitching several multiview panoramas into one 3DGS scene could extend it to larger spaces, a direction the paper flags as future work.
Editorial inference: for XR developers, the result argues for investing in Gaussian-level selection and interaction primitives instead of treating mesh conversion as the default integration path.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Here's the bottom line: this is a legitimate small empirical comparison of three 3DGS pipelines for XR virtual environments, and the qualitative material is worth a look for people working on VR scene creation. But the central quantitative claim—that novel view synthesis beats mesh reconstruction on presence—is not backed by the reported numbers.

What's actually new: the authors integrate three known methods (vanilla 3DGS, SuGaR-based mesh extraction, DreamScene360-style panoramas) into Unity for HMD use and run an eight-person SUS study. That integration is real work, and the qualitative observations about motion blur, incomplete reconstructions, and panoramic viewpoint restrictions are plausible and useful. The discussion section is honest: they note mesh reconstruction can't compete with TSDF on distant objects and that the panorama approach fails on free movement. That counts for something.

The soft spots are mostly statistical. With n=8, the SUS differences between novel view (4.9±1.2) and mesh (4.5±1.1) are small; the stress-test estimate of d≈0.35 is right, and without a paired test or effect sizes the ordering is within noise. The paper also gives no objective rendering quality metrics, tests only one scene, and describes each pipeline in a few sentences, so the implementation-effort confound the reader flagged is real. If the mesh pipeline was tuned less carefully, the gap could be an artifact. Even under equal tuning, the evidence can't distinguish the two top methods. The 70 FPS claim is less fragile, but even there no variance is reported.

Citation pattern is fine; the one self-citation is a position paper and isn't used to construct the results. The SUS findings are external ratings, so no circularity problem.

Who's this for? XR developers who want a quick sense of what 3DGS-based VE creation can do today, and researchers doing follow-up comparisons. It deserves a serious referee because the comparison question is worth answering, but it needs a revision with more participants, inferential statistics, and ideally objective quality metrics and more scenes. As it stands, treat the headline claim as preliminary.

Referee Report

4 major / 4 minor

Summary. The manuscript presents a comparative study of three 3D Gaussian Splatting (3DGS) pipelines for creating virtual environments for extended reality: a novel-view-synthesis approach, a mesh reconstruction approach (SuGaR with Poisson surface reconstruction), and a panoramic 3D scene approach. The authors render the resulting environments in Unity, display them on a Meta Quest 3 HMD, and report qualitative impressions, rendering speed (approximately 70 FPS for all approaches), and a user study with 8 participants using the Slater-Usoh-Steed (SUS) questionnaire. Based on Table 1, they claim that the novel view synthesis approach achieves the highest presence scores and conclude that 3DGS is feasible for real-time XR virtual environments, while also discussing limitations and future directions.

Significance. If the comparative claims were statistically and experimentally well supported, this would be a useful contribution to the XR and 3DGS communities, since practical evaluations of 3DGS in head-mounted displays are scarce. The paper has concrete strengths: it considers three distinct reconstruction/rendering pipelines, uses an actual HMD rather than only a desktop display, reports a standard presence questionnaire, and includes qualitative examples of artifacts. However, the evidence is currently preliminary. The headline claim that novel view synthesis provides the best user presence rests on a small single-scene user study with no inferential statistics, and the feasibility claim rests on a one-line FPS report without measurement details. The manuscript is therefore more convincing as an experience report than as a rigorous comparative evaluation.

major comments (4)

[Section 3, Table 1] The central claim that "the novel view synthesis approach achieves the highest SUS Count and Mean metrics" is not statistically supported by the reported data. Table 1 gives means and standard deviations for n=8, but no p-values, confidence intervals, effect sizes, or inferential tests are provided. The decisive comparison is Novel View Synthesis versus Mesh Reconstruction: SUS Mean 4.9±1.2 versus 4.5±1.1 (Cohen's d ≈ 0.35), Q2 4.3±1.3 versus 4.0±1.4, and Q6 4.0±2.0 versus 3.8±1.8. With eight participants, differences of this magnitude are likely to be within sampling error, and no correction is applied for the multiple SUS questions and aggregate metrics. The paper should either report appropriate paired significance tests (e.g., Wilcoxon signed-rank test with a stated correction), present confidence intervals, or explicitly soften the conclusion to a preliminary observation. As written, the observed ordering between the two leading approaches could be noise.
[Section 3] The feasibility claim "all approaches achieve approximately 70 FPS" is reported without the information needed to assess it. The authors do not state how FPS was measured (e.g., in-HMD measurement, screen capture, or Unity Profiler), how many runs were averaged, whether the 70 FPS figure is a mean, median, or peak value, or whether it varied across the approaches and during user movement. Since real-time XR feasibility is one of the two main claims, this needs a measurement protocol and per-approach results with variance. Additionally, the evaluation uses only a single scene, so the generality of both the FPS and the user-study findings is unclear.
[Section 2] The comparison suffers from a potential implementation-effort confound. Each pipeline is described in only a few sentences with no hyperparameters, training times, numbers of Gaussians, mesh resolutions, or failure cases. The user study is therefore comparing not just the three conceptual approaches but also the authors' proficiency with each pipeline. For example, if the panoramic pipeline was under-tuned relative to the novel-view pipeline, then the large SUS differences (SUS Mean 1.9±0.9 versus 4.9±1.2) would be an artifact of implementation quality rather than an intrinsic property of the approach. The authors should provide more implementation details or acknowledge that the comparison is demonstration-oriented rather than a controlled benchmark.
[Section 4] The statement that "3DGS-based mesh reconstruction could hardly compete with traditional TSDF methods" is made without presenting the results or protocol of the "additional tests" that support it. This is a comparative claim and should either be backed by data (even a small quantitative comparison) or be moved to the limitations/future-work paragraph without an unverifiable comparison.

minor comments (4)

[Introduction] The phrase "conducting a comparable study" should be "conducting a comparative study", and "system evaluations" in Section 5 should be clarified because the only quantitative system evaluation reported is the FPS statement.
[Section 3] The text has typographical spacing errors: "theSlater" and "Slater , Usoh" should be "the Slater" and "Slater, Usoh". Also, the SUS acronym is used for the Slater-Usoh-Steed questionnaire rather than the more common System Usability Scale; the authors should explicitly define the scoring of "SUS Count" and "SUS Mean" in the text, since these are not standard from the acronym alone.
[Table 1] Table 1 would benefit from a clearer layout: the column header "Virtual environments created using the approaches of" is awkward, and the table should state which questions Q1 through Q6 correspond to, since the SUS questionnaire items are not included in the paper.
[References] Reference [7] is a self-citation to the authors' position paper; this is not a problem for circularity, but the phrase "straightforward scene manipulation [7]" cites a position paper instead of a technical paper demonstrating manipulability, which would be a stronger support.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the comparative results come from external user ratings and FPS measurements, not from fitted parameters or self-referential equations.

full rationale

The paper's claims are empirical and externally grounded. The central comparative claim (novel view synthesis achieving the highest SUS Count and Mean) is supported by Table 1, which is a direct tabulation of human questionnaire responses (n=8), not a quantity derived from the paper's own definitions. The real-time feasibility claim (~70 FPS) is an external hardware measurement. The only self-citation is reference [7], used for the motivational statement that 3DGS gives 'straightforward scene manipulation'; this is not load-bearing for the measured outcomes and no derivation or prediction reduces to that citation. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported, and no known result is repackaged. The absence of significance testing for small SUS differences is a statistical-correctness concern about how strongly the ordering is established, but it is not a circularity concern.

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

The paper introduces no new parameters or entities; it relies on existing 3DGS pipelines, a pre-trained depth estimator, the SUS questionnaire, and a Unity shader. The load-bearing assumptions are that these components are valid and that the implementation effort is comparable across the three compared approaches.

assumptions (4)

domain assumption The referenced 3DGS, SuGaR, and DreamScene360 implementations are applied correctly and effectively.
The paper's conclusions rest on the reliability of prior implementations cited in Section 2, without independent verification.
domain assumption The SUS questionnaire is a valid measure of presence for comparing VE creation approaches.
SUS scores are used as the primary quantitative outcome (Section 3, Table 1) but no validity check is reported.
domain assumption The open-source Unity shader renders trained 3DGS scenes without significant artifacts that bias the comparison.
All three pipelines are rendered in Unity via different processes (Section 2); if the shader pipeline is lossy, comparisons could be skewed.
domain assumption The evaluated scene and hardware setup are representative of XR virtual environment use.
Only one scene and one hardware configuration are used; the paper treats the qualitative results as general findings (Section 3).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Creating Virtual Environments with 3D Gaussian Splatting: A Comparative Study." pith.science (2026). https://pith.science/paper/23QO34DO

@misc{pith2026250109302,
  author       = {Pith},
  title        = {Pith review of: Creating Virtual Environments with 3D Gaussian Splatting: A Comparative Study},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/23QO34DO}},
  note         = {Machine review of arXiv:2501.09302}
}

read the original abstract

3D Gaussian Splatting (3DGS) has recently emerged as an innovative and efficient 3D representation technique. While its potential for extended reality (XR) applications is frequently highlighted, its practical effectiveness remains underexplored. In this work, we examine three distinct 3DGS-based approaches for virtual environment (VE) creation, leveraging their unique strengths for efficient and visually compelling scene representation. By conducting a comparable study, we evaluate the feasibility of 3DGS in creating immersive VEs, identify its limitations in XR applications, and discuss future research and development opportunities.

Figures

Figures reproduced from arXiv: 2501.09302 by the authors.

**Figure 1.** We develop three different 3D Gaussian Splatting approaches for virtual environment creation, where the corresponding [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Qualitative comparisons of VE rendering results. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 10 canonical work pages

[1]

write newline

" write newline "" before.all 'output.state := STRINGS d dc pc.str doi.urlstr INTEGERS nth.dash doi.string.pos doilen FUNCTION add.doi doi empty "\ doi 'd := "" 'doi.urlstr := #0 'nth.dash := #1 'doi.string.pos := d text.length 'doilen := nth.dash #3 < d doi.string.pos #1 substring 'dc := #1 doi.string.pos + 'doi.string.pos := dc "/" = #1 nth.dash + 'nth....
[2]

write newline

" write newline "" before.all 'output.state := STRINGS d dc pc.str doi.urlstr INTEGERS nth.dash doi.string.pos doilen FUNCTION add.doi doi empty "\ doi 'd := "" 'doi.urlstr := #0 'nth.dash := #1 'doi.string.pos := d text.length 'doilen := nth.dash #3 < d doi.string.pos #1 substring 'dc := #1 doi.string.pos + 'doi.string.pos := dc "/" = #1 nth.dash + 'nth....
[3]

write newline

" write newline "" before.all 'output.state := STRINGS d dc pc.str doi.urlstr INTEGERS nth.dash doi.string.pos doilen FUNCTION add.doi doi empty "\ doi 'd := "" 'doi.urlstr := #0 'nth.dash := #1 'doi.string.pos := d text.length 'doilen := nth.dash #3 < d doi.string.pos #1 substring 'dc := #1 doi.string.pos + 'doi.string.pos := dc "/" = #1 nth.dash + 'nth....
[4]

write newline

" write newline "" before.all 'output.state := STRINGS d dc pc.str doi.urlstr INTEGERS nth.dash doi.string.pos doilen FUNCTION add.doi doi empty "\ doi 'd := "" 'doi.urlstr := #0 'nth.dash := #1 'doi.string.pos := d text.length 'doilen := nth.dash #3 < d doi.string.pos #1 substring 'dc := #1 doi.string.pos + 'doi.string.pos := dc "/" = #1 nth.dash + 'nth....
[5]

Gu \'e don and V

A. Gu \'e don and V. Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. CVPR , 2024

work page 2024
[6]

Guédon and V

A. Guédon and V. Lepetit. Gaussian frosting: Editable complex radiance fields with real-time rendering. In ECCV , 2024

work page 2024
[7]

Izadi et al

S. Izadi et al. Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In ACM UIST , 2011

work page 2011
[8]

Kazhdan, M

M. Kazhdan, M. Bolitho, and H. Hoppe. Poisson surface reconstruction. In Eurographics/ACM SGP , volume 7, 2006

work page 2006

Show all 14 references

[9]

Kerbl, G

B. Kerbl, G. Kopanas, T. Leimk \"u hler, and G. Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. , 42(4):139--1, 2023

2023
[10]

Mildenhall, P

B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM , 65(1):99--106, 2021

2021
[11]

S. Qiu, B. Xie, Q. Liu, and P.-A. Heng. Advancing extended reality with 3d gaussian splatting: Innovations and prospects. In IEEE AIxVR , 2025

2025
[12]

J. L. Sch\" o nberger, E. Zheng, M. Pollefeys, and J.-M. Frahm. Pixelwise view selection for unstructured multi-view stereo. In ECCV , 2016

2016
[13]

M. Usoh, E. Catena, S. Arman, and M. Slater. Using presence questionnaires in reality. Presence , 9(5):497--503, 2000

2000
[14]

Zhou et al

S. Zhou et al. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In ECCV , 2024

2024

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

write newline

" write newline "" before.all 'output.state := STRINGS d dc pc.str doi.urlstr INTEGERS nth.dash doi.string.pos doilen FUNCTION add.doi doi empty "\ doi 'd := "" 'doi.urlstr := #0 'nth.dash := #1 'doi.string.pos := d text.length 'doilen := nth.dash #3 < d doi.string.pos #1 substring 'dc := #1 doi.string.pos + 'doi.string.pos := dc "/" = #1 nth.dash + 'nth....

[2] [2]

write newline

" write newline "" before.all 'output.state := STRINGS d dc pc.str doi.urlstr INTEGERS nth.dash doi.string.pos doilen FUNCTION add.doi doi empty "\ doi 'd := "" 'doi.urlstr := #0 'nth.dash := #1 'doi.string.pos := d text.length 'doilen := nth.dash #3 < d doi.string.pos #1 substring 'dc := #1 doi.string.pos + 'doi.string.pos := dc "/" = #1 nth.dash + 'nth....

[3] [3]

write newline

" write newline "" before.all 'output.state := STRINGS d dc pc.str doi.urlstr INTEGERS nth.dash doi.string.pos doilen FUNCTION add.doi doi empty "\ doi 'd := "" 'doi.urlstr := #0 'nth.dash := #1 'doi.string.pos := d text.length 'doilen := nth.dash #3 < d doi.string.pos #1 substring 'dc := #1 doi.string.pos + 'doi.string.pos := dc "/" = #1 nth.dash + 'nth....

[4] [4]

write newline

" write newline "" before.all 'output.state := STRINGS d dc pc.str doi.urlstr INTEGERS nth.dash doi.string.pos doilen FUNCTION add.doi doi empty "\ doi 'd := "" 'doi.urlstr := #0 'nth.dash := #1 'doi.string.pos := d text.length 'doilen := nth.dash #3 < d doi.string.pos #1 substring 'dc := #1 doi.string.pos + 'doi.string.pos := dc "/" = #1 nth.dash + 'nth....

[5] [5]

Gu \'e don and V

A. Gu \'e don and V. Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. CVPR , 2024

work page 2024

[6] [6]

Guédon and V

A. Guédon and V. Lepetit. Gaussian frosting: Editable complex radiance fields with real-time rendering. In ECCV , 2024

work page 2024

[7] [7]

Izadi et al

S. Izadi et al. Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In ACM UIST , 2011

work page 2011

[8] [8]

Kazhdan, M

M. Kazhdan, M. Bolitho, and H. Hoppe. Poisson surface reconstruction. In Eurographics/ACM SGP , volume 7, 2006

work page 2006