Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Reflections Unlock: Geometry-Aware Reflection Disentanglement in 3D Gaussian Splatting for Photorealistic Scenes Rendering

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Ref-Unlock splits 3D Gaussian scenes into transmitted and reflected branches, beating GS-based reflection methods on real reflective scenes.

desk verdict Useful mask-free reflection disentanglement for 3DGS, but the headline LPIPS edge over NeRFReN is a byproduct of tuning on the evaluation set. read the letter →

arxiv 2507.06103 v1 pith:JD2SDGSH submitted 2025-07-08 cs.CV

Jiayi Song , Zihan Ye , Qingyuan Zhou , Weidong Yang , Ben Fei , Jingyi Xu , Ying He , Wanli Ouyang This is my paper · ORCID
classification cs.CV
keywords novelviewsynthesis3DGaussianSplattingreflectiondisentanglementsphericalharmonicsremovalpseudo-depthsupervisionreal-timerenderingediting
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

The paper tries to show that 3D Gaussian Splatting can render reflective scenes faithfully if the scene is explicitly decomposed into a transmitted component and a reflected component, rather than forcing one set of Gaussians to absorb reflections as if they were geometry. It argues that this disentanglement, supervised by pseudo-reflection-free images from a single-image reflection-removal model and regularized by pseudo-depth and a bilateral smoothness constraint, lets a real-time Gaussian renderer approach the quality of slower NeRF-based reflection models. If true, the result matters because it brings editable, photorealistic rendering of mirrors, glass, and glossy surfaces into the real-time domain without needing manual reflection masks.

What carries the argument

The central object is the dual-branch Gaussian representation Ref-Unlock = {µ, Σ, D, cref, αref, ctrans, αtrans, βref}, where each Gaussian carries separate spherical-harmonic color and opacity for reflection and transmission, and a reflection confidence βref in [0,1] that accumulates into a per-pixel reflection map Mref through alpha-blending with an energy-conservation constraint Mref + Mtrans = 1. The final color is a reflection-map-weighted fusion of the two branches. The load-bearing mechanism is the reflection-removal module that generates pseudo-clean images Iclean and supervises the transmitted branch (Eq. 13); the ablation shows that without it the reflection/transmission separation collapses, so the entire disentanglement is carried by external single-image reflection removal and pseudo-depth supervision.

What would settle it

Take a real reflective scene where DSRNet visibly misclassifies a mirror reflection as transmission (as Sec. 6 admits happens), render Ref-Unlock's transmission branch alone, and check whether the mirror-reflected object still appears there; if it does, the claimed disentanglement has not occurred and the framework is simply copying the teacher's error.

Watch

Extended reading notes

Core claim

The paper's central claim is that a dual-branch representation in 3DGS, where each Gaussian carries separate color and opacity for reflected and transmitted light plus a learnable reflection confidence, can physically separate reflections from scene geometry and render them with fidelity comparable to NeRF-based reflection methods. The decomposition is driven by a photometric loss that matches the transmitted branch to pseudo reflection-free images generated by DSRNet, by depth supervision from Depth Anything v2, and by bilateral smoothness on depth and reflection maps. On the RFFR dataset, the method reaches 34.365 PSNR, 0.9488 SSIM, and 0.2111 LPIPS, beating every GS-based baseline and beating NeRFReN on LPIPS; on Shiny Blender it posts the best LPIPS among all listed methods (0.0767) and the strongest PSNR among GS-based methods (30.057). The same explicit reflection representation enables user-controlled reflection scaling in text-prompted regions without manual masks.

Load-bearing premise

The whole decomposition collapses if the external reflection-removal model (DSRNet) gives wrong clean images, because the transmitted branch is trained to match those images and the ablation shows separation fails without that supervision.

Editorial extensions

If this is right

  • If the central claim holds, real-time 3DGS renderers can handle mirrors, glass, and semi-reflective surfaces without manual masks, closing a known gap where reflections get baked into geometry as blurry artifacts.
  • The explicit reflection map makes reflection intensity editable per region: attenuating or amplifying reflections by a user coefficient changes only the reflected content, leaving the transmission branch untouched.
  • The method's near-parity with NeRF-based reflection models on LPIPS suggests that a fast explicit representation can substitute for slow volumetric optimization in reflective novel-view synthesis.
  • Because the decomposition is unsupervised with respect to ground-truth reflection layers, the framework can be applied to new real-world scenes by running the same external pseudo-label generators, without per-scene parameter tuning.
  • The high-degree (5th-order) spherical harmonics on both branches provide a direct way to capture high-frequency specular details, with the paper reporting consistent PSNR and LPIPS gains as SH degree increases on RFFR.

Reading between the lines

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

  • The paper's gains over vanilla 3DGS are largely inherited from external supervision: if DSRNet mislabels reflected content as transmission (which Sec. 6 concedes it sometimes does), Ref-Unlock is trained to reproduce that error, so the decomposition quality is bounded by the pseudo-label quality.
  • A testable extension is to replace the DSRNet pseudo-clean images with a zero-shot reflection-removal model; if the paper's RFFR LPIPS advantage is preserved, the framework is robust to the choice of external teacher, whereas a drop would confirm the teacher is the bottleneck.
  • The reflection map and separated branches could be reused as a training-signal generator for downstream tasks like reflection removal or relighting, since they provide per-scene pseudo-labels that are view-consistent by construction.
  • The depth-smoothness constraint's reliance on rendered transmission colors means that errors in the transmission branch can feed back into depth regularization; a failure cascade is plausible in scenes where the transmission branch is itself misled.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. Ref-Unlock is a 3D Gaussian Splatting framework for novel view synthesis of reflective scenes. It decomposes the scene into transmitted and reflected branches, each with its own color and opacity, and introduces a learnable per-Gaussian reflection confidence that is accumulated into a pixel-wise reflection map. High-order spherical harmonics (degree 5) are used for both branches to capture high-frequency reflections. The transmitted branch is supervised by pseudo reflection-free images produced by DSRNet, and the geometry is regularized by pseudo-depth from Depth Anything v2 plus a bilateral smoothness constraint. The paper reports experiments on the RFFR and Shiny Blender datasets, claiming that Ref-Unlock significantly outperforms GS-based baselines and is competitive with NeRF-based methods, while also enabling reflection editing via a text-prompted segmentation model.

Significance. If the reported numbers are reliable, Ref-Unlock is a practically valuable contribution: it brings explicit reflection disentanglement to the real-time 3DGS paradigm without manual reflection masks, and it ships with a public code release, which supports reproducibility. The use of public benchmarks and externally pretrained pseudo-label generators (DSRNet, Depth Anything v2) means the headline comparisons are not circular. However, the central comparative claim is currently weakened by the evaluation protocol: hyperparameters are selected on the same RFFR scenes used for the final numbers, and the specific LPIPS superiority over NeRFReN appears to depend on that selection. The disentanglement also relies heavily on an external reflection-removal model whose failure modes are acknowledged in the paper. These issues are fixable with a proper validation protocol and more careful claims, so the work merits revision rather than rejection.

major comments (3)
  1. [Section 5, Tables 4-6] All ablation and hyperparameter experiments are performed on the entire RFFR dataset, and the final numbers in Table 1 are reported on that same dataset. The choices S_bItrans=10, lambda_depth=30, and SH degree=5 are made by optimizing metrics on the evaluation set. This is load-bearing for the LPIPS claim: Table 4 shows that S_bItrans=1 yields higher PSNR (34.557) but LPIPS 0.2250, which is worse than NeRFReN's 0.2152; only by selecting the configuration with the best LPIPS does Ref-Unlock surpass NeRFReN. Since the baselines use default or officially released hyperparameters, the comparison is biased in Ref-Unlock's favor. Please provide a held-out validation split or scene-level cross-validation and re-report the final numbers, or explicitly disclose that hyperparameters were tuned on the test set and present the resulting uncertainty in the claims.
  2. [Eq. (5) and Eq. (6)] The reflection map in Eq. (5) accumulates beta_i^ref multiplied by alpha_i^trans, where alpha_trans is defined as the transmission opacity. Since the reflection branch has its own opacity alpha_ref, using the transmission opacity in the reflection map is either a typo (should be alpha_i^ref) or a fundamental mis-specification of the compositing model. In addition, Eq. (6) asserts M_ref + M_trans = 1, but M_trans is never defined and it is not shown that the proposed accumulation satisfies this identity. Please correct the equation, define M_trans explicitly, and provide a derivation of the conservation relation.
  3. [Section 6 and Table 3] The paper concedes in Section 6 that DSRNet can misclassify reflected content such as mirror reflections as transmission, and Table 3 indicates that without the RRM loss the reflection/transmission separation collapses (Model A vs. the full model, if the checkmarks are correctly interpreted). This means the disentanglement is effectively carried by an external single-image reflection-removal model whose errors are propagated into the trained Gaussians and into the reflection-editing results of Section 4.6. The paper should quantify how DSRNet's failures affect the separation and editing on the two benchmarks (e.g., error maps or per-scene analysis), or explicitly scope the claims to scenes where DSRNet succeeds. Without this, the editing demonstration of Section 4.6 may scale mislabeled content rather than true reflections.
minor comments (6)
  1. [Table 3] The checkmark alignment in Table 3 is inconsistent with the text: the text says Model A removes the Reflection Removal Module, but the table appears to show Model A with the RRM checkmark present and only the reflection-map loss missing. Please make the table columns unambiguous so each ablation variant is clear.
  2. [Table 2 title] The title of Table 2 contains a typo: 'ShniyBlender' should be 'Shiny Blender'.
  3. [Figure 6 caption] The caption of Figure 6 refers to 'ReFNeRF'; this should be 'Ref-NeRF' for consistency with the text and references.
  4. [Eq. (8)] In Eq. (8), the phrase 'b target function' contains a typo; it should read 'the target function'.
  5. [Section 4.4 and abstract] The abstract claims the method 'significantly outperforms classical GS-based reflection methods,' but on Shiny Blender the margins over 3DGS and R3DG are small in several scenes and metrics; 'significantly' should be supported by statistical testing or softened.
  6. [Section 4.3] The paper emphasizes real-time rendering but does not report training or rendering time for Ref-Unlock or the baselines; a timing table would support the efficiency claims made in the introduction and Section 4.4.2.

Circularity Check

1 steps flagged · score 6.0 of 10

RFFR hyperparameters are tuned on the full benchmark and the same benchmark is then used to support the NeRF-competitive claim, so the reported LPIPS advantage is a selected optimum, not a prediction.

  1. fitted input called prediction [Section 5.2 (Table 4), Section 4.4.1, Section 4.3]
    "We conduct three sets of experiments with S_bItrans = 1, 10, and 20 to decide the optimal parameter for the transmission strength. ... when the transmission strength is set to 10, the model achieves the best SSIM and LPIPS scores, and we therefore select it as the optimal parameter. ... Ref-Unlock outperforms NeRFReN—the method specifically proposed for this dataset—in terms of LPIPS, indicating better perceptual quality."

    The final RFFR comparison (Table 1) uses S_bItrans=10, chosen in Sec. 5.2 because it gives the best SSIM and LPIPS on the full RFFR dataset. Table 4 shows that S_bItrans=1 yields higher PSNR (34.557) but LPIPS 0.2250, worse than NeRFReN's 0.2152; S_bItrans=10 yields 0.2111, the exact margin used to claim LPIPS superiority over NeRFReN. Thus the headline 'outperforms NeRFReN in LPIPS' claim is not an unbiased held-out prediction but the selected optimum on the same evaluation set. Likewise, lambda_depth=30 and SH degree=5 are chosen by best PSNR/LPIPS on the same six scenes (Tables 5 and 6), and baselines run with default hyperparameters (Sec. 4.3).

full rationale

The core derivation is not circular: the dual-branch SH decomposition, reflection map, and loss functions are defined from the model's own parameters; DSRNet and Depth Anything v2 are external pretrained models used for supervision, not the authors' own prior results; and there is no self-citation chain or imported uniqueness theorem. Shiny Blender results use the same hyperparameters without tuning on that benchmark, so they remain externally anchored. However, the RFFR evaluation is compromised: Sec. 5 performs all hyperparameter ablations on the entire RFFR dataset and then the same dataset produces the headline Table 1. The LPIPS value 0.2111 used to claim superiority over NeRFReN is the selected optimum of S_bItrans on the test set, with the neighboring setting giving 0.2250. This makes one of the paper's central claims (NeRF-competitive perceptual quality) a fitted selection rather than a genuine prediction. The large PSNR margin over GS-based baselines is less likely to be an artifact of this selection, so the circularity is partial rather than total.

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

The method is a composition of existing representations plus a few tuned constants; it introduces no new physics. Free parameters are the five loss weights and the SH degree, fixed by ablation on the evaluation datasets, plus a learned reflection confidence whose semantics are unvalidated. Axioms include the per-pixel energy-conservation compositing (Eq. 6) and strong trust in two external priors (DSRNet, Depth Anything v2) whose failure modes the paper documents itself. The bilateral weight on rendered transmission (Eq. 16) is self-referential. No invented physical entities; the closest is the reflection confidence, which has no independent evidence.

free parameters (5)
  • SH degree n = 5
    Chosen by ablation on the full RFFR dataset (Table 6); gains over degree 3 are small (34.250 to 34.365 PSNR).
  • lambda_depth (pseudo-depth weight) = 30
    Largest loss weight; selected by ablation on the evaluation dataset (Table 5), so the final RFFR numbers partly reflect tuning on the test set.
  • S_bItrans (transmission branch strength) = 10
    Selected by ablation (Table 4); S=1 gives higher PSNR but visually poorer disentanglement, so the choice blends quantitative and qualitative criteria.
  • Loss weights lambda_bI, lambda_init, lambda_bi, lambda_ref = 0.8, 0.01, 0.001, 0.001
    Hand-set in Sec. 4.3 without sensitivity analysis; the paper claims the 0.001 weights are essential despite being tiny.
  • Reflection confidence beta_ref (per-Gaussian) = learned
    The core learned quantity defining the reflection map (Eq. 5); its semantics as a reflection probability are not validated against any independent measurement.
assumptions (5)
  • domain assumption Per-pixel energy conservation cM_ref + cM_trans = 1 with the reflection map accumulated using transmission opacity alpha_trans (Eqs. 5 and 6).
    Energy conservation is a physical principle, but the specific compositing rule, weighting reflection contribution by transmission opacity, is a modeling choice not derived from the cited optics texts (Hecht, Pharr et al.).
  • domain assumption DSRNet pseudo-reflection-free images are valid targets for the transmitted branch (Eq. 13).
    Load-bearing external prior; the ablation (Table 3, Model A) shows removing it collapses disentanglement, while Sec. 6 admits DSRNet misclassifies mirror reflections as transmission and blurs details.
  • domain assumption Depth Anything v2 pseudo-depth is a valid geometric prior near reflective surfaces (Eq. 14).
    Section 3.4 states monocular depth estimators may assign depth values that contradict physical reality, so the prior is trusted with the largest weight (lambda_depth=30) while being acknowledged as unreliable in the regions the method targets.
  • ad hoc to paper Bilateral smoothness weight uses rendered transmitted images rather than ground-truth color (Eq. 16).
    Self-referential regularizer: the weight function is computed from the very transmission estimate the regularizer is meant to stabilize. A deliberate deviation from NeRFReN, but presented without a stability analysis.
  • standard math Exponential SH approximation error bound with unspecified constants A and b (Eq. 8).
    The bound is a standard approximation-theory result (Dai 2013), but the constants are not instantiated, and the derived 30-degree angular resolution (180/(5+1)) is a heuristic not tied to Eq. 8.
invented entities (1)
  • Reflection confidence beta_ref and pixel reflection map cM_ref
    purpose: Per-Gaussian and per-pixel continuous weighting that separates reflected from transmitted radiance and enables later reflection scaling edits (Sec. 4.6).
    No external measurement validates the reflection map; correctness is inferred only from final rendering quality, which jointly tests all components. The editing demo (Fig. 8) is qualitative.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reflections Unlock: Geometry-Aware Reflection Disentanglement in 3D Gaussian Splatting for Photorealistic Scenes Rendering." pith.science (2026). https://pith.science/paper/JD2SDGSH

@misc{pith2026250706103,
  author       = {Pith},
  title        = {Pith review of: Reflections Unlock: Geometry-Aware Reflection Disentanglement in 3D Gaussian Splatting for Photorealistic Scenes Rendering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JD2SDGSH}},
  note         = {Machine review of arXiv:2507.06103}
}
read the original abstract

Accurately rendering scenes with reflective surfaces remains a significant challenge in novel view synthesis, as existing methods like Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) often misinterpret reflections as physical geometry, resulting in degraded reconstructions. Previous methods rely on incomplete and non-generalizable geometric constraints, leading to misalignment between the positions of Gaussian splats and the actual scene geometry. When dealing with real-world scenes containing complex geometry, the accumulation of Gaussians further exacerbates surface artifacts and results in blurred reconstructions. To address these limitations, in this work, we propose Ref-Unlock, a novel geometry-aware reflection modeling framework based on 3D Gaussian Splatting, which explicitly disentangles transmitted and reflected components to better capture complex reflections and enhance geometric consistency in real-world scenes. Our approach employs a dual-branch representation with high-order spherical harmonics to capture high-frequency reflective details, alongside a reflection removal module providing pseudo reflection-free supervision to guide clean decomposition. Additionally, we incorporate pseudo-depth maps and a geometry-aware bilateral smoothness constraint to enhance 3D geometric consistency and stability in decomposition. Extensive experiments demonstrate that Ref-Unlock significantly outperforms classical GS-based reflection methods and achieves competitive results with NeRF-based models, while enabling flexible vision foundation models (VFMs) driven reflection editing. Our method thus offers an efficient and generalizable solution for realistic rendering of reflective scenes. Our code is available at https://ref-unlock.github.io/.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Confidence matters: Leveraging Multi-view Geometric Priors for GS-based Reconstruction

    cs.CV 2026-08 conditional novelty 6.0 of 10

    Confidence-weighted multi-view geometric priors from VGGT improve 3D Gaussian splatting reconstruction on specular objects, cutting normal MAE on Shiny Blender from 3.23 to 1.23 degrees.

Reference graph

Works this paper leans on

55 extracted references · 52 canonical work pages · cited by 1 Pith paper

  1. [1]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp 5855--5864

    Barron JT, Mildenhall B, Tancik M, et al (2021) Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp 5855--5864

    work page 2021

  2. [2]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp 4217--4229

    Chan ER, Nagano K, Chan MA, et al (2023) Generative novel view synthesis with 3d-aware diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp 4217--4229

    work page 2023

  3. [3]

    arXiv preprint arXiv:230913039

    Chen X, Liu J, Zhao H, et al (2023) Nerrf: 3d reconstruction and view synthesis for transparent and specular objects with neural refractive-reflective fields. arXiv preprint arXiv:230913039

    work page 2023

  4. [4]

    Springer

    Dai F (2013) Approximation theory and harmonic analysis on spheres and balls. Springer

    work page 2013

  5. [5]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 12882--12891

    Deng K, Liu A, Zhu JY, et al (2022) Depth-supervised nerf: Fewer views and faster training for free. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 12882--12891

    work page 2022

  6. [6]

    IEEE Transactions on Visualization and Computer Graphics

    Fei B, Xu J, Zhang R, et al (2024) 3d gaussian splatting as new era: A survey. IEEE Transactions on Visualization and Computer Graphics

    work page 2024

  7. [7]

    arXiv preprint arXiv:231116043

    Gao J, Gu C, Lin Y, et al (2023) Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv preprint arXiv:231116043

    work page 2023

  8. [8]

    arXiv preprint arXiv:221000379

    Gao K, Gao Y, He H, et al (2022) Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:221000379

    work page 2022

Show all 55 references
  1. [9]

    Green R (2003) Spherical harmonic lighting: The gritty details. Tech. rep., Sony Computer Entertainment America

    2003

  2. [10]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 5354--5363

    Gu \'e don A, Lepetit V (2024) Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 5354--5363

    2024

  3. [11]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 18409--18418

    Guo YC, Kang D, Bao L, et al (2022) Nerfren: Neural radiance fields with reflections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 18409--18418

    2022

  4. [12]

    Addison Wesley

    Hecht E (2002) Optics. Addison Wesley

    2002

  5. [13]

    2310.10650

    Holland LV, Bliersbach R, Müller JU, et al (2023) Tram-nerf: Tracing mirror and near-perfect specular reflections through neural radiance fields. 2310.10650

    2023 arXiv

  6. [14]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp 13138--13147

    Hu Q, Guo X (2023) Single image reflection separation via component synergy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp 13138--13147

    2023

  7. [15]

    In: ACM SIGGRAPH 2024 conference papers, pp 1--11

    Huang B, Yu Z, Chen A, et al (2024) 2d gaussian splatting for geometrically accurate radiance fields. In: ACM SIGGRAPH 2024 conference papers, pp 1--11

    2024

  8. [16]

    International Journal of Computer Vision 132(4):1310--1335

    Jain N, Kumar S, Van Gool L (2024) Learning robust multi-scale representation for neural radiance fields from unposed images. International Journal of Computer Vision 132(4):1310--1335

    2024

  9. [17]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 5322--5332

    Jiang Y, Tu J, Liu Y, et al (2024) Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 5322--5332

    2024

  10. [18]

    ACM Transactions on Graphics 42(4):1--14

    Kerbl B, Kopanas G, Leimk \"u hler T, et al (2023) 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4):1--14

    2023

  11. [19]

    In: Computer Graphics Forum, Wiley Online Library, pp 29--43

    Kopanas G, Philip J, Leimk \"u hler T, et al (2021) Point-based neural rendering with per-view optimization. In: Computer Graphics Forum, Wiley Online Library, pp 29--43

    2021

  12. [20]

    Artificial Intelligence Review 58(7):1--53

    Li F, Ma J, Liang HN, et al (2025) A comprehensive survey of specularity detection: state-of-the-art techniques and breakthroughs. Artificial Intelligence Review 58(7):1--53

    2025

  13. [21]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 21644--21653

    Liang Z, Zhang Q, Feng Y, et al (2024) Gs-ir: 3d gaussian splatting for inverse rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 21644--21653

    2024

  14. [22]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp 6036--6045

    Lin CY, Fu Q, Merth T, et al (2024) Fastsr-nerf: improving nerf efficiency on consumer devices with a simple super-resolution pipeline. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp 6036--6045

    2024

  15. [23]

    arXiv preprint arXiv:240511921

    Liu J, Tang X, Cheng F, et al (2024) Mirrorgaussian: Reflecting 3d gaussians for reconstructing mirror reflections. arXiv preprint arXiv:240511921

    2024

  16. [24]

    International Journal of Computer Vision pp 1--21

    Liu X, Li S, Gao Y (2025) Image matting and 3d reconstruction in one loop. International Journal of Computer Vision pp 1--21

    2025

  17. [25]

    arXiv preprint arXiv:231213102

    Ma L, Agrawal V, Turki H, et al (2023) Specnerf: Gaussian directional encoding for specular reflections. arXiv preprint arXiv:231213102

    2023

  18. [26]

    arXiv preprint arXiv:231213729

    Malarz D, Smolak W, Tabor J, et al (2023) Gaussian splatting with nerf-based color and opacity. arXiv preprint arXiv:231213729

    2023

  19. [27]

    arXiv preprint arXiv:240401168

    Meng J, Li H, Wu Y, et al (2024) Mirror-3dgs: Incorporating mirror reflections into 3d gaussian splatting. arXiv preprint arXiv:240401168

    2024

  20. [28]

    Communications of the ACM 65(1):99--106

    Mildenhall B, Srinivasan PP, Tancik M, et al (2021) Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1):99--106

    2021

  21. [29]

    arXiv preprint arXiv:200613846

    Nilsson J, Akenine-M \"o ller T (2020) Understanding ssim. arXiv preprint arXiv:200613846

    2020

  22. [30]

    Morgan Kaufmann

    Pharr M, Jakob W, Humphreys G (2016) Physically Based Rendering: From Theory to Implementation. Morgan Kaufmann

    2016

  23. [31]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 10318--10327

    Pumarola A, Corona E, Pons-Moll G, et al (2021) D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 10318--10327

    2021

  24. [32]

    In: Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pp 497--500

    Ramamoorthi R, Hanrahan P (2001) An efficient representation for irradiance environment maps. In: Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pp 497--500

    2001

  25. [33]

    In: Proceedings of the IEEE/CVF international conference on computer vision, pp 14335--14345

    Reiser C, Peng S, Liao Y, et al (2021) Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In: Proceedings of the IEEE/CVF international conference on computer vision, pp 14335--14345

    2021

  26. [34]

    ACM Transactions on Graphics (ToG) 41(4):1--14

    R \"u ckert D, Franke L, Stamminger M (2022) Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (ToG) 41(4):1--14

    2022

  27. [35]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 4104--4113

    Schonberger JL, Frahm JM (2016) Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 4104--4113

    2016

  28. [36]

    In: ACM Transactions on Graphics (TOG), ACM, pp 527--536

    Sloan PP, Kautz J, Snyder J (2002) Precomputed radiance transfer for real-time rendering in dynamic, low-frequency lighting environments. In: ACM Transactions on Graphics (TOG), ACM, pp 527--536

    2002

  29. [37]

    In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, pp 5481--5490

    Verbin D, Hedman P, Mildenhall B, et al (2022) Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, pp 5481--5490

    2022

  30. [38]

    In: SIGGRAPH Asia 2024 Conference Papers, pp 1--10

    Verbin D, Srinivasan PP, Hedman P, et al (2024) Nerf-casting: Improved view-dependent appearance with consistent reflections. In: SIGGRAPH Asia 2024 Conference Papers, pp 1--10

    2024

  31. [39]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp 9065--9076

    Wang G, Chen Z, Loy CC, et al (2023) Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp 9065--9076

    2023

  32. [40]

    International Journal of Computer Vision pp 1--24

    Wang J, Lyu Z, Fei B, et al (2024) Slide: A unified mesh and texture generation framework with enhanced geometric control and multi-view consistency. International Journal of Computer Vision pp 1--24

    2024

  33. [41]

    arXiv preprint arXiv:240519671

    Xiang H, Li X, Lai X, et al (2024) Gaussianroom: Improving 3d gaussian splatting with sdf guidance and monocular cues for indoor scene reconstruction. arXiv preprint arXiv:240519671

    2024

  34. [42]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp 18430--18439

    Xu Y, Peng S, Yang C, et al (2022) 3d-aware image synthesis via learning structural and textural representations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp 18430--18439

    2022

  35. [43]

    Advances in Neural Information Processing Systems 37:21875--21911

    Yang L, Kang B, Huang Z, et al (2024) Depth anything v2. Advances in Neural Information Processing Systems 37:21875--21911

    2024

  36. [44]

    arXiv preprint arXiv:241219282

    Yao Y, Zeng Z, Gu C, et al (2024) Reflective gaussian splatting. arXiv preprint arXiv:241219282

    2024

  37. [45]

    arXiv preprint arXiv:241024204

    Ye K, Gao C, Li G, et al (2024 a ) Geosplatting: Towards geometry guided gaussian splatting for physically-based inverse rendering. arXiv preprint arXiv:241024204

    2024

  38. [46]

    In: ACM SIGGRAPH 2024 Conference Papers, pp 1--10

    Ye K, Hou Q, Zhou K (2024 b ) 3d gaussian splatting with deferred reflection. In: ACM SIGGRAPH 2024 Conference Papers, pp 1--10

    2024

  39. [47]

    arXiv preprint arXiv:240316964

    Yu M, Lu T, Xu L, et al (2024) Gsdf: 3dgs meets sdf for improved rendering and reconstruction. arXiv preprint arXiv:240316964

    2024

  40. [48]

    In: Proceedings of the 31st ACM International Conference on Multimedia, pp 4606--4615

    Zeng J, Bao C, Chen R, et al (2023) Mirror-nerf: Learning neural radiance fields for mirrors with whitted-style ray tracing. In: Proceedings of the 31st ACM International Conference on Multimedia, pp 4606--4615

    2023

  41. [49]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 5449--5458

    Zhang X, Bi S, Sunkavalli K, et al (2022) Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 5449--5458

    2022

  42. [50]

    arXiv preprint arXiv:240620076

    Zhang Y, Cheng T, Zhu L, et al (2024) Evf-sam: Early vision-language fusion for text-prompted segment anything model. arXiv preprint arXiv:240620076

    2024

  43. [51]

    arXiv preprint arXiv:250305182

    Zhou Q, Gong Y, Yang W, et al (2025) Mgsr: 2d/3d mutual-boosted gaussian splatting for high-fidelity surface reconstruction under various light conditions. arXiv preprint arXiv:250305182

    2025

  44. [52]

    Artificial Intelligence Review 58(1):30

    Zhu H, Zhang Z, Zhao J, et al (2024) Scene reconstruction techniques for autonomous driving: a review of 3d gaussian splatting. Artificial Intelligence Review 58(1):30

    2024

  45. [53]

    International Journal of Computer Vision 133(2):611--627

    Zuo X, Samangouei P, Zhou Y, et al (2025) Fmgs: Foundation model embedded 3d gaussian splatting for holistic 3d scene understanding. International Journal of Computer Vision 133(2):611--627

    2025

  46. [54]

    , " * write output.state after.block = add.period write newline

    ENTRY address archive author booktitle chapter doi edition editor eid eprint howpublished institution journal key keywords month note number organization pages publisher school series title type url volume year archivePrefix primaryClass adsurl adsnote version label extra.labe...

  47. [55]

    write newline

    " write newline "" before.all 'output.state := FUNCTION add.period duplicate empty 'skip "." * add.blank if FUNCTION if.digit duplicate "0" = swap duplicate "1" = swap duplicate "2" = swap duplicate "3" = swap duplicate "4" = swap duplicate "5" = swap duplicate "6" = swap dupl...

Pith tools

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