Pith. sign in

REVIEW 4 major objections 6 minor 39 references

RePaintGS: Reference-Guided Gaussian Splatting for Realistic and View-Consistent 3D Scene Inpainting

T0 review · 4 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 RePaintGS claims that one user-picked inpainted view, warped across viewpoints and weighted by perceptual similarity, makes 3D scene inpainting realistic and view-consistent.

desk verdict RePaintGS is a plausible incremental method for view-consistent 3D inpainting, but the SPIn-NeRF evaluation leaks the reference image from the test set and the depth-based confidence is self-referential. read the letter →

arxiv 2507.08434 v1 pith:D76NPZCV submitted 2025-07-11 cs.CV

Ji Hyun Seo , Byounhyun Yoo , Gerard Jounghyun Kim This is my paper
classification cs.CV
keywords 3DsceneinpaintingGaussiansplattingobjectremovalreference-guidedreconstructionmulti-viewconsistencydepth-basedwarpingLPIPSimage
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 claims that 3D scene inpainting — filling the background revealed when an object is removed from a multi-view scene — can be made realistic and view-consistent by anchoring the whole reconstruction to a single user-selected inpainted image. It argues that previous approaches fail because per-view image inpainting produces one of many plausible completions, so blending them with pixel-level or perceptual losses loses fine detail or drifts semantically. RePaintGS instead warps the reference inpainting into every other view using estimated depth, scores each view's perceptual similarity to the warped reference as a confidence, and uses those scores to weight the view's contribution to geometry and to guide appearance toward the reference. If the claim holds, object removal in radiance-field scenes can preserve the fine details and appearance of one chosen completion while remaining coherent from arbitrary viewpoints.

What carries the argument

The load-bearing mechanism is the confidence-weighted, geometry-consistent warping loop. Confidence is computed by depth-warping the reference inpainting into a target view, bilateral-filtering both patches to suppress texture, and evaluating $conf_j = 1-\mathrm{LPIPS}(I^{\mathrm{BF}}_{\mathrm{ref}\to j}|_P, I^{\mathrm{BF}}_j|_P)$; the scalar controls the per-view training weight $w_{\mathrm{inp}}=\sigma(\alpha(conf-\beta))$ and therefore how much that view's inpainted depth and normals shape the geometry. Appearance transfer is gated by the set $P' = \{p \in P \cap M : |\hat{D}_{\mathrm{ref}}(p) - \hat{D}_{\mathrm{tar}\to\mathrm{ref}}(p)| \le \tau\}$, which excludes pixels where the optimized geometry disagrees with the reference geometry, so pseudo-ground truth is applied only where the reconstruction has caught up with the reference.

What would settle it

Take a scene with a large occlusion that is not visible from the chosen reference view, or add realistic noise to the aligned depth maps after alignment, then run the full pipeline; if the inpainted region still comes out sharp and consistent, the claimed reliance on reference visibility and depth-guided warping is wrong, while visible ghosts, blur, or floating Gaussians would confirm the stated mechanism.

Watch

Extended reading notes

Core claim

RePaintGS establishes that a reference-view inpainting can be turned into a globally consistent 3D reconstruction through two linked operations: confidence-weighted multi-view fusion and progressive geometry-consistent warping. After an initial 3D Gaussian Splatting pass trained on masked images exposes the background behind the removed object, each rendered view is inpainted and given a monocular depth map, and the user picks one inpainted view as the reference. The reference is warped to each other view with the projective mapping $p_{j\to i}\approx K R_i R_j^{-1} D_j(p_j) K^{-1}p_j$, then compared against that view's own inpainting under LPIPS after bilateral filtering, yielding $conf_j = 1 - \mathrm{LPIPS}(\cdot)$. These confidence values weight the depth, normal, and appearance losses so that unreliable inpaintings contribute less. During optimization the reference is warped again to each target view, but only into pixels where the rendered depth and the reprojected target depth agree within a threshold $\tau$ — the geometry-consistent set $P'$ — and the warped reference, Poisson-blended into the target, serves as pseudo-ground truth. The result is a single 3DGS model whose inpainted region matches the reference appearance across viewpoints.

Load-bearing premise

The hole left by the removed object must be visible in the chosen reference view, and the monocular depth estimates must be accurate enough that warping the reference into the other views lands in the right places.

Editorial extensions

If this is right

  • Object-removal editing for radiance fields no longer requires inpainted views to agree with one another; a single reference view can arbitrate the final appearance.
  • Fine textures and structural details from the chosen reference should survive across wide viewpoint changes, where LPIPS-only blending blurs them.
  • Confidence is recomputed during optimization as geometry improves, so the pipeline can start from rough multi-view inpaintings and progressively harden geometry and appearance.
  • The preprocessing modules — segmentation, image inpainting, and monocular depth estimation — are interchangeable, so replacing any of them with a stronger model should directly improve the final reconstruction.

Reading between the lines

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

  • The stated visibility assumption implies the method will struggle exactly where inpainting is hardest: holes that are occluded in the reference view itself, or visible only under very different angles; a stress test would pick a reference view with minimal overlap of the target hole.
  • Because low-confidence views are downweighted, the final scene's appearance is essentially inherited from the reference; if the user picks an outlier completion, RePaintGS would confidently reproduce that outlier rather than correct it.
  • A natural extension, not explored in the paper, is to automate reference selection by choosing the inpainted view with highest median confidence against all others, which would remove the user interaction and could be tested on the same datasets.
  • The Poisson-blending step ties success to the hole being a single connected region; the paper's own limitation discussion suggests multi-object holes and residual shadows would defeat the seamlessness claim.
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

4 major / 6 minor

Summary. The paper proposes RePaintGS, a 3D Gaussian Splatting method for 3D scene inpainting after object removal. The pipeline first reconstructs the background with an initial 3DGS stage, applies per-image segmentation, inpainting, and monocular depth estimation, and then derives per-view confidence scores by warping a user-selected reference inpainted view to each other view and measuring LPIPS similarity. An 'inpaint-3DGS' stage then optimizes the inpainted region with confidence-weighted depth and normal losses plus appearance losses against a progressively geometry-filtered warped reference. Experiments on five SPIn-NeRF scenes, two MipNeRF360 scenes, and three custom Blender scenes compare against reimplemented SPIn-NeRF, Gaussian Grouping, and Infusion, with ablations of the confidence weighting and of the individual loss terms.

Significance. The problem is practically important, and the method is a reasonable design: anchoring the inpainted reconstruction to a single reference view and weighting other views by their geometric similarity is a principled way to mitigate multi-view inpainting inconsistency. The progressive geometry-consistent warping with the P' mask is a sensible mechanism, and the paper is transparent about its reimplementation choices and limitations. However, the current evidence has two load-bearing weaknesses: the SPIn-NeRF quantitative evaluation uses a ground-truth test image as the reference, and the ablation of the warped-reference color loss shows mixed metrics. In addition, the geometry-consistency filter is self-referential because both depths compared in P' are trained toward the same monocular depth prior. These issues are fixable with additional experiments rather than fatal to the idea, so the appropriate decision is major revision.

major comments (4)
  1. [§4.2 (reference selection)] For the SPIn-NeRF experiment, the reference image is taken from the test set, whose images are the clean background ground truth, whereas the method is designed to use one user-selected inpainted image (Sec. 3.2). This gives the pseudo-ground-truth warp (Eqs. 10-11) access to the exact target appearance in the main quantitative comparison, biasing the comparison in favor of RePaintGS and making it impossible to interpret Table 2 as evaluating the proposed reference-guided inpainting protocol. Please rerun the SPIn-NeRF evaluation with a reference drawn from the inpainted training views, and state the reference-selection rule for all datasets.
  2. [§3.3, Eq. (8) and P′] The geometry-consistency mask P′ compares the rendered depth of the target view reprojected to the reference with the rendered depth of the reference, and the depth loss Eq. (8) trains both rendered depths toward the same monocular depth D_mono. In the inpainted region there are no COLMAP sparse points, so the scale/shift alignment of Eq. (3) cannot anchor D_mono; a smooth bias there is absorbed by the learned geometry and is not rejected by the threshold τ. Thus the 'geometrically consistent' region is not independently validated, and the paper's geometric-fidelity claim remains vulnerable to exactly the failure the confidence-weighting is meant to detect. Please add an external geometry check (e.g., multi-view stereo depth or plane-sweep validation) and report how often P′ agrees with it.
  3. [Table 5 (ablation of LiC)] The ablation of the warped-reference color loss (w/o LiC) improves SSIM (0.8208 vs 0.7913) and LPIPS (0.2209 vs 0.2267) relative to the full model, with only FID favoring the full model. Since LiC is the component that transfers reference appearance, this is not a minor numerical discrepancy; it undermines the central claim that the warped-reference appearance guidance improves realism. The qualitative argument in Fig. 10 should be quantified, or the loss weighting/implementation reconsidered, before the claim can be accepted.
  4. [§4.1, §4.2, Tables 2-3] SPIn-NeRF was reimplemented on 3DGS rather than run as published, and all numbers are single-run means without variances or significance tests. The comparison with original SPIn-NeRF (LPIPS 0.394 vs 0.487) is reported only as a one-line justification. Please report original SPIn-NeRF numbers for the same scenes, provide confidence intervals or multiple seeds, and include the closest reference-guided baseline, Mirzaei et al. [26], in the quantitative comparison.
minor comments (6)
  1. [Tables 2 and 3] The captions say 'Qualitative results' but the tables contain quantitative metrics; please rename them.
  2. [Eq. (6)] R_i and R_j are called extrinsic matrices, but the formula uses only rotations; please use full camera poses or clarify the notation.
  3. [§3.2-§3.3] The hyperparameters α, β, τ, λ_D, λ_iN, λ_iLPIPS and the bilateral filter settings are not given; please provide values and a sensitivity analysis for τ.
  4. [Figure 2] Figure 2 contains a stray Korean annotation ('RGB이미지 밝기를 +20, 40% 키운 버전') that should be removed.
  5. [Figure 7] The caption states (a)-(e) are from SPIn-NeRF, but the text says some are from MipNeRF360; please align the caption with the text.
  6. [§4.3 and Tables 2-3] It is unclear whether FID is computed per test image or on a pooled set; with only five or three scenes and bounding-box crops, the FID values should be reported with the number of samples used.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: reference-guided supervision is explicit, and headline metrics are computed on held-out ground-truth views.

full rationale

The paper's headline comparisons are externally grounded: SSIM, LPIPS, and FID in Tables 2 and 3 are computed on held-out test views (Section 4.2), and the reference view used as guidance is excluded from those evaluation views. The reference-guided mechanism is presented as an explicit optimization objective rather than as a prediction: Eqs. (10)-(11) directly compare the render against a warped reference, so matching the reference is the loss's definition, not a discovered result. The confidence score in Eq. (7) and the pseudo-ground truth use the same reference, which means reference-aligned appearance is enforced by construction, but this is a transparent design choice and does not contaminate the held-out comparison. The depth-dependence concern (Eq. (8) and the P' filter both rely on monocular depth or rendered depth trained against it) is a real robustness limitation acknowledged in Section 4.5, but the depth prior is an external model (Depth Anything v2), not a quantity fitted to the paper's own outputs, so it is not a circular derivation. No load-bearing self-citation or author-imported uniqueness argument appears in the reference list. The ablation in Table 5 is mixed but bears on effectiveness, not circularity. Overall: no significant circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the accuracy of monocular depth, the visibility of the reference, and the validity of LPIPS as a geometric consistency proxy. These are domain assumptions rather than standard math, and they are partially acknowledged in Section 4.5. No invented entities are introduced.

free parameters (3)
  • alpha, beta of confidence sigmoid winp = not reported
    Map confidence conf to inpainting weight via winp = sigmoid(alpha*(conf-beta)); values are chosen by hand and not listed in the paper.
  • tau, depth consistency threshold = not reported
    Controls the set P' of pixels deemed geometrically consistent for warping in Section 3.3; not specified numerically.
  • loss weights lambda_D, lambda_iN, lambda_iLPIPS = not reported
    Weights in Ltotal = LC + lambda_D LD + lambda_iN LiN + lambda_iLPIPS LiLPIPS; values are not reported.
assumptions (5)
  • domain assumption Most regions requiring inpainting are visible in the reference view
    Section 3.2.1 states this assumption directly; if the reference does not see the occluded region, warping cannot supply pseudo-ground truth for it.
  • domain assumption Monocular depth estimates are accurate enough for warping and confidence evaluation
    Eq. (6) warps the reference with estimated depth, and Eq. (7) confidence inherits any depth errors. Section 4.5 concedes dependence on depth.
  • domain assumption LPIPS on bilateral-filtered patches is a valid proxy for geometric/content consistency
    Section 3.2.2 uses (1-LPIPS) as confidence; the strong bilateral filter is intended to suppress texture differences, but this proxy is assumed, not validated independently.
  • domain assumption Z-buffer and bilinear interpolation fill warping holes adequately
    Section 3.2.1 says unmatched pixels are filled via bilinear interpolation; this can hallucinate content in large holes, affecting confidence and pseudo-ground truth.
  • domain assumption The user-selected reference is semantically appropriate
    Section 3.2 says the reference 'reflects the target that the user intends to build,' so an unhelpful reference will propagate its errors to all views.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RePaintGS: Reference-Guided Gaussian Splatting for Realistic and View-Consistent 3D Scene Inpainting." pith.science (2026). https://pith.science/paper/D76NPZCV

@misc{pith2026250708434,
  author       = {Pith},
  title        = {Pith review of: RePaintGS: Reference-Guided Gaussian Splatting for Realistic and View-Consistent 3D Scene Inpainting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D76NPZCV}},
  note         = {Machine review of arXiv:2507.08434}
}
read the original abstract

Radiance field methods, such as Neural Radiance Field or 3D Gaussian Splatting, have emerged as seminal 3D representations for synthesizing realistic novel views. For practical applications, there is ongoing research on flexible scene editing techniques, among which object removal is a representative task. However, removing objects exposes occluded regions, often leading to unnatural appearances. Thus, studies have employed image inpainting techniques to replace such regions with plausible content - a task referred to as 3D scene inpainting. However, image inpainting methods produce one of many plausible completions for each view, leading to inconsistencies between viewpoints. A widely adopted approach leverages perceptual cues to blend inpainted views smoothly. However, it is prone to detail loss and can fail when there are perceptual inconsistencies across views. In this paper, we propose a novel 3D scene inpainting method that reliably produces realistic and perceptually consistent results even for complex scenes by leveraging a reference view. Given the inpainted reference view, we estimate the inpainting similarity of the other views to adjust their contribution in constructing an accurate geometry tailored to the reference. This geometry is then used to warp the reference inpainting to other views as pseudo-ground truth, guiding the optimization to match the reference appearance. Comparative evaluation studies have shown that our approach improves both the geometric fidelity and appearance consistency of inpainted scenes.

Figures

Figures reproduced from arXiv: 2507.08434 by the authors.

Figure 1
Figure 1. Our method provides high-fidelity and view-consistent 3D scene inpainting using 3DGS, even in complex scenes with severe inconsistencies across [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Limitation in image inpainting. For complex scenes, image inpainting Figure 3: Limitation in image inpainting. For complex scenes, image inpainting [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Confidence evaluation of inpainted region. Each region’s confidence [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figures from the paper (7 more)
Figure 6
Figure 6. Figure 6: Geometry-Consistent Warping. When warping a reference view to a [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Visual comparison of test set renderings. This figure presents sample renderings from the novel test view: (a)–(c) are from SPIn-NeRF, (d)–(e) from Figure 7: Visual comparison of test set renderings. This figure presents sample renderings from the novel test view: (a)–…
Figure 7
Figure 7. Figure 7: Visual comparison of test set renderings. This figure presents sample renderings from the novel test view: (a)–(c) are from SPIn-NeRF, (d)–(e) from Figure 7: Visual comparison of test set renderings. This figure presents sample renderings from the novel test view: (a)–…
Figure 8
Figure 8. Figure 8: Qualitative comparison on three novel views. This figure demonstrates the e [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: High-fidelity appearance from a single reference view. We present sample novel views from wide-view datasets (MipNeRF360 [2] and our custom scenes), Figure 9: High-fidelity appearance from a single reference view. We present sample novel views from wide-view datasets (…
Figure 10
Figure 10. Figure 10: Qualitative examples from the ablation study on loss term combinations. The figure shows sample results on our custom dataset. Excluding normal, [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Example of degraded performance. Due to the reliance on inpainted [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 10 canonical work pages

  1. [26]

    Mirzaei, T

    A. Mirzaei, T. Aumentado-Armstrong, M. A. Brubaker, J. Kelly, A. Levinshtein, K. G. Derpanis, I. Gilitschenski, Reference-guided con- trollable inpainting of neural radiance fields, in: 2023 IEEE /CVF Inter- national Conference on Computer Vision (ICCV), IEEE Computer Soci- ety, Los Alamitos, CA, USA, 2023, pp. 17769–17779. doi:10.1109/ ICCV51070.2023.016...

    arXiv 2023

  2. [1]

    Mildenhall, P

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, R. Ng, Nerf: Representing scenes as neural radiance fields for view syn- thesis, Commun. ACM 65 (1) (2021) 99–106. doi:10.1145/3503250. URL https://doi.org/10.1145/3503250

    doi:10.1145/3503250 2021

  3. [3]

    M ¨uller, A

    T. M ¨uller, A. Evans, C. Schied, A. Keller, Instant neural graphics prim- itives with a multiresolution hash encoding, ACM Trans. Graph. 41 (4) (2022) 102:1–102:15. doi:10.1145/3528223.3530127. URL https://doi.org/10.1145/3528223.3530127

    arXiv 2022

  4. [4]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimkuehler, G. Drettakis, 3d gaussian splat- ting for real-time radiance field rendering, ACM Trans. Graph. 42 (4) (jul 2023). doi:10.1145/3592433. URL https://doi.org/10.1145/3592433

    doi:10.1145/3592433 2023

  5. [6]

    Cheng, X

    K. Cheng, X. Long, K. Yang, Y . Yao, W. Yin, Y . Ma, W. Wang, X. Chen, Gaussianpro: 3d gaussian splatting with progressive propagation, in: arXiv preprint arXiv:2402.14650, 2024. doi:10.48550/arXiv.2402. 14650. URL https://doi.org/10.48550/arXiv.2402.14650

    arXiv doi:10.48550/arxiv.2402.14650 2024

  6. [8]

    Rudnev, M

    V . Rudnev, M. Elgharib, W. Smith, L. Liu, V . Golyanik, C. Theobalt, Nerf for outdoor scene relighting, in: Computer Vision – ECCV 2022, Springer Nature Switzerland, Cham, 2022, pp. 615–631. doi:10.1007/ 978-3-031-19787-1{\_}35 . URL https://doi.org/10.1007/978-3-031-19787-1{_}35

    doi:10.1007/978-3-031-19787-1 2022

  7. [10]

    J. Gao, C. Gu, Y . Lin, Z. Li, H. Zhu, X. Cao, L. Zhang, Y . Yao, Relightable 3d gaussians: Realistic point cloud relighting with brdf decomposition and ray tracing, in: Computer Vision – ECCV 2024, Springer Nature Switzerland, Cham, 2025, pp. 73–89. doi:10.1007/ 978-3-031-72995-9{\_}5 . URL https://doi.org/10.1007/978-3-031-72995-9{_}5

    doi:10.1007/978-3-031-72995-9 2024

  8. [12]

    K. Liu, F. Zhan, M. Xu, C. Theobalt, L. Shao, S. Lu, Stylegaussian: In- stant 3d style transfer with gaussian splatting, in: SIGGRAPH Asia 2024 Technical Communications, SA ’24, Association for Computing Machin- ery, New York, NY , USA, 2024.doi:10.1145/3681758.3698002. URL https://doi.org/10.1145/3681758.3698002

    arXiv 2024

Show all 39 references
  1. [13]

    B. Yang, Y . Zhang, Y . Xu, Y . Li, H. Zhou, H. Bao, G. Zhang, Z. Cui, Learning object-compositional neural radiance field for editable scene rendering, in: 2021 IEEE /CVF International Conference on Computer Vision (ICCV), IEEE Computer Society, Los Alamitos, CA, USA, 2021, p...

    2021

  2. [16]

    E. Chan, M. Monteiro, P. Kellnhofer, J. Wu, G. Wetzstein, pi-gan: Peri- odic implicit generative adversarial networks for 3d-aware image synthe- sis, in: arXiv, 2020. doi:10.1109/CVPR46437.2021.00574. URL https://doi.org/10.1109/CVPR46437.2021.00574

    2020

  3. [17]

    Suvorov, E

    R. Suvorov, E. Logacheva, A. Mashikhin, A. Remizova, A. Ashukha, A. Silvestrov, N. Kong, H. Goka, K. Park, V . Lempitsky, Resolution- robust large mask inpainting with fourier convolutions, in: Proceedings of the IEEE /CVF Winter Conference on Applications of Computer Vi- sion...

    2022

  4. [18]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, B. Ommer, High- resolution image synthesis with latent di ffusion models, in: Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recog- nition (CVPR), 2022, pp. 10684–10695. doi:10.1109/CVPR52688. 2022.01042. U...

    2022

  5. [19]

    Podell, Z

    D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. M ¨uller, J. Penna, R. Rombach, Sdxl: Improving latent di ffusion models for high-resolution image synthesis (2023). arXiv:2307.01952, doi: 10.48550/arXiv.2307.01952. URL https://doi.org/10.48550/arXiv.2307.01952

    2023 doi arXiv

  6. [21]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, O. Wang, The unreason- able effectiveness of deep features as a perceptual metric, in: Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2018. doi:10.1109/CVPR.2018.00068. URL https://doi.or...

    2018

  7. [22]

    Y . Yin, Z. Fu, F. Yang, G. Lin, Or-nerf: Object removing from 3d scenes guided by multiview segmentation with neural radiance fields (2023). arXiv:2305.10503, doi:10.48550/arXiv.2305.10503. URL https://doi.org/10.48550/arXiv.2305.10503

    2023 doi arXiv

  8. [23]

    Huang, H

    J. Huang, H. Yu, J. Zhang, H. Nait-Charif, Point’n move: In- teractive scene object manipulation on gaussian splatting radiance fields (2024). arXiv:https://digital-library.theiet.org/ doi/pdf/10.1049/ipr2.13190, doi:10.1049/ipr2.13190. URL https://doi.org/10.1049/ipr2.13190

    2024 doi

  9. [24]

    D. Wang, T. Zhang, A. Abboud, S. S ¨usstrunk, Inpaintnerf360: Text- guided 3d inpainting on unbounded neural radiance fields (2023).arXiv: 2305.15094, doi:10.48550/arXiv.2305.15094. URL https://doi.org/10.48550/arXiv.2305.15094

    2023 doi arXiv

  10. [25]

    M. Ye, M. Danelljan, F. Yu, L. Ke, Gaussian grouping: Segment and edit anything in 3d scenes, in: Computer Vision - ECCV 2024, Springer-Verlag, Berlin, Heidelberg, 2024, p. 162–179. doi:10.1007/ 978-3-031-73397-0{\_}10 . URL https://doi.org/10.1007/978-3-031-73397-0{_}10

    2024 doi

  11. [27]

    Shen, H.-K

    I.-C. Shen, H.-K. Liu, B.-Y . Chen, Nerf-in: Free-form inpainting for pre- trained nerf with rgb-d priors, IEEE Computer Graphics and Applications 44 (2) (2024) 100–109. doi:10.1109/MCG.2023.3336224

    2024

  12. [29]

    Z. Liu, H. Ouyang, Q. Wang, K. L. Cheng, J. Xiao, K. Zhu, N. Xue, Y . Liu, Y . Shen, Y . Cao, Infusion: Inpainting 3d gaussians via learn- ing depth completion from diffusion prior (2024). arXiv:2404.11613, doi:10.48550/arXiv.2404.11613. URL https://doi.org/10.48550/arXiv.2404.11613

    2024 doi arXiv

  13. [30]

    Fischer, J

    R. Fischer, J. Roßkamp, T. Hudcovic, A. Schlegel, G. Zachmann, Inpaint- ing of depth images using deep neural networks for real-time applica- tions, in: G. Bebis, G. Ghiasi, Y . Fang, A. Sharf, Y . Dong, C. Weaver, Z. Leo, J. J. LaViola Jr., L. Kohli (Eds.), Advances in Visual...

    2023 doi

  14. [31]

    D. Wang, T. Zhang, A. Abboud, S. S¨usstrunk, Innerf360: Text-guided 3d- consistent object inpainting on 360-degree neural radiance fields, in: Pro- ceedings of the IEEE /CVF Conference on Computer Vision and Pattern 15 Recognition (CVPR), 2024. doi:10.1109/CVPR52733.2024.01205...

    2024

  15. [33]

    Chung, J

    J. Chung, J. Oh, K. M. Lee, Depth-regularized optimization for 3d gaus- sian splatting in few-shot images, in: Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition (CVPR) Work- shops, 2024, pp. 811–820. doi:10.1109/CVPRW63382.2024.00086. URL https...

    2024

  16. [34]

    Turkulainen, X

    M. Turkulainen, X. Ren, I. Melekhov, O. Seiskari, E. Rahtu, J. Kannala, Dn-splatter: Depth and normal priors for gaussian splatting and meshing, in: 2025 IEEE/CVF Winter Conference on Applications of Computer Vi- sion (W ACV), 2025, pp. 2421–2431.doi:10.1109/WACV61041.2025. 00...

    2025

  17. [35]

    Z. Zhu, Z. Fan, Y . Jiang, Z. Wang, Fsgs: Real-time few-shot view syn- thesis using gaussian splatting, in: Computer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Pro- ceedings, Part XXXIX, Springer-Verlag, Berlin, Heidelberg, 2024, ...

    2024 doi

  18. [36]

    Roessle, J

    B. Roessle, J. T. Barron, B. Mildenhall, P. P. Srinivasan, M. Nießner, Dense depth priors for neural radiance fields from sparse input views, in: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR), 2022, pp. 12882–12891. doi:10.1109/CVPR52688. 2022.01...

    2022

  19. [37]

    J. L. Sch ¨onberger, J.-M. Frahm, Structure-from-motion revisited, in: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 4104–4113. doi:10.1109/CVPR.2016.445. URL https://doi.org/10.1109/CVPR.2016.445

    2016 doi

  20. [38]

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. R ¨adle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V . Alwala, N. Carion, C.-Y . Wu, R. Girshick, P. Doll ´ar, C. Feichtenhofer, Sam 2: Segment anything in images and videos (2024). arXiv:2408.00714, do...

    2024 doi arXiv

  21. [39]

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, H. Zhao, Depth anything v2, arXiv:2406.09414 (2024). doi:10.48550/arXiv.2406. 09414. URL https://doi.org/10.48550/arXiv.2406.09414

    2024 doi arXiv

  22. [40]

    Tomasi, R

    C. Tomasi, R. Manduchi, Bilateral filtering for gray and color im- ages, in: Sixth International Conference on Computer Vision (IEEE Cat. No.98CH36271), 1998, pp. 839–846. doi:10.1109/ICCV.1998. 710815

    1998 doi

  23. [41]

    M.-S. Kwak, J. Song, S. Kim, Geconerf: few-shot neural radiance fields via geometric consistency, in: Proceedings of the 40th International Con- ference on Machine Learning (ICML), 2023. URL https://dl.acm.org/doi/10.5555/3618408.3619152

    2023

  24. [42]

    P ´erez, M

    P. P ´erez, M. Gangnet, A. Blake, Poisson image editing, in: ACM SIG- GRAPH 2003 Papers, SIGGRAPH ’03, Association for Computing Ma- chinery, New York, NY , USA, 2003, p. 313–318. doi:10.1145/ 1201775.882269. URL https://doi.org/10.1145/1201775.882269

    2003

  25. [43]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, B. Ommer, High- resolution image synthesis with latent di ffusion models, in: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 10674–10685. doi:10.1109/CVPR52688.2022. 01042. URL https://doi.or...

    2022

  26. [44]

    URL https://www.blender.org

    Blender Online Community, Blender 3.5.1 - a 3D modelling and render- ing package, Blender Foundation, Amsterdam, The Netherlands (2023). URL https://www.blender.org

    2023

  27. [45]

    blenderkit.com, accessed: 2025-04

    BlenderKit, Online asset library for blender, https://www. blenderkit.com, accessed: 2025-04

    2025

  28. [46]

    Heusel, H

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, S. Hochreiter, Gans trained by a two time-scale update rule converge to a local nash equi- librium, in: Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, Curran Associates Inc...

    2017

  29. [47]

    Z. Wang, A. Bovik, H. Sheikh, E. Simoncelli, Image quality assessment: from error visibility to structural similarity, IEEE Transactions on Image Processing 13 (4) (2004) 600–612. doi:10.1109/TIP.2003.819861

    2004

  30. [49]

    Kumari, B

    N. Kumari, B. Zhang, R. Zhang, E. Shechtman, J.-Y . Zhu, Multi-concept customization of text-to-image diffusion, in: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 1931–

    2023

  31. [1941]

    URL https://doi.org/10.1109/CVPR52729.2023.00192 16

    doi:10.1109/CVPR52729.2023.00192. URL https://doi.org/10.1109/CVPR52729.2023.00192 16

    2023

Pith tools

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