Pith. sign in

REVIEW 4 major objections 6 minor 63 references

3D-GIMP shows that a single generative inpainting of a reference view, propagated across all views with a 3D-aware PatchMatch, is enough to match multi-view diffusion object removal in quality while being far cheaper and more view-consisten

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

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

T0 review

2026-08-01 09:23 UTC pith:DGTCTZJA

load-bearing objection Single-reference diffusion plus PatchMatch propagation is a sensible speed play, but the view-consistency claim rides on a depth prior that is neither named nor shown to be robust. the 4 major comments →

arxiv 2607.20789 v1 pith:DGTCTZJA submitted 2026-07-22 cs.CV cs.GR

3D-GIMP: When 3D Gaussian Inpainting Meets PatchMatch

Xuening Tian , Dieter Schmalstieg , Shohei Mori This is my paper
classification cs.CV cs.GR
keywords 3D Gaussian Splattingobject removalPatchMatchdepth completionPoisson blendingmulti-view consistencyimage inpaintingnovel view synthesis
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

3D-GIMP sets out to fix a central weakness of diffusion-based 3D scene editing: when every training view is inpainted independently, the generative model hallucinates different details per view, and the 3D reconstruction turns blurry or distorted. The paper's proposal is to invert the usual order—perform a single generative inpainting on one key reference view, then treat that result as an appearance prior and propagate it to all other views with a 3D-aware PatchMatch algorithm. A Poisson-blending-based depth completion step aligns the reference depth scale so that the patch correspondences reproject correctly. On 360° benchmarks the method matches or beats per-view diffusion in PSNR, LPIPS, and FID while cutting inpainting time from roughly an hour to six minutes and improving cross-view consistency.

Core claim

The central claim is that high-fidelity, view-consistent object removal in 3D Gaussian Splatting does not require per-view generative inpainting. Instead, the paper builds a patch mapping field in the inpainted reference view and reprojects it through a scale-consistent depth map, so every target view is filled by sampling and blending local patches that already agree with the scene's geometry and photometry. The paper argues this 'reconstructive consistency' avoids hallucination drift and preserves high-frequency details at arbitrary resolution, and supports it with experiments where the proposed pipeline reaches 19.09 PSNR / 0.248 LPIPS / 53.11 FID on the IMFine dataset in 6 minutes, compa

What carries the argument

The load-bearing mechanism is a 3D-aware PatchMatch mapping field: a nearest-neighbour field f that explains each inpainted pixel in the reference view by a similar patch elsewhere in the same image, with a visibility score biasing the search toward patches that stay visible in many views. This field is propagated to other cameras by reprojecting matched pairs through a depth map that is completed inside the removal region by a discrete Poisson equation, whose gradient guidance comes from a scaled monocular disparity prior and whose boundary values come from the 2DGS rasterizer. The mapping field lets the method resample local patches in each target view, then blends them via splatting and d

Load-bearing premise

The weakest link is the Poisson-completed depth map inside the removed region: if the scaled monocular prior (with the fixed sf=0.1) does not reconstruct the true geometry, the patch reprojections misalign and the cross-view consistency that the method's claim rests on degrades, as the paper itself acknowledges for 'extreme cases'.

What would settle it

Run the pipeline on a scene with a known wrong monocular depth prior inside the ROI (e.g., a mirror, a transparent object, or a strongly concave surface), and measure PSNR of propagated inpainted views: if the patch-reprojected patches are visibly misaligned and the view-consistency advantage over per-view diffusion disappears, the central claim fails on such geometry. A cheaper check: vary sf around 0.1 on the 'rocks' scene and observe whether the reported FID/PSNR gap closes—if a single fixed sf cannot serve varied shapes, the claim of scale-consistent propagation is conditional, not general

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Per-view diffusion inpainting for 3DGS object removal can be replaced by one generative pass plus geometric propagation, cutting inpainting cost by about 90% on the tested scenes.
  • Because propagation is geometric and patch-based, the method keeps high-frequency details and view-dependent effects (such as specular reflections) that simple image warping loses, and it scales to higher resolutions without additional generative passes.
  • The patch mapping field acts as a structural prior: feeding it to a diffusion refiner (as in the paper's Ours+LeftRefill ablation) reduces hallucinations and yields the best reported PSNR and LPIPS, at higher runtime.
  • Better multi-view consistency translates directly to cleaner novel-view synthesis, since the inpainted Gaussian primitives are back-projected from views that already agree.

Where Pith is reading between the lines

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

  • The 'generate once, propagate through geometry' recipe generalizes beyond object removal: the same single-reference edit plus PatchMatch propagation could handle object insertion, texture replacement, or reflection editing in 3DGS, since the mechanism only requires an edited reference and a completed depth map.
  • The scene-specific scale factor sf is the obvious knob to automate: a search that maximizes photometric consistency of propagated patches across neighboring views would remove the manual tuning and likely improve robustness on hard geometry.
  • The explicit mapping field could be replaced or refined by a learned correspondence network, potentially making the propagation differentiable and end-to-end trainable while keeping the geometric grounding.
  • The visibility-score weighting suggests a principled reference-view selection criterion: pick the camera whose view of the removed region is most visible across the training set, which the paper notes would reduce sampling gaps and stretching artifacts.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. 3D-GIMP proposes a hybrid object-removal pipeline for 3D Gaussian Splatting. Instead of running a diffusion inpainting model on every training view, it performs a single generative inpainting on one reference view, constructs a PatchMatch mapping field from the inpainted reference, completes the reference depth map inside the removal region with a Poisson equation guided by scaled monocular disparity, and then propagates the mapping to all other views via 3D warping. The resulting per-view inpainted images are used to re-initialize and finetune the Gaussian primitives. The paper reports about a 10x speedup over per-view diffusion baselines on IMFine (6 min vs 1 h) and claims competitive quality with better view consistency. The method is evaluated on IMFine, 360-USID, and Mip-NeRF 360 scenes, including an ablation and a comparison with a video-diffusion baseline.

Significance. If the central claims hold, the single-reference + PatchMatch propagation design is a meaningful efficiency contribution to 3D Gaussian object removal: it reduces the dependence on stochastic multi-view diffusion and achieves a large practical speedup. The reported FID gains on IMFine and 360-USID, and the hybrid 'Ours + LeftRefill' variant that improves LPIPS, are encouraging. However, the paper's strongest conceptual claim — multi-view consistency — is never directly measured, and the entire propagation chain rests on a depth-completion step whose monocular prior is left unidentified and whose scale parameter is tuned per test scene. The significance of the results is therefore conditional on additional validation of the depth propagation and a direct consistency metric.

major comments (4)
  1. [Section 3.3, Eqs. (4)-(6); Section 5; Supp. A.4] The depth-completion step that drives the propagation is load-bearing, but its core components are not specified or validated. The monocular disparity prior d_m is never identified (no model name, source, or license), and the scale factor s_f is set to 0.1 as 'the most consistent quality across all test scenes' (Supp. A.4) while Section 5 admits it is 'currently scene-specific' and that when the solver fails, 'the precision of subsequent patch-warping steps is significantly degraded.' Since Algorithm 1 warps both p0 and q0 through w_j(·) using D(p0), any depth error directly misplaces the copied patches and undermines the asserted multi-view consistency. Please identify the monocular depth model, report sensitivity of the final metrics to s_f across all scenes (not just 'rocks'), and provide a quantitative check of completed depth, e.g., against ground-truth depth or by correlating depth
  2. [Abstract and Section 4.3 (Table 1)] The paper repeatedly claims that 3D-GIMP 'outperforms ... in view consistency' (Abstract, Section 1, Section 4.3), but no quantitative sanity metric for multi-view agreement is reported anywhere in Tables 1-4. The only consistency-adjacent number is SSIM in the GEN3C comparison (Table 3), which is computed on a single view's object mask, not across views. This is a central, load-bearing claim; it should be tested directly, e.g., by measuring pixel variance or pairwise PSNR/SSIM between overlapping inpainted views, or by rendering novel views and comparing them to ground truth when available. Without such a metric, the claimed advantage over multi-view diffusion methods is not demonstrated.
  3. [Table 1] The quality claims are only partially supported by the numbers. On IMFine, Ours (2DGS) has a worse LPIPS than AuraFusion360 (0.248 vs 0.173); on 360-USID, Ours (2DGS) has a worse PSNR (15.02 vs 16.22). The abstract says 'competitive inpainting quality', which is defensible when FID and runtime are considered, but the text in Section 4.3 should qualify the claim rather than implying global superiority. Moreover, the results are averaged over 8 scenes (IMFine) and 7 scenes (360-USID) with a scene-tuned s_f; per-scene results with standard deviations are needed to assess whether the averages are driven by a few favorable scenes. Please provide the per-scene tables with significance information.
  4. [Section 4.2, Supp. A.3] The evaluation protocol for the reference view is ambiguous and potentially unfair. The paper states 'we use the first ground truth frame from the testing set as the reference view' and Supp. A.3 says 'we use the same view as the ground truth for comparison.' If the reference view is a clean ground-truth background image, then the diffusion pass on the reference is not actually exercised, and methods that do not receive such a reference (e.g., SPInNeRF, NeRFiller, GScream) are at a disadvantage. Please clarify whether the 'ground truth frame' still contains the object to be removed, how the reference is used by each baseline, and what exactly is meant by 'same view as the ground truth.' If a clean reference is used, justify this protocol or adopt a common protocol with and without a reference.
minor comments (6)
  1. [Section 4.4] Typos: 'ketchen' should be 'kitchen' (twice in Section 4.4 and in the Table 2 caption).
  2. [Section 3.3, Eq. (7)] The visibility score is defined as v(p) in the text but Eq. (7) writes v(q); the notation is inconsistent and should be aligned.
  3. [Algorithm 1] ReTargeting, PixelSampling, SortingAndBlending, and UpdateCameraState are not defined in the main text or the supplement. Please add pseudocode or a precise citation for each.
  4. [Section 3.3] The sentence 'Ideally, we wish to construct the patch mapping field f only once in the reference view and reuse these assignments in all reference views I_j' uses 'reference views' where it should be 'target views'.
  5. [Supp. A.4, Figure 11] The axis annotation '0.0137 / 0.1655 Disp.' is unclear. Please specify the units and what the two numbers denote.
  6. [Section 4.4] The sentence 'While the core of our method serves as a highly efficient standalone solution, establishes a new state-of-the-art.' is incomplete and grammatically broken; it also overstates the evidence given the mixed results in Table 1.

Circularity Check

0 steps flagged

No circularity found: the PatchMatch propagation and Poisson depth completion are an empirical construction benchmarked externally; the scene-specific scale factor is a robustness limitation, not a definitional loop.

full rationale

The paper's derivations do not define any predicted quantity in terms of its own inputs. The patch mapping field f is computed on the inpainted reference view and propagated via reprojection w_j(p), using the completed depth map D from Eqs. (4)-(6). That depth map is filled from a monocular disparity prior scaled by s_f with rasterizer boundary conditions; it is not fitted to the PSNR/LPIPS/FID targets. No uniqueness theorem or load-bearing self-citation is invoked: the only self-references (InpaintFusion [Mori et al. 2020], Good Keyframes [Mori et al. 2022]) are motivational or about reference-view selection, not used to force the algorithm. The central claim is validated against external baselines (AuraFusion360, IMFine, GEN3C, etc.) on held-out ground-truth views, so the reported numbers are not equivalent to the method's definitions. I do flag the admitted limitation in Sec. 5 and Supp. A.4: the depth-completion scale s_f=0.1 was chosen as 'most consistent quality across all test scenes,' and the paper states that when the Poisson solver fails, 'the precision of subsequent patch-warping steps is significantly degraded.' That is a genuine robustness and possible test-set-tuning concern, but it is not circularity: the parameter is not disguised as a prediction, and the propagation pipeline remains a constructive recipe whose success depends on the unknown monocular depth model. For the same reason, the use of the first ground-truth frame as reference view is a data-protocol question, not a reduction of the output to the input.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

No invented physical or conceptual entities: the 'mapping field' is the standard nearest-neighbor field from PatchMatch, and the visibility score is a standard multi-view consistency measure. The ledger is dominated by hand-set hyperparameters (sf, s, lambda_s, lambda_v, sigma_0) plus a small set of domain assumptions about 2DGS depth accuracy, patch photometric consistency across views, and the copy-and-paste ceiling. The two most load-bearing assumptions - patch photometric consistency across views and local correctness of the unnamed monocular depth prior - are both self-acknowledged as fragile in Sec. 5.

free parameters (6)
  • sf (scale of monocular disparity guidance) = 0.1
    Log-scale factor in the Poisson guidance field (Eq. 5); empirically chosen because it 'yielded the most consistent quality across all test scenes' (Supp. A.4); Sec. 5 admits the optimal value is scene-specific, making the reported results partly a function of hand-tuning on the evaluation scenes.
  • s (guard-band enlargement) = 1.2
    Sec. 3.2: 'We empirically chose s=1.2, as it needs to be larger than the projected ROI to have a clear removal boundary.' Directly defines the inpainted region T and hence the mask over which PSNR/LPIPS are computed.
  • lambda_s (spatial smoothness weight) = 0.001
    Cost function Eq. 8; fixed in Sec. 4.1; controls smoothness/clustering of the patch field.
  • lambda_v (visibility weight) = not reported
    Cost function Eq. 8 weights the multi-view visibility term; the value is never reported, so the effective cost function is not fully reproducible.
  • sigma_0 (SH outlier weight strictness) = not reported
    Supplementary Eq. 12 controls the weighting of photometric outliers when initializing SH coefficients for new Gaussian primitives; value unreported.
  • Reference camera selection = first test-frame camera
    Supplementary A.3 shows per-scene PSNR varies by several dB with reference choice and some scenes are 'more sensitive to camera choice'; the main tables use a single fixed choice, so reported numbers depend on this selection.
axioms (6)
  • domain assumption 2D Gaussian Splatting ray-splat intersection produces accurate, scale-consistent depth and normal maps for the visible scene.
    Sec. 3.1: the approach switches from 3DGS to 2DGS to obtain reliable depth; the rasterized boundary depth d_g anchors the Poisson solve (Eq. 6). If 2DGS depth is biased, all warping inherits the bias.
  • domain assumption Surface patches in the reference view remain photometrically similar in neighboring views, so a mapping field computed once can be reprojected and reused.
    Sec. 3.3: 'Local surface patches typically exhibit high photometric correlation in neighboring views'; this premise makes cross-view propagation valid, and the paper admits failures under occlusion and limited views (Sec. 5).
  • domain assumption The disoccluded content in the reference view can be represented as copies of visible patches from an unoccluded source region.
    PatchMatch copies from I0\T0; Sec. 3 states 'the copy-and-paste nature of patch-match restricts the content in the inpainted area to elements that are already present in one of the views'. Unique background that never appears outside the ROI cannot be reconstructed.
  • standard math Poisson equation with Dirichlet boundary conditions and scaled monocular guidance yields scale-consistent completed depth.
    Eqs. 4-6: standard Poisson blending guarantees seamless gradient-domain interpolation into the boundary, but correctness relative to true geometry depends on the guidance field, i.e., on the monocular prior and sf.
  • domain assumption The monocular depth prior d_m is locally correct up to latent scale.
    Sec. 3.3 introduces d_m via zero-shot depth estimators; the model is never identified, and if its local gradients are wrong the completed depth is wrong (failure admitted in Sec. 5).
  • domain assumption A diffusion-based reference inpainter (Nano Banana Pro / Gemini 2.5) provides a trustworthy appearance prior.
    Sec. 4.1: the reference-view inpainting is outsourced to a closed commercial API; the paper's quality ceiling is this black box, and any hallucination in the reference view is propagated to every view.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of 3D-GIMP: When 3D Gaussian Inpainting Meets PatchMatch." pith.science (2026). https://pith.science/paper/DGTCTZJA

@misc{pith2026260720789,
  author       = {Pith},
  title        = {Pith review of: 3D-GIMP: When 3D Gaussian Inpainting Meets PatchMatch},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DGTCTZJA}},
  note         = {Machine review of arXiv:2607.20789}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recent advances in 3D scene editing have leveraged iterative diffusion models to update input views. However, this process is computationally expensive and struggles to produce sharp details. Meanwhile, ``hallucination drift'' frequently introduces multi-view inconsistencies, leading to structural artifacts when rendering novel viewpoints. To address this problem, we present 3D-GIMP (3D Gaussian Inpainting Meets Patch Matching), a novel hybrid paradigm designed for high-fidelity object removal in 3D Gaussian Splatting. Instead of diffusing every view, 3D-GIMP performs a single generative inpainting on a key reference view, which serves as an appearance prior. We then introduce a 3D-aware PatchMatch algorithm to propagate these reference textures across all remaining views via correspondence matching, effectively bypassing the stochastic nature of frame-by-frame diffusion. By prioritizing reconstructive consistency over iterative generation, 3D-GIMP maintains high-frequency details across arbitrary resolutions while ensuring a mathematically consistent 3D reconstruction. Our experiments demonstrate that 3D-GIMP not only achieves competitive inpainting quality as previous methods using diffusion in multiple views, but also outperforms these methods in rendering speed and view consistency.

Figures

Figures reproduced from arXiv: 2607.20789 by Dieter Schmalstieg, Shohei Mori, Xuening Tian.

Figure 1
Figure 1. Figure 1: The proposed pipeline for view-consistent object removal in 3DGS. We initiate object removal in a reference view [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Random hallucinations of an image inpainting [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Data flow of our proposed pipeline. Our method selects a training view as a reference for rendering both before and [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of the process of warping the patch [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Generative refinement. Our method maintains struc [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison between depth completion of AuraFu [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: We also tested a hybrid configuration which combines our [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Per-scene PSNR (mean ± std) across 8 evaluation scenes. A.4 Scale parameter 𝑠𝑓 for Depth Completion While the Poisson solver is effective for depth completion, they often produce over-smoothed results, particularly in irregular re￾gions along Dirichlet boundaries. To address this, we introduce a scaling factor sf that adaptively enhances or compresses the gradi￾ents derived from the inferred monocular dept… view at source ↗
Figure 9
Figure 9. Figure 9: Per-scene LPIPS score (mean ± std) across 8 evalua￾tion scenes. dataset [Shi et al. 2025] ( [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Ablation study results. Our method effectively inpaints the scene by leveraging surrounding pixels from the available [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Impact of the scale parameter 𝑠𝑓 . Each image shows the disparity map generated by our method for different scale factors. The parameter 𝑠𝑓 regulates the influence of the monocular guidance field 𝑑𝑚 relative to the boundary constraints. In our experiments, we empirically set 𝑠𝑓 = 0.1, as this value yielded the most consistent quality across all test scenes [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative results compared with one of most representative 360 inpainting work AuraFusion360 [PITH_FULL_IMAGE:figures/full_fig_p013_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: We provide qualitative comparisons against state-of-the-art methods on IMFine dataset. While InFusion [Liu et al [PITH_FULL_IMAGE:figures/full_fig_p014_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Comparison of 3DGS inpainting methods across three downsampled resolution ( [PITH_FULL_IMAGE:figures/full_fig_p015_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative comparisons with GEN3C [Ren et al [PITH_FULL_IMAGE:figures/full_fig_p015_15.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

63 extracted references · 5 linked inside Pith

  1. [1]

    Structured Variational Inference Procedures and their Realizations (as incol)

    Dan Geiger and Christopher Meek. Structured Variational Inference Procedures and their Realizations (as incol). Proceedings of Tenth International Workshop on Artificial Intelligence and Statistics, The Barbados

  2. [2]

    Stan W. Smith. An experiment in bibliographic mark-up: Parsing metadata for XML export. Proceedings of the 3rd. annual workshop on Librarians and Computers. 2010

    2010

  3. [3]

    Derpanis and Jonathan Kelly and Marcus A

    Ashkan Mirzaei and Tristan Aumentado-Armstrong and Konstantinos G. Derpanis and Jonathan Kelly and Marcus A. Brubaker and Igor Gilitschenski and Alex Levinshtein , year=

  4. [4]

    2025 , pages =

    Wu, Chung-Ho and Chen, Yang-Jung and Chen, Ying-Huan and Lee, Jie-Ying and Ke, Bo-Hsu and Mu, Chun-Wei Tuan and Huang, Yi-Chuan and Lin, Chin-Yang and Chen, Min-Hung and Lin, Yen-Yu and Liu, Yu-Lun , title =. 2025 , pages =

    2025

  5. [5]

    GScream: Learning 3D Geometry and Feature Consistent Gaussian Splatting for Object Removal , author=

  6. [6]

    arXiv preprint arXiv:2404.11613 , year=

    InFusion: Inpainting 3D Gaussians via Learning Depth Completion from Diffusion Prior , author=. arXiv preprint arXiv:2404.11613 , year=

    Pith/arXiv arXiv

  7. [7]

    NeRFiller: Completing Scenes via Generative 3D Inpainting , author =

  8. [8]

    2025 , eprint=

    Perspective-aware 3D Gaussian Inpainting with Multi-view Consistency , author=. 2025 , eprint=

    2025

  9. [9]

    2025 , eprint=

    IMFine: 3D Inpainting via Geometry-guided Multi-view Refinement , author=. 2025 , eprint=

    2025

  10. [10]

    Inpaint360GS: Efficient Object-Aware 3D Inpainting via Gaussian Splatting for 360

    Shaoxiang Wang and Shihong Zhang and Christen Millerdurai and Rüdiger Westermann and Didier Stricker and Alain Pagani , year=. Inpaint360GS: Efficient Object-Aware 3D Inpainting via Gaussian Splatting for 360. 2511.06457 , archivePrefix=

    arXiv

  11. [11]

    InstaInpaint: Instant 3D-Scene Inpainting with Masked Large Reconstruction Model , author=

  12. [12]

    arXiv preprint arXiv:2408.00714 , year=

    SAM 2: Segment Anything in Images and Videos , author=. arXiv preprint arXiv:2408.00714 , year=

    Pith/arXiv arXiv

  13. [13]

    arXiv:2304.02643 , year=

    Segment Anything , author=. arXiv:2304.02643 , year=

    Pith/arXiv arXiv

  14. [14]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering , journal = TOG, number =

    Kerbl, Bernhard and Kopanas, Georgios and Leimk. 3D Gaussian Splatting for Real-Time Radiance Field Rendering , journal = TOG, number =

  15. [15]

    PixMix: A real-time approach to high-quality Diminished Reality , year=

    Herling, Jan and Broll, Wolfgang , booktitle=. PixMix: A real-time approach to high-quality Diminished Reality , year=

  16. [16]

    2009 , volume =

    Barnes, Connelly and Shechtman, Eli and Finkelstein, Adam and Goldman, Dan B , title =. 2009 , volume =

    2009

  17. [17]

    2025 , eprint=

    GSFix3D: Diffusion-Guided Repair of Novel Views in Gaussian Splatting , author=. 2025 , eprint=

    2025

  18. [18]

    3d gaussian inpainting with depth-guided cross-view consistency , author=

  19. [19]

    and Perez, P

    Criminisi, A. and Perez, P. and Toyama, K. , booktitle=CVPR, title=. 2003 , volume=

    2003

  20. [20]

    2024 , isbn =

    Ye, Mingqiao and Danelljan, Martin and Yu, Fisher and Ke, Lei , title =. 2024 , isbn =

    2024

  21. [21]

    On the uncertainty of self-supervised monocular depth estimation , author =

  22. [22]

    NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis , author=

  23. [23]

    Gaussianeditor: Swift and controllable 3d editing with gaussian splatting , author=

  24. [24]

    Reference-guided Controllable Inpainting of Neural Radiance Fields , author=

  25. [25]

    MVIP-NeRF: Multi-view 3D Inpainting on NeRF Scenes via Diffusion Prior , author=

  26. [26]

    2D Gaussian Splatting for Geometrically Accurate Radiance Fields , author=

  27. [27]

    InpaintFusion: Incremental RGB-D Inpainting for 3D Scenes , volume=

    Mori, Shohei and Erat, Okan and Broll, Wolfgang and Saito, Hideo and Schmalstieg, Dieter and Kalkofen, Denis , journal=. InpaintFusion: Incremental RGB-D Inpainting for 3D Scenes , volume=

  28. [28]

    2015 , volume=

    Ružić, Tijana and Pižurica, Aleksandra , journal=TIP, title=. 2015 , volume=

    2015

  29. [29]

    Patch-Based Image Inpainting via Two-Stage Low Rank Approximation , year=

    Guo, Qiang and Gao, Shanshan and Zhang, Xiaofeng and Yin, Yilong and Zhang, Caiming , journal=. Patch-Based Image Inpainting via Two-Stage Low Rank Approximation , year=

  30. [30]

    Multimedia Tools Appl

    Lu, Hongyang and Liu, Qiegen and Zhang, Minghui and Wang, Yuhao and Deng, Xiaohua , title =. Multimedia Tools Appl. , month = mar, pages =. 2018 , publisher =

    2018

  31. [31]

    2014 , volume=

    Lu, Si and Ren, Xiaofeng and Liu, Feng , booktitle=CVPR, title=. 2014 , volume=

    2014

  32. [32]

    Atapour Abarghouei, Amir and Breckon, Toby , year =

  33. [33]

    and Firman, Michael and Vicente, Sara , title =

    Weder, Silvan and Garcia-Hernando, Guillermo and Monszpart, \'Aron and Pollefeys, Marc and Brostow, Gabriel J. and Firman, Michael and Vicente, Sara , title =. 2023 , pages =

    2023

  34. [34]

    arXiv preprint arXiv:2109.07161 , year=

    Resolution-robust Large Mask Inpainting with Fourier Convolutions , author=. arXiv preprint arXiv:2109.07161 , year=

    Pith/arXiv arXiv

  35. [35]

    Towards An End-to-End Framework for Flow-Guided Video Inpainting , author=

  36. [36]

    DVI: Depth Guided Video Inpainting for Autonomous Driving , author=

  37. [37]

    2024 , volume=

    Corneanu, Ciprian and Gadde, Raghudeep and Martinez, Aleix M , booktitle=WACV, title=. 2024 , volume=

    2024

  38. [38]

    2025 , eprint=

    Efficient Zero-Shot Inpainting with Decoupled Diffusion Guidance , author=. 2025 , eprint=

    2025

  39. [39]

    2024 , eprint=

    ObjectDrop: Bootstrapping Counterfactuals for Photorealistic Object Removal and Insertion , author=. 2024 , eprint=

    2024

  40. [40]

    2021 , eprint=

    High-Resolution Image Synthesis with Latent Diffusion Models , author=. 2021 , eprint=

    2021

  41. [41]

    Gaussian Grouping: Segment and Edit Anything in 3D Scenes , author=

  42. [42]

    arXiv preprint arXiv:2212.03245 , year=

    Differentiable Point-Based Radiance Fields for Efficient View Synthesis , author=. arXiv preprint arXiv:2212.03245 , year=

    Pith/arXiv arXiv

  43. [43]

    Srinivasan and Peter Hedman and Ricardo Martin-Brualla and Jonathan T

    Ben Mildenhall and Dor Verbin and Pratul P. Srinivasan and Peter Hedman and Ricardo Martin-Brualla and Jonathan T. Barron , year=

  44. [44]

    2025 , eprint =

    Gemma 3 Technical Report , author =. 2025 , eprint =

    2025

  45. [45]

    Nano Banana Pro: Gemini 3 Pro Image Generation , year =

  46. [46]

    2024 , eprint=

    From Noise to Nuance: Advances in Deep Generative Image Models , author=. 2024 , eprint=

    2024

  47. [47]

    2018 , eprint=

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric , author=. 2018 , eprint=

    2018

  48. [48]

    Heusel, Martin and Ramsauer, Hubert and Unterthiner, Thomas and Nessler, Bernhard and Hochreiter, Sepp , title =. Proc. NeurIPS , pages =. 2017 , publisher =

    2017

  49. [49]

    Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields , author=

  50. [50]

    Poisson image editing , year =

    P\'. Poisson image editing , year =

  51. [51]

    2024 , pages=

    Yang, Lihe and Kang, Bingyi and Huang, Zilong and Xu, Xiaogang and Feng, Jiashi and Zhao, Hengshuang , booktitle=CVPR, title=. 2024 , pages=

    2024

  52. [52]

    Mori, Shohei and Schmalstieg, Dieter and Kalkofen, Denis , journal=TVCG, title=

  53. [53]

    Scaffold-gs: Structured 3d gaussians for view-adaptive rendering , author=

  54. [54]

    2024 , eprint=

    Text-to-3D using Gaussian Splatting , author=. 2024 , eprint=

    2024

  55. [55]

    Haque, Ayaan and Tancik, Matthew and Efros, Alexei and Holynski, Aleksander and Kanazawa, Angjoo , title =

  56. [56]

    IPSJ Transactions on Computer Vision and Applications , author =

    A survey of diminished reality:. IPSJ Transactions on Computer Vision and Applications , author =

  57. [57]

    and Tancik, Matthew and Barron, Jonathan T

    Mildenhall, Ben and Srinivasan, Pratul P. and Tancik, Matthew and Barron, Jonathan T. and Ramamoorthi, Ravi and Ng, Ren , title =. Commun. ACM , month = dec, pages =. 2021 , volume =

    2021

  58. [58]

    Look Outside the Room: Synthesizing A Consistent Long-Term 3D Scene Video from A Single Image , author=

  59. [59]

    2023 , eprint=

    Zero-1-to-3: Zero-shot One Image to 3D Object , author=. 2023 , eprint=

    2023

  60. [60]

    2023 , eprint=

    Zero123++: a Single Image to Consistent Multi-view Diffusion Base Model , author=. 2023 , eprint=

    2023

  61. [61]

    Generative Camera Dolly: Extreme Monocular Dynamic Novel View Synthesis , author=

  62. [62]

    GEN3C: 3D-Informed World-Consistent Video Generation with Precise Camera Control , author=

  63. [63]

    LeftRefill: Filling Right Canvas based on Left Reference through Generalized Text-to-Image Diffusion Model , author=

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.