Pith. sign in

REVIEW 4 major objections 4 minor 2 cited by

SplatFill claims a depth-guided Gaussian-splatting inpainting pipeline that produces sharper, more view-consistent scene fills than prior NeRF- and 3DGS-based methods while training 24.5% faster.

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 →

T0 review · deepseek-v4-flash

2026-08-04 21:40 UTC pith:AAZCLMD2

load-bearing objection A well-designed 3DGS inpainting method with strong internal ablations, but the SOTA claim outruns the evidence: masked-region tie with the one retrained baseline, missing direct comparison to the other two 3DGS methods. the 4 major comments →

arxiv 2509.07809 v1 pith:AAZCLMD2 submitted 2025-09-09 cs.CV

SplatFill: 3D Scene Inpainting via Depth-Guided Gaussian Splatting

classification cs.CV
keywords 3D scene inpaintingGaussian Splattingdepth-guided inpaintingmulti-view consistencyobject removalselective guided inpaintingmonocular depth supervisionperceptual quality
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.

SplatFill sets out to show that 3D scene inpainting does not require a slow NeRF optimization or a blanket multi-view 2D inpainting pass: starting from one diffusion-inpainted reference, its depth-guided Gaussian-splatting optimizer can place new Gaussians correctly and then repair only the views where rendered depth disagrees most with monocular depth. The paper claims this beats existing NeRF- and 3DGS-based inpainting methods on perceptual fidelity (full-image LPIPS 0.25, FID 29.76 on the SPIn-NeRF benchmark) while cutting training time by about 24.5% relative to the fastest prior Gaussian-splatting baseline. The contribution matters because object removal and scene editing in captured real-world scenes need sharp, geometrically coherent fill that is also quick enough to iterate on. The method's three levers are a scale-shift-free depth clustering loss, a crop-focused depth loss, and an iterative selective-guided-inpainting loop, with an object-contrastive loss that also attaches usable segmentation features to Gaussians.

Core claim

The paper's central claim is that monocular depth, used the right way, can replace explicit depth inpainting and learned scale-shift alignment for Gaussian-splatting inpainting. Its soft depth clustering loss bins monocular depth and penalizes the spread of rendered depth around each bin's mean, so rendered and estimated depths are compared on the same scale; the crop-focused variant concentrates this supervision on masked areas. On top of that, the selective guided inpainting loop renders all views, picks the view with the largest accumulated rendered-versus-monocular depth error in masked regions, thresholds the error's gradient to localize inconsistent patches, re-inpaints only those patc

What carries the argument

Three mechanisms carry the argument. The Soft Depth Clustering Loss compares rendered depth against the mean rendered depth inside each monocular-depth bin, making supervision scale- and shift-invariant by construction; the Crop-Focused Depth Loss applies the same binning to randomly expanded crops around the masked area with Gaussian weighting, concentrating supervision where geometry is hardest. Selective Guided Inpainting is the refinement engine: it accumulates masked-region absolute error between rendered and monocular depth per view, chooses the worst view, thresholds the gradient magnitude of that error, dilates the result, and re-inpaints only that localized patch, then adds the view

Load-bearing premise

The pipeline starts by randomly picking one reference view for the initial 2D inpainting; if that view is a poor seed, the depth and segmentation losses cannot recover, and the method has no fallback re-selection.

What would settle it

A direct test: run SplatFill on each SPIn-NeRF scene with multiple different random choices of reference view and record the spread of masked-region LPIPS; if one bad random reference systematically produces visible artifacts that SGI does not repair, the random-selection premise fails. A separate test: capture a scene with a moving shadow or reflection inside the inpainted region; since the paper states it does not explicitly handle illumination changes, this should expose inconsistencies and mark the boundary of the claim.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Object removal from a captured multi-view set becomes a roughly 34-minute Gaussian-splatting procedure with sharp, viewpoint-consistent inpainted regions rather than a slow NeRF optimization.
  • Because the depth loss needs no scale/shift calibration, a practitioner could swap the monocular depth estimator without re-tuning alignment, and imported depth priors transfer directly.
  • The selective refinement loop means only the most inconsistent view is re-inpainted each round, so the number of expensive diffusion calls scales with residual error, not with the number of views.
  • The per-Gaussian segmentation features provide a usable by-product: selecting or masking one object in the 3D representation for editing comes from the same learned field used during inpainting.

Where Pith is reading between the lines

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

  • If SplatFill is sound, the random-reference limitation is the first place to improve: a cheap view-selection heuristic (for example, maximize visible unmasked area or minimize monocular-depth uncertainty) could shrink result variance more than any loss change.
  • The SGI view selector is driven purely by depth error, so it may miss appearance-only inconsistencies; combining it with rendered-feature or perceptual error could make the refinement loop catch texture artifacts too.
  • The depth-clustering loss should transfer to other 3D representations such as meshes or NeRFs, since it needs only rendered depth and a monocular prior, though the paper only demonstrates it for 3DGS.
  • The object-insertion results in the supplementary material suggest the same pipeline can serve as a general scene-editing tool, not just object removal; per-object inpainting remains a stated future step.

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 / 4 minor

Summary. The paper proposes SplatFill, a 3D Gaussian Splatting inpainting method for object removal in multi-view scenes. The pipeline selects a single reference view, inpaints it with a diffusion-based 2D model, and then optimizes 3D Gaussians under photometric supervision plus two proposed depth losses (Soft Depth Clustering Loss and Crop-Focused Depth Loss) and an Object-Aware Contrastive Loss. A Selective Guided Inpainting (SGI) stage iteratively re-inpaints only the views/regions with the largest rendered-vs-monocular depth inconsistencies. Experiments on the SPIn-NeRF dataset report full-image improvements over GScream in SSIM, LPIPS, and FID, identical masked-region scores to GScream, and a 24.5% training-time reduction. The paper claims state-of-the-art perceptual quality for both NeRF- and 3DGS-based inpainting.

Significance. If the claims are substantiated, SplatFill would be a useful engineering contribution: it combines explicit depth-guided supervision that avoids scale-shift alignment, an object-aware segmentation field, and a selective refinement loop, and it demonstrates faster training than GScream. The evaluation uses the external SPIn-NeRF benchmark with held-out test views, and the supervision signals come from off-the-shelf depth and segmentation models rather than from the test set, so circularity is not a concern. However, the central state-of-the-art claim is currently under-supported: the two closest 3DGS baselines (InFusion, RefFusion) are not quantitatively evaluated, the masked-region numbers are identical to GScream, and the full-region differences are small and reported without variance or significance tests. The paper therefore needs additional experiments and/or more cautious claims before the central assertion can be accepted.

major comments (4)
  1. [Sec. 2.3, Sec. 4.1, Tables 1-2] InFusion [16] and RefFusion [21] are described in Sec. 2.3 as the closest 3DGS-based inpainting methods, and Sec. 4.1 states 'We compare our results to [16,21,29]'. However, Tables 1 and 2 report quantitative results only for GScream, SPIn-NeRF, and OR-NeRF. The abstract's claim that SplatFill 'surpasses existing NeRF-based and 3DGS-based inpainting methods' is therefore unverified for the two most relevant 3DGS baselines. The authors should add quantitative comparisons to InFusion and RefFusion under the same protocol, or explicitly restrict the SOTA claim to the methods actually compared.
  2. [Tables 1-2, Sec. 4.2] The numerical support for 'state-of-the-art perceptual quality' is weak. In Table 2 the masked-region scores are identical to GScream (PSNR 15.67, SSIM 0.21, LPIPS 0.54). In Table 1 the full-region improvements are PSNR 20.46 vs 20.45, SSIM 0.63 vs 0.58, and LPIPS 0.25 vs 0.28, with no error bars, per-scene breakdown, or significance tests. The FID improvement (29.76 vs 36.72) is a full-image metric; masked-region FID is not reported, so it does not establish perceptual quality inside the inpainted area. The authors should provide repeated-seed statistics and masked-region FID, or soften the 'surpasses' language to 'comparable'.
  3. [Sec. 5, Tables 1-2] The Limitations section states that the initial reference view is selected randomly and that its quality 'has a strong influence' on the result. Since all reported numbers appear to come from a single run/seed, the claimed improvements may be an artifact of a favorable random choice. The authors should evaluate robustness across multiple reference-view selections, or implement the more informed heuristic they mention. Without this, the quantitative claims are not stable and the SOTA comparison is not reliable.
  4. [Sec. 4.2, Tables 1-2] The SPIn-NeRF and OR-NeRF numbers are borrowed from GScream's paper, with an asterisk, and the text concedes these 'may not fully reflect differences in conditions'. The indirect inference that matching GScream implies beating these baselines is not rigorous. The authors should either retrain SPIn-NeRF and OR-NeRF under the same evaluation protocol, or present the borrowed values only as indicative and avoid making comparative claims based on them.
minor comments (4)
  1. [Sec. 4.1] The text says 'four baseline approaches' but then lists only GScream, SPIn-NeRF, and OR-NeRF. Either include a fourth baseline or correct the count.
  2. [Sec. 2.2] Typo: 'pionnering' should be 'pioneering'.
  3. [Eq. (2), Sec. 3.2] The per-bin loss minimizes absolute deviation from the bin's mean rendered depth, effectively encouraging constant depth within each bin. The authors should clarify why flattening within-bin depth variation is desirable, and report sensitivity to the number of depth bins, which is a free parameter.
  4. [Supplementary, Fig. 5] The caption says 'spanning two columns' while the text describes 'three columns in each row'. Please make the description consistent.

Circularity Check

0 steps flagged

No significant circularity; central claim is evaluated on external held-out benchmark and does not reduce to its own inputs.

full rationale

We inspected the derivation chain. SplatFill's claimed results are measured on the SPIn-NeRF held-out test views, with supervision from off-the-shelf monocular depth (Depth Anything), SAM2 masks, and a pre-trained 2D diffusion inpainter. The depth losses (SDCL and CFDL) and the SGI refinement compare rendered depth to monocular depth; they are training losses, not predictions derived from test targets. No fitted constant is renamed as a prediction, and no loss or equation is defined in terms of the perceptual metrics it claims to improve. The only self-citation is the object-aware contrastive loss borrowed from the authors' prior ICPR paper [4], used as a 16-dimensional per-Gaussian feature and contrastive clustering loss. This is a component of the method, is not isolated by ablation, and the state-of-the-art claim does not logically reduce to that citation; rather, it is supported by external benchmark numbers. The limitation about randomly selecting the reference view is acknowledged and is a robustness/correctness issue, not a circularity. No circular step could be quoted with the required specificity.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The method depends on three off-the-shelf priors (monocular depth, SAM2, 2D diffusion) and on several hand-tuned hyperparameters that are not fully disclosed. No new physical or conceptual entities are introduced.

free parameters (7)
  • kappa (CFDL weight) = 25
    Balances global SDCL and localized CFDL in Eq. (4); set empirically without sensitivity analysis.
  • Baseline training steps = 8000
    Number of optimization steps before SGI; fixed empirically.
  • SDCL computation frequency = every 59 iterations
    Hyperparameter controlling how often the depth clustering loss is applied; see supplementary.
  • CFDL computation frequency = every 9 iterations
    Hyperparameter for the crop-focused loss frequency; see supplementary.
  • Depth bin count = not reported
    Number of evenly spaced bins for SDCL/CFDL masks; not disclosed, empirically chosen.
  • SGI gradient threshold = not reported
    Threshold for binary mask from gradient magnitude in Eq. (8); not disclosed, chosen empirically.
  • Number of SGI iterations = not reported
    Loop count for selective refinement; not disclosed.
axioms (5)
  • domain assumption Monocular depth from Depth Anything provides a reliable geometric prior for 3D scene geometry.
    Used in SDCL, CFDL, and SGI (Sec. 3.2, 3.4). If depth is inaccurate, the loss and refinement target wrong geometry.
  • domain assumption SAM2 segmentation masks give object-consistent clusters for contrastive learning.
    OACL (Eq. 5) groups Gaussians using SAM2 masks; inaccurate masks degrade object alignment (Sec. 3.3).
  • domain assumption The 2D diffusion inpainting model [9] produces plausible content for masked regions.
    The initial reference and SGI updated views depend on this model; poor 2D inpainting propagates into the 3D result (Sec. 3.1, 3.4).
  • domain assumption The MCMC-based 3DGS optimizer [14] is a suitable base and the custom CUDA depth rendering is correct.
    The implementation extends MCMC to render depth and augment features; correctness is assumed (Sec. 4.1, supplementary).
  • domain assumption SPIn-NeRF test views provide valid ground truth for inpainting evaluation.
    The 40 object-free views per scene are used to measure PSNR/SSIM/LPIPS/FID (Sec. 4.1).

pith-pipeline@v1.3.0-alltime-deepseek · 13563 in / 16154 out tokens · 161695 ms · 2026-08-04T21:40:38.736260+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of SplatFill: 3D Scene Inpainting via Depth-Guided Gaussian Splatting." pith.science (2026). https://pith.science/paper/AAZCLMD2

@misc{pith2026250907809,
  author       = {Pith},
  title        = {Pith review of: SplatFill: 3D Scene Inpainting via Depth-Guided Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AAZCLMD2}},
  note         = {Machine review of arXiv:2509.07809}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

3D Gaussian Splatting (3DGS) has enabled the creation of highly realistic 3D scene representations from sets of multi-view images. However, inpainting missing regions, whether due to occlusion or scene editing, remains a challenging task, often leading to blurry details, artifacts, and inconsistent geometry. In this work, we introduce SplatFill, a novel depth-guided approach for 3DGS scene inpainting that achieves state-of-the-art perceptual quality and improved efficiency. Our method combines two key ideas: (1) joint depth-based and object-based supervision to ensure inpainted Gaussians are accurately placed in 3D space and aligned with surrounding geometry, and (2) we propose a consistency-aware refinement scheme that selectively identifies and corrects inconsistent regions without disrupting the rest of the scene. Evaluations on the SPIn-NeRF dataset demonstrate that SplatFill not only surpasses existing NeRF-based and 3DGS-based inpainting methods in visual fidelity but also reduces training time by 24.5%. Qualitative results show our method delivers sharper details, fewer artifacts, and greater coherence across challenging viewpoints.

Figures

Figures reproduced from arXiv: 2509.07809 by Alessio Del Bue, Mahtab Dahaghin, Matteo Toso, Milind G. Padalkar.

Figure 1
Figure 1. Figure 1: The 3D scene inpainting problem: Starting from a set of images with known poses and masks occluding the area to be in￾painted, we train a model to incrementally fill the masked regions with high-frequency geometric and texture details that remain con￾sistent with the unmasked scene elements. The resulting method is consistent even under strong viewpoint changes. which were either occluded by other elements… view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline. (a) We begin with a set of multi-view training images, each with its corresponding binary mask and camera pose. (b) One image is selected as the reference and inpainted using a diffusion-based 2D inpainting model, providing plausible content for the masked region. (c) Monocular depth estimation is performed for each image, producing depth maps that guide the 3D spatial placement of Gaussians. (d)… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative evaluation across multiple scenes (each column shows a different scene). For each method, the first row shows the full inpainted view, and the second row provides zoomed-in crops of selected regions from that view. The top block corresponds to GScream, the middle block to our proposed method, and the bottom block to the ground truth, where the target objects were physically removed from the rea… view at source ↗
Figure 4
Figure 4. Figure 4: Object insertion results. (a) Inpainted reference view [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of inpainting results spanning two columns. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison of depth supervision strategies. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of multi-inpainting strategies. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

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

  1. BEA-GS: BEyond RAdiance Supervision in 3DGS for Precise Object Extraction

    cs.CV 2026-05 unverdicted novelty 7.0

    BEA-GS achieves superior object boundary segmentation in 3D Gaussian Splatting by introducing two new losses that adjust geometry of visible and non-visible Gaussians based on semantics.

  2. Depth Estimators Are Implicit Neural Fields for 3D Scene Geometry Inpainting and Reconstruction

    cs.CV 2026-07 conditional novelty 4.0

    NDF treats a fixed-image depth estimator as an implicit field and optimizes it on observed depth at test time, improving inpainting accuracy and cross-view consistency.

Reference graph

Works this paper leans on

41 extracted references · 32 canonical work pages · cited by 2 Pith papers · 2 internal anchors

  1. [1]

    PatchMatch: A randomized correspon- dence algorithm for structural image editing.ACM Trans

    Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. PatchMatch: A randomized correspon- dence algorithm for structural image editing.ACM Trans. Graph., 28(3):24, 2009. 1, 2

  2. [2]

    Image inpainting

    Marcelo Bertalmio, Guillermo Sapiro, Vincent Caselles, and Coloma Ballester. Image inpainting. InProceedings of the 27th Annual Conference on Computer Graphics and Inter- active Techniques, page 417–424. Addison-Wesley, 2000. 1, 2

  3. [3]

    Mvinpainter: Learning multi-view consistent inpainting to bridge 2d and 3d editing.arXiv preprint arXiv:2408.08000, 2024

    Chenjie Cao, Chaohui Yu, Fan Wang, Xiangyang Xue, and Yanwei Fu. Mvinpainter: Learning multi-view consistent inpainting to bridge 2d and 3d editing.arXiv preprint arXiv:2408.08000, 2024. 2

  4. [4]

    Contrastive gaussian clustering for weakly super- vised 3d scene segmentation

    Myrna Castillo, Mahtab Dahaghin, Matteo Toso, and Alessio Del Bue. Contrastive gaussian clustering for weakly super- vised 3d scene segmentation. InPattern Recognition: 27th International Conference, ICPR 2024, Kolkata, India, De- cember 1–5, 2024, Proceedings, Part XXIII, page 114–130,

  5. [5]

    Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Mano- lis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu

    Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Mano- lis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Tech- nological Institute at ...

  6. [6]

    Mvip- nerf: Multi-view 3d inpainting on nerf scenes via diffusion prior

    Honghua Chen, Chen Change Loy, and Xingang Pan. Mvip- nerf: Multi-view 3d inpainting on nerf scenes via diffusion prior. InCVPR, 2024. 2, 3

  7. [7]

    Latentpaint: Image inpainting in latent space with diffusion models

    Ciprian Corneanu, Raghudeep Gadde, and Aleix M Mar- tinez. Latentpaint: Image inpainting in latent space with diffusion models. In2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4322–4331,

  8. [8]

    Criminisi, P

    A. Criminisi, P. Perez, and K. Toyama. Region filling and object removal by exemplar-based image inpainting.IEEE Transactions on Image Processing, 13(9):1200–1212, 2004. 1, 2

  9. [9]

    Scaling recti- fied flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- fied flow transformers for high-resolution image synthesis. InForty-first international conference on machine learning,

  10. [10]

    PixMix: A real-time ap- proach to high-quality diminished reality

    Jan Herling and Wolfgang Broll. PixMix: A real-time ap- proach to high-quality diminished reality. In2012 ieee in- ternational symposium on mixed and augmented reality (is- mar), pages 141–150. IEEE, 2012. 1, 2

  11. [11]

    GANs trained by a two time-scale update rule converge to a local nash equi- librium.Advances in neural information processing systems, 30, 2017

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equi- librium.Advances in neural information processing systems, 30, 2017. 6

  12. [12]

    Globally and locally consistent image completion.ACM Transactions on Graphics (ToG), 36(4):1–14, 2017

    Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Globally and locally consistent image completion.ACM Transactions on Graphics (ToG), 36(4):1–14, 2017. 2

  13. [13]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4), 2023. 1

  14. [14]

    3d gaussian splat- ting as markov chain monte carlo

    Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Wei- wei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splat- ting as markov chain monte carlo. InAdvances in Neural Information Processing Systems (NeurIPS), 2024. Spotlight Presentation. 6, 2

  15. [15]

    Taming latent diffusion model for neural radiance field inpainting

    Chieh Hubert Lin, Changil Kim, Jia-Bin Huang, Qinbo Li, Chih-Yao Ma, Johannes Kopf, Ming-Hsuan Yang, and Hung- Yu Tseng. Taming latent diffusion model for neural radiance field inpainting. InEuropean Conference on Computer Vi- sion (ECCV), 2024. 2, 3

  16. [16]

    InFusion: Inpainting 3d gaussians via learn- ing depth completion from diffusion prior.arXiv preprint arXiv:2404.11613, 2024

    Zhiheng Liu, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Jie Xiao, Kai Zhu, Nan Xue, Yu Liu, Yujun Shen, and Yang Cao. InFusion: Inpainting 3d gaussians via learn- ing depth completion from diffusion prior.arXiv preprint arXiv:2404.11613, 2024. 2, 3, 4

  17. [17]

    Repaint: Inpaint- ing using denoising diffusion probabilistic models

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpaint- ing using denoising diffusion probabilistic models. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11451–11461, 2022. 1, 2

  18. [18]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. InProceedings of the European Conference on Com- puter Vision (ECCV), 2020. 1

  19. [19]

    Reference-guided control- lable inpainting of neural radiance fields

    Ashkan Mirzaei, Tristan Aumentado-Armstrong, Marcus A Brubaker, Jonathan Kelly, Alex Levinshtein, Konstantinos G Derpanis, and Igor Gilitschenski. Reference-guided control- lable inpainting of neural radiance fields. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 17815–17825, 2023. 2, 3

  20. [20]

    SPIn-NeRF: Multiview segmentation and perceptual inpainting with neural radiance fields

    Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstanti- nos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein. SPIn-NeRF: Multiview segmentation and perceptual inpainting with neural radiance fields. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 20669–20679,

  21. [21]

    RefFusion: Reference adapted diffusion models for 3d scene inpainting, 2024

    Ashkan Mirzaei, Riccardo De Lutio, Seung Wook Kim, David Acuna, Jonathan Kelly, Sanja Fidler, Igor Gilitschen- ski, and Zan Gojcic. RefFusion: Reference adapted diffusion models for 3d scene inpainting, 2024. 2, 4

  22. [22]

    Context encoders: Feature learning by inpainting

    Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. InProceedings of the IEEE con- 9 ference on computer vision and pattern recognition, pages 2536–2544, 2016. 2

  23. [23]

    Inpaint3D: 3D Scene Content Generation using 2D Inpainting Diffusion

    Kira Prabhu, Jane Wu, Lynn Tsai, Peter Hedman, Dan B Goldman, Ben Poole, and Michael Broxton. Inpaint3D: 3d scene content generation using 2d inpainting diffusion.arXiv preprint arXiv:2312.03869, 2023. 2, 3

  24. [24]

    SAM 2: Segment anything in images and videos,

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junt- ing Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao- Yuan Wu, Ross Girshick, Piotr Doll´ar, and Christoph Feicht- enhofer. SAM 2: Segment anything in images and videos,

  25. [25]

    High-Resolution Image Synthesis with Latent Diffusion Models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models . In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674–10685, 2022. 1, 2

  26. [26]

    Geometry-Aware Diffusion Models for Multiview Scene Inpainting

    Ahmad Salimi, Tristan Aumentado-Armstrong, Marcus A Brubaker, and Konstantinos G Derpanis. Geometry-aware diffusion models for multiview scene inpainting.arXiv preprint arXiv:2502.13335, 2025. 2

  27. [27]

    Resolution-robust large mask inpainting with fourier convolutions

    Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3172– 3182, 2022. 2

  28. [28]

    InNeRF360: Text-Guided 3D-Consistent Object Inpainting on 360-degree Neural Radiance Fields

    Dongqing Wang, Tong Zhang, Alaa Abboud, and Sabine S¨usstrunk. InNeRF360: Text-Guided 3D-Consistent Object Inpainting on 360-degree Neural Radiance Fields. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3

  29. [29]

    Learning 3D geometry and feature consistent gaussian splat- ting for object removal

    Yuxin Wang, Qianyi Wu, Guofeng Zhang, and Dan Xu. Learning 3D geometry and feature consistent gaussian splat- ting for object removal. InEuropean Conference on Com- puter Vision, pages 1–17. Springer, 2025. 2, 3, 4, 6

  30. [30]

    Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004. 6

  31. [31]

    Nerfiller: Completing scenes via generative 3d inpainting

    Ethan Weber, Aleksander Holynski, Varun Jampani, Saurabh Saxena, Noah Snavely, Abhishek Kar, and Angjoo Kanazawa. Nerfiller: Completing scenes via generative 3d inpainting. InCVPR, 2024. 2, 3

  32. [32]

    Smartbrush: Text and shape guided object inpaint- ing with diffusion model

    Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpaint- ing with diffusion model. In2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22428–22437, 2023. 1, 2

  33. [33]

    Depth anything: Unleashing the power of large-scale unlabeled data

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371–10381, 2024. 4

  34. [34]

    OR- NeRF: Object removing from 3d scenes guided by multiview segmentation with neural radiance fields.arXiv preprint arXiv:2305.10503, 2023

    Youtan Yin, Zhoujie Fu, Fan Yang, and Guosheng Lin. OR- NeRF: Object removing from 3d scenes guided by multiview segmentation with neural radiance fields.arXiv preprint arXiv:2305.10503, 2023. 2, 3, 6

  35. [35]

    Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S. Huang. Generative image inpainting with contex- tual attention. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5505–5514, 2018. 2

  36. [36]

    Free-form image inpainting with gated con- volution

    Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas Huang. Free-form image inpainting with gated con- volution. In2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4470–4479, 2019. 2

  37. [37]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018. 6

  38. [38]

    Large scale image comple- tion via co-modulated generative adversarial networks.arXiv preprint arXiv:2103.10428, 2021

    Shengyu Zhao, Jonathan Cui, Yilun Sheng, Yue Dong, Xiao Liang, Eric I Chang, and Yan Xu. Large scale image comple- tion via co-modulated generative adversarial networks.arXiv preprint arXiv:2103.10428, 2021. 2 10 SplatFill: 3D Scene Inpainting via Depth-Guided Gaussian Splatting Supplementary Material

  39. [39]

    flowers on the bench

    Object Insertion In addition to 3D inpainting, our framework is capable of seamlessly inserting new objects into the reconstructed 3D scene. For object insertion, we adopt a strategy analogous to our inpainting approach. Specifically, the user provides a textual prompt to Stable Diffusion (e.g., “flowers on the bench”) to generate a reference view with th...

  40. [40]

    flowers on the bench

    Additional Ablation Effect of Selective Guided InpaintingTo evaluate the impact of our Selective Guided Inpainting (SGI) strategy, we compare two methods: one that relies solely on the in- painted reference view as guidance and our full SGI ap- proach. Figure 5 is organized into two rows corresponding to these methods, with three columns in each row. In t...

  41. [41]

    Our imple- mentation builds upon a modified version of the MCMC framework [14], where we extended the CUDA rasteriza- tion functions to render depth information alongside color

    Additional Implementation Details In this section, we provide further details regarding the implementation of our SplatFill framework. Our imple- mentation builds upon a modified version of the MCMC framework [14], where we extended the CUDA rasteriza- tion functions to render depth information alongside color. This modification allows us to compute the r...