acceptodds
Under review as a conference paper at ICLR 2027

WINGS🪽: Reference-Free Gaussian Splatting Inpainting with 3D-Native Generative Priors

Abstract

Inpainting 3D Gaussian Splatting scenes, a key challenge in 3D editing, requires generating plausible content within a masked region of 3D space. Prior approaches rely on 2D diffusion models to produce one or several inpainted reference views, making them susceptible to challenges associated with multi-view inconsistency and lengthy optimization times. Departing from these approaches, we introduce a reference-free Gaussian splatting inpainting method operating natively in 3D. Our method leverages the embedding space of a large, pre-trained 3D prior, combined with a structure completion network to feed a generative prior which reconstructs the missing region's geometry and appearance. Performing content generation entirely in 3D, it avoids the need to reconcile inconsistencies of multiple inpainted reference images, and is faster than related 2D-based methods. We demonstrate the effectiveness of our method qualitatively and quantitatively, through extensive experiments and a user study. To the best of our knowledge, this work is the first Gaussian splatting inpainting method to operate in the learned representation space of a 3D-native generative prior without relying on inpainted reference views.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
gaussian splatting, inpainting, structured latents, rectified flow

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_NCFvOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

WINGS addresses object removal and hole completion in 3D Gaussian Splatting scenes. Instead of generating one or more 2D inpainted reference images and fitting the 3D scene to them, it uses the pretrained TRELLIS generative model as a 3D-native prior.

Critical strengths
  1. The 3D-native fomulation is genuenly useful. Generating the missing content directly in a learned 3D latent representation is well positioned to solve the inconsistent completions across views of 2D-based inpainting methods. The method is technically coherent and the human study suggests that this architectural change produces perceptually meaningful improvements.

  2. The 3D-native formulation that generates the missing Gaussians conditioned directly on the retained scene context in TRELLIS’s latent space is in my opnion very clever. This avoids relying on independently inpainted 2D views and naturally promotes multi-view consistency.

  3. Strong experimental breadth and component validation. The paper evaluates three benchmarks, four methods, automated and human metrics, runtime, failure cases, and detailed ablations. The attention masking, structure completion, axis alignment, and decoder adaptation each receive empirical support.

Critical weaknesses
  1. The strongest generalization claim is not convincingly tested. The structure completion model is trained on samples filtered to favor geometry inferable from nearby context. Although the external benchmarks are not subject to this filtering, performance is not analyzed by geometric complexity, leaving generalization to irregular or semantically ambiguous holes unclear.

  2. The gains over Inpaint360GS are modest on most automated metrics. The stronger user-study result is promising, but participant details and statistical uncertainty are limited. The evaluation also uses clean predefined masks, so robustness to realistic segmentation errors remains unclear.

  3. The paper's novelty is somewhat limited. Its contribution is primarily an engineering adaptation of existing 3D latent inpainting techniques to photorealistic 3DGS scenes, rather than a fundamentally new formulation.

Questions for authors
  1. Since TRELLIS already supports RePaint-style regional editing, and VoxHammer and InpaintSLat also edit structured 3D latents, could you clarify WINGS’s key technical novelty and provide a direct comparison under the same scenes and masks?

  2. How does WINGS perform on irregular structures that do not satisfy the training-data curation criteria, or when the input 3D mask is imperfect? A breakdown by geometric complexity or mask quality would help establish generalization.

  3. Can WINGS support controllable editing tasks (i.e. Xiao et al., “Localized Gaussian Splatting Editing with Contextual Awareness,” WACV 2025.) such as changing an object’s color or inserting a specified object? If not, is the current method limited to unconditional completion after object removal?

3 comments
AuthorsOpenReview ↗

Part 1/2 - Response to Reviewer NCFv

We thank Reviewer NCFv for their constructive remarks and feedback.

Q1: Since TRELLIS already supports RePaint-style regional editing, and VoxHammer and InpaintSLat also edit structured 3D latents, could you clarify WINGS’s key technical novelty and provide a direct comparison under the same scenes and masks?

Our method performs inpainting of 3D Gaussian splatting scenes by operating in the latent space of the TRELLIS generative prior. It differs from existing 3D latent inpainting techniques such as TRELLIS's RePaint [3, 9], VoxHammer [10] and InpaintSLat [11] in several aspects:

  • We propose scene-specific decoder adaptation to address the domain gap between the object-centric training distribution of the generative prior and photorealistic scene-level assets, while the aforementioned latent inpainting techniques operate in latent space without proposing such a mechanism, yielding poor perceptual consistency between existing context and generated Gaussians (see Table (j)). Our ablation study (Tables 3 and (c)) demonstrates this contribution's importance, reducing FID from 64.36 to 49.95.
  • Our learned structure completion network also outperforms TRELLIS GS\mathcal{G}_S+RePaint on sparse structure prediction (Figure 4) and on final inpainting results (Table 3). Notably, we observed that TRELLIS GSG_S+RePaint tends to generate novel geometry in the masked region, consistent with the existing context at the boundary but unnecessary for the inpainting task, specifically for object removal.
  • Our attention masking mechanism differs on principle from the one proposed by VoxHammer: we propose to restrict target-to-target attention during part of the SLAT flow integration to promote information sharing between context and target latents, whose effectiveness is demonstrated by our ablation study. On the other hand, VoxHammer caches key and value vectors during inverse noising at the structure and SLAT stages to guarantee consistency of the unedited region. This mechanism is not needed for our method, since our use of the RePaint algorithm explicitly preserves context latents at each integration step, and the original context Gaussians are preserved in the final inpainted scene.
  • InpaintSLat optimises the initial noised representation at both structure and SLAT generation stage to guarantee faithful reconstruction during flow integration, motivated by the observation that TRELLIS is sensitive to initial noise conditions. Unlike InpaintSLat, WINGS directly predicts masked region geometry using a learned structure completion network and performs reference free, context-conditional inpainting via RePaint sampling and attention masking, and does not require initial noise optimisation.
  • Finally, all three aforementioned methods mostly evaluate results on object-centric assets. Departing from these, we explicitly repurpose the TRELLIS generative prior for inpainting of photorealistic scenes, enabling usage in a new application domain.

In addition, we report in Table (j) automated perceptual metrics for the original TRELLIS RePaint-style editing and VoxHammer (we were unable to find a peer-reviewed manuscript or published code for InpaintSLat).

Table (j): Average metrics with equal-scene weighting, Ours, TRELLIS, VoxHammer

MethodLPIPS (↓)FID (↓)Local-FID (↓)KID (×10⁻³) (↓)
TRELLIS0.2434 [-0.0012 / +0.0012]70.741 [-2.777 / +2.777]71.781 [-4.925 / +4.925]28.277 [-2.913 / +2.913]
VoxHammer0.2546 [-0.0012 / +0.0012]93.427 [-7.655 / +7.655]96.733 [-7.464 / +7.464]49.401 [-8.676 / +8.676]
Ours0.2273 [-0.0002 / +0.0002]49.667 [-0.616 / +0.604]47.242 [-0.624 / +0.623]8.107 [-0.714 / +0.689]

Table (j) demonstrates that TRELLIS and VoxHammer, while effective latent inpainting methods, are not competitive with existing 3D Gaussian splatting inpainting methods for object removal in photorealistic scenes. Qualitatively, we observed a strong perceptual mismatch between existing and generated Gaussians as well as unwanted geometry in the target region. Notably, TRELLIS is outperformed by all ablated components of WINGS (see Tables 3 and (c)), supporting the effectiveness of our proposed mechanisms for axis-alignment, structure completion, conditional latent sampling and scene-specific decoder adaptation.

We also note that VoxHammer required substantial adaptation of the publicly available codebase to perform inpainting of 3DGS scene editing: as the published framework requires a 3D asset and a mask mesh as input, we adapted the code to accept the same SLAT representation encoded by WINGS and associated mask. [... to be continued in Part 2/2]

AuthorsOpenReview ↗

Part 2/2 - Response to Reviewer NCFv

In order to provide equivalent comparison with the other inpainting methods reported in our paper, we used LaMa for inpainting the reference image provided to the TRELLIS flow models GS\mathcal{G}_S and GL\mathcal{G}_L. We finally performed decoding to 3DGS and merged with the original scene in an identical way to our WINGS pipeline. However, we were not able to obtain satisfactory results on photorealistic 360° 3D Gaussian splatting scene inpainting with VoxHammer, highlighting the challenges of adapting latent inpainting techniques to scene-level environments.

Q2: How does WINGS perform on irregular structures that do not satisfy the training-data curation criteria, or when the input 3D mask is imperfect?

We thank the reviewer for their question. As an example of irregular structure, we suggest to look at results for scenes cone_red and cone_yellow from the Inpaint360 dataset. The former's masked region lies in the corner of a beveled square structure, while the latter's spans steps of a staircase. WINGS is competitive with the strongest baselines on these two representative examples, with respectively 75% and 45.5% preference over Inpaint360GS. It is nonetheless true that the majority of existing benchmark datasets for 3D Gaussian splatting inpainting mostly involve geometrically regular underlying structures and we propose to list the construction of a more challenging benchmark as a relevant future work. We encourage the reviewer to evaluate the qualitative results for those representative examples in the electronic supplemental. Additionally, our structure completion network is robust to geometrically complex holes, as reported in Figure 12.

Our evaluation protocol considers the user-provided 3D masks as a strict geometric constraint, and are assumed to contain the entirety of the unwanted content. As creative applications progressively become more 3D-native, we decided to focus on evaluating inpainting performance against masks with no important irregularities or imperfections: however, our isolated voxel filtering mechanism described in the supplementary material (Appendix B.1, Context encoding) is able to discard floating artifacts at encoding time, improving the robustness of our method to potential 3D mask imperfections.

Q3: Can WINGS support controllable editing tasks [...] such as changing an object’s color or inserting a specified object?

WINGS is designed for reference-free object removal inpainting in 3D Gaussian splatting scenes, and does not currently support controllable editing tasks, due to the design choices of our method: (1) our structure completion network is explicitly trained to predict plausible geometric content in the masked region, and (2) our conditional latent sampling procedure promotes context consistency rather than novel appearance in generated voxels.

Extending WINGS to controllable editing tasks would require modifications to both of these components, allowing prompt-controlled geometry generation by retaining the original TRELLIS GS\mathcal{G}_S flow model and performing SLAT generation based on conditioning text or images. Our proposed scene-specific decoder adaptation mechanism could be retained to promote perceptual consistency at the boundary between existing and novel content. We consider the reviewer's comment a particularly interesting future work direction, allowing 3D-native generative models to be used for editing of photorealistic 3D Gaussian splatting scenes.

[3] Xiang J, Lv Z, Xu S, Deng Y, Wang R, Zhang B, et al. Structured 3D Latents for Scalable and Versatile 3D Generation. In: 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) [Internet]. 2025 [cited 2026 Sep 26]. p. 21469–80. Available from: https://ieeexplore.ieee.org/document/11094803 doi:10.1109/CVPR52734.2025.02000

[9] Lugmayr A, Danelljan M, Romero A, Yu F, Timofte R, Van Gool L. RePaint: Inpainting using Denoising Diffusion Probabilistic Models. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) [Internet]. 2022 [cited 2026 Sep 26]. p. 11451–61. Available from: https://ieeexplore.ieee.org/document/9880056 doi:10.1109/CVPR52688.2022.01117

[10] Li L, Huang Z, Feng H, Zhuang G, Chen R, Guo C, et al. VoxHammer: Training-Free Precise and Coherent 3D Editing in Native 3D Space. In. IEEE Computer Society; 2026 [cited 2026 Sep 26]. p. 1281–92. Available from: https://www.computer.org/csdl/proceedings-article/3dv/2026/731200b281/2gPtTdhIemA doi:10.1109/3DV69130.2026.00126

[11] Chung J, Lee S, Lee KM. InpaintSLat: Inpainting Structured 3D Latents via Initial Noise Optimization [Internet]. arXiv; 2026 [cited 2026 Sep 7]. Available from: http://arxiv.org/abs/2605.00664 doi:10.48550/arXiv.2605.00664

Reviewer_NCFvOpenReview ↗

Maintain my rating of 3.

Thanks for the clear responses. I will maintain my rating of 3.

Reviewer_5YN4OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a reference-free 3D Gaussian inpainting method, named WINGS, which operates in the learned representation space of a 3D-native generative prior. Specifically, this method builds upon the TRELLIS pipeline. It first extracts context SLATs using the TRELLIS VAE encoder and employs a structure-completion network to predict the completed sparse structure. The completed sparse structure is then fed into the flow transformer with the proposed attention-masking strategy to generate inpainted SLATs. Furthermore, this method introduces a scene-conditioned decoder adaptation procedure to bridge the domain gap between TRELLIS's object-centric assets and scene-level reconstructions.

Critical strengths

This paper proposes a reference-free 3D Gaussian inpainting method that leverages a 3D-native generative prior rather than relying on 2D diffusion priors.

Critical weaknesses
  1. This paper claims that the method leverages a 3D-native generative prior for 3D scene inpainting. However, the key geometry inpainting capability seems to come from the structure-completion network (3D U-Net), which is trained as an occupancy predictor. It remains unclear whether this occupancy predictor plays a more significant role in the inpainting process than the generative prior from TRELLIS.
  2. In Figure 1, the method leverages a frozen VAE encoder to extract context SLATs. Since the TRELLIS VAE encoder is pretrained on object-centric assets, it remains unclear whether there is a domain gap between voxel features extracted from real-world 3DGS scenes and those from TRELLIS's object-centric assets.
  3. The experiments lack evaluations on forward-facing datasets, such as the SPIn-NeRF [1] dataset. In addition, the number of compared baselines is relatively limited, lacking comparisons with representative methods such as NeRFiller [2], GScream [3], AuraFusion360 [4], and InstaInpaint [5].
  4. The robustness of the method to scenes with large masked regions, such as indoor scenes demonstrated in NeRFiller[2], remains unclear.
  5. The sources of the training datasets are not clearly specified.
  6. The Introduction provides relatively few citations to relevant prior work.

[1] Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. CVPR 2023. [2] Nerfiller: Completing scenes via generative 3d inpainting. CVPR 2024. [3] Learning 3d geometry and feature consistent gaussian splatting for object removal. ECCV 2024. [4] Aurafusion360: Augmented unseen region alignment for reference-based 360 unbounded scene inpainting. CVPR 2025. [5] Instainpaint: Instant 3d-scene inpainting with masked large reconstruction model. NeurIPS 2025.

Questions for authors

Please refer to the Critical Weaknesses.

3 comments
AuthorsOpenReview ↗

Part 1/3 - Response to Reviewer 5YN4

We thank Reviewer 5YN4 for their constructive feedback and comments.

Q1: It remains unclear whether this occupancy predictor plays a more significant role in the inpainting process than the generative prior from TRELLIS.

Our proposed occupancy predictor and prior-based conditional inpainting mechanism are two distinct, but complementary, stages of our inpainting method, mirroring the two-stage framework of TRELLIS. The structure prediction network generates the geometry of the masked region and the conditional inpainting stage assigns SLAT features to the resulting inpainted structure. As such, both stages play complementary roles, and both contribute significantly to the final inpainting quality. As displayed in our ablation study (Tables 3 and (c)), removing the structure prediction network increases FID from 49.95 to 61.65; while removing attention masking increases it to 53.41. Additionally, our structure prediction network only predicts the spatial support of generated content, whose appearance is subsequently generated via sampling the TRELLIS SLAT flow model GL\mathcal{G}_L: therefore, using the generative prior is indispensable to performing the inpainting task.

Q2: It remains unclear whether there is a domain gap between voxel features extracted from real-world 3DGS scenes and those from TRELLIS's object-centric assets.

We thank the reviewer for this observation. We investigate the existence of a domain gap between SLAT features obtained by encoding photorealistic 3DGS scenes and object-centric assets. To this end, we select 50 random assets from the Toys4k dataset [2] and encode them using the frozen TRELLIS VAE encoder, before normalizing the features according to the values recommended by TRELLIS [3], resulting in 520,250 SLAT tokens. We perform the same process for 50 randomly cropped subscenes sampled from the COR-NeRF, Inpaint360 and 360-USID datasets, yielding 774,079 SLAT tokens.

Evaluating distributional distances between object-centric and scene-level SLAT features reveals a consistent difference between the object-centric and scene-level features, as reported in Table (f):

Table (f): Distributional distance metrics between object-centric and scene-level SLAT features

Wasserstein-2 distanceSquared MMD distanceKL-divergence
Object / Scene1.72380.08352.170
Object / Object0.5620.00760.256

The MMD distance is evaluated with a RBF kernel using the median heuristic, and KL-divergence is computed from the scene-level distribution to the object-centric distribution (Object / Scene). The same metrics computed on two disjoint object-centric asset collections from Toys4k (Object / Object) exhibit much lower distributional distances. Table (f) therefore highlights a systematic distributional gap between object-centric and scene-level SLAT features, but this mismatch can partially be attributed to differences in the encoded semantic content. To further the analysis, we evaluate the domain mismatch at the level of the SLAT flow model GLG_L. We perform this by deterministically noising SLAT features and computing the mean squared error of the SLAT flow model's predicted velocity against the true velocity obtained from the inverse flow trajectory. Table (g) describes the flow velocity MSE at different noise levels.

Table (g): Flow velocity MSE, null conditioning, object-centric vs scene-level assets.

t=0.05t=0.05t=0.25t=0.25t=0.5t=0.5t=0.75t=0.75t=0.95t=0.95
Object-centric0.5430.2230.1650.1670.267
Scene-level0.6520.3050.2370.2480.475

Table (g) shows that scene-level assets exhibit a consistently higher velocity prediction error, indicating less reliable sampling behavior on scene-level SLAT features. Our proposed attention masking strategy is motivated to address this, by constraining the target region latents to attend to the context SLATs, steering the trajectory of the flow model. The effectiveness of our strategy against the original TRELLIS SLAT flow model GL\mathcal{G}_L is displayed in our ablation study. Our revised manuscript will quantitatively discuss this SLAT feature mismatch.

AuthorsOpenReview ↗

Part 2/3 - Response to Reviewer 5YN4

Q3: The experiments lack evaluations on forward-facing datasets, [...] the number of compared baselines is relatively limited.

We acknowledge the reviewer's observation. Our method, similarly to existing related works [4, 5], focuses on unbounded 360° Gaussian splatting scene inpainting, rather than forward-facing scene inpainting. We will clarify the scope of our work in the revised manuscript by specifically mentioning this in the Introduction.

We also welcome the reviewer's suggestion of incorporating more baselines. While NeRFiller is specialized to radiance field inpainting, InstaInpaint and GScream are inpainting methods specifically devised for forward-facing inpainting. InstaInpaint's authors have acknowledged that their method does not apply to 360° scene inpainting [6], and GScream has been established as "fundamentally limited in 360° inpainting due to [its] lack of multi-view awareness" [5]. However, we welcome the reviewer's suggestion of including AuraFusion360: it is a very relevant baseline, and including it substantially increases the experimental breadth of our paper. We report in Table (h) the results of AuraFusion360 on the COR-NeRF, Inpaint360 and 360-USID datasets, compared against WINGS.

Table (h): Perceptual metrics for AuraFusion360 and WINGS

MethodLPIPS (↓)FID (↓)Local-FID (↓)KID (×10⁻³) (↓)
AuraFusion3600.24560.44563.14713.06
Ours0.22749.66747.2428.107

AuraFusion ranks third on all reported scene-averaged metrics, behind Ours and Inpaint360GS, and in front of Gaussian Grouping and InFusion. We thank the reviewer for proposing to include this baseline: the revised manuscript will include full results for AuraFusion360, including multi-seed runs with confidence intervals.

Q4: The robustness of the method to scenes with large masked regions [...] remains unclear.

We evaluate the robustness of our method by evaluating metric performance and human preference against the strongest baseline, Inpaint360GS, with respect to masked region size across all three benchmarked datasets. The 29 benchmarked scenes are split in two categories characterising masked region extent: (1) small, 17 scenes with masked regions spanning from 0.1% to 1% of the entire extent of the unbounded 360° scene in the XY plane, and (2) large, 12 scenes with masks spanning from 1% to 4.2% of the entire scene extent. The evaluated scene with the largest masked region is the car scene from the Inpaint360 dataset. Comparative performance of WINGS and the strongest baseline, Inpaint360GS, is reported in Table (i):

Table (i): WINGS vs Inpaint360GS as a function of masked region size

Masked region sizeMethodLPIPS ↓FID ↓Local-FID ↓KID × 10³ ↓Scene-mean human preference (%) ↑
Small: 0–1%WINGS0.227 [-0.00012, +0.00013]52.58 [-0.76, +0.87]52.13 [-0.78, +0.85]7.524 [-0.942, +0.961]67.7
Inpaint360GS0.226 [-0.00003, +0.00003]50.58 [-0.16, +0.18]52.46 [-0.34, +0.31]5.151 [-0.082, +0.085]32.3
Large: >1–4.215%WINGS0.227 [-0.00040, +0.00038]45.54 [-0.63, +0.67]40.31 [-0.52, +0.51]8.932 [-0.599, +0.673]74.9
Inpaint360GS0.237 [-0.00005, +0.00006]49.89 [-0.08, +0.08]50.27 [-0.09, +0.10]8.660 [-0.046, +0.043]25.1

We see that WINGS and Inpaint360GS obtain comparable perceptual metrics on the subset of 17 scenes with small masked regions. Notably, WINGS obtains an average of 67.7% human preference against Inpaint360GS and a lower Local-FID, despite higher LPIPS, FID and KID metrics.

On the subset of 12 scenes with larger masked regions, WINGS clearly outperforms Inpaint360GS on LPIPS, FID, Local-FID and human preference, with an average over scenes of 74.9%. These results suggest that our method is robust to large masks in 3D Gaussian splatting scenes. While performance on extremely large masks such as those described in NeRFiller [9] has not yet been evaluated, further characterisation of our method's capabilities in challenging environments will be the object of future work; specifically, TRELLIS.1's fixed 64364^3 voxel grid, with 32 decoded Gaussians per voxel, may limit reconstruction fidelity when encoding large masked regions.

AuthorsOpenReview ↗

Part 3/3 - Response to Reviewer 5YN4

Q5: The sources of the training datasets are not clearly specified.

The internal training dataset used for our structure prediction network consists in 900 internal scene-level 3D Gaussian splatting scenes. Appendix C.1 details our dataset curation process. Although we cannot share our internal training dataset, the code we will publicly release upon acceptance will include the exact computational procedure required to regenerate our training dataset from publicly available data, such as SuperSplat scenes licensed under CC BY 4.0.

Q6: The Introduction provides relatively few citations to relevant prior work.

We thank the reviewer for this comment. We will add citations to relevant existing work in the Introduction section, specifically to TRELLIS [3] and to relevant 3D Gaussian splatting and radiance field inpainting and editing methods [4, 5, 7, 8, 9].

[2] Stojanov S, Thai A, Rehg JM. Using Shape to Categorize: Low-Shot Learning with an Explicit Shape Bias. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) [Internet]. 2021 [cited 2026 Nov 17]. p. 1798–808. Available from: https://ieeexplore.ieee.org/document/9577863 doi:10.1109/CVPR46437.2021.00184

[3] Xiang J, Lv Z, Xu S, Deng Y, Wang R, Zhang B, et al. Structured 3D Latents for Scalable and Versatile 3D Generation. In: 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) [Internet]. 2025 [cited 2026 Sep 26]. p. 21469–80. Available from: https://ieeexplore.ieee.org/document/11094803 doi:10.1109/CVPR52734.2025.02000

[4] Wu CH, Chen YJ, Chen YH, Lee JY, Ke BH, Mu CWT, et al. AuraFusion360: Augmented Unseen Region Alignment for Reference-based 360° Unbounded Scene Inpainting. In: 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) [Internet]. 2025 [cited 2026 Sep 26]. p. 16366–76. Available from: https://ieeexplore.ieee.org/document/11092624 doi:10.1109/CVPR52734.2025.01526

[5] Wang S, Zhang S, Millerdurai C, Westermann R, Stricker D, Pagani A. Inpaint360GS: Efficient Object-Aware 3D Inpainting via Gaussian Splatting for 360° Scenes. In: 2026 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) [Internet]. IEEE; 2026 [cited 2026 Sep 26]. p. 117–27. Available from: https://ieeexplore.ieee.org/abstract/document/11492045/

[6] InstaInpaint authors, Official comment, https://openreview.net/forum?id=8EhH3A9hoP&noteId=92xgSRXS9C

[7] Mirzaei A, Aumentado-Armstrong T, Derpanis KG, Kelly J, Brubaker MA, Gilitschenski I, et al. SPIn-NeRF: Multiview Segmentation and Perceptual Inpainting with Neural Radiance Fields. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) [Internet]. Vancouver, BC, Canada: IEEE; 2023 [cited 2026 Sep 26]. p. 20669–79. Available from: https://ieeexplore.ieee.org/document/10205488/ doi:10.1109/CVPR52729.2023.01980

[8] Chen Y, Chen Z, Zhang C, Wang F, Yang X, Wang Y, et al. GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splatting. In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) [Internet]. 2024 [cited 2026 Sep 26]. p. 21476–85. Available from: https://ieeexplore.ieee.org/document/10655963 doi:10.1109/CVPR52733.2024.02029

[9] Weber E, Holynski A, Jampani V, Saxena S, Snavely N, Kar A, et al. NeRFiller: Completing Scenes via Generative 3D Inpainting. In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) [Internet]. 2024 [cited 2026 Nov 17]. p. 20731–41. Available from: https://ieeexplore.ieee.org/abstract/document/10657147 doi:10.1109/CVPR52733.2024.01959

Reviewer_kVRAOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes WINGS, a reference-free 3DGS inpainting method that generates the masked region directly in the structured latent space of TRELLIS, without any 2D-inpainted reference views. It encodes a locally cropped, orientation-aligned context into structured latents and predicts the masked region’s sparse voxel occupancy with a 3D residual U-Net trained on 900 curated scene-level 3DGS captures. It then samples masked latents from the frozen flow model via RePaint combined with target-to-target self-attention masking, and decodes them with a per-scene LoRA-adapted decoder before merging with the context. Evaluated on 29 scenes from COR-NeRF, Inpaint360 and 360-USID, the method reports the best aggregate perceptual metrics and a strong human preference over Gaussian Grouping, InFusion and Inpaint360GS.

Critical strengths
  1. The paper is well-written and clearly motivated. To my knowledge, this paper is one of the the first work to perform scene-level 3DGS inpainting inside the latent space of a 3D-native generative prior without inpainted reference views, which sidesteps reconciling inconsistent 2D hallucinations. This is a timely and interesting direction.

  2. The method carefully addresses the domain gap between an object-centric prior and photorealistic scenes. The structure-completion U-Net clearly outperforms TRELLIS G_S + RePaint and quadric extrapolation across masked-voxel ratios (Fig. 4, Fig. 13). The decoder adaptation fixes appearance mismatch at the boundary (Table 3). The diagnosis of the flow model’s axis-alignment bias, localized to G_L rather than the VAE (Fig. 10b), is a useful empirical finding for anyone building on TRELLIS.

Critical weaknesses
  1. The quantitative gains over the strongest baseline are marginal and depend on how results are aggregated.
  • Against Inpaint360GS, LPIPS (0.235 vs. 0.236), KID (6.08 vs. 6.16) and PSNR (21.95 vs. 21.94) are effectively tied, and SSIM is lower (0.640 vs. 0.644).
  • Per dataset, the method is worse on KID on both COR-NeRF (3.63 vs. 2.82) and Inpaint360 (4.88 vs. 3.10). Table 2 averages dataset means, which overweights the 7-scene 360-USID set. Averaging over all 29 scenes instead reverses the KID ranking (about 5.58 vs. 5.27).
  • The paper’s superiority therefore rests largely on the user study. Meanwhile, the self-attention masking layers and timesteps, the context ratio of 6, and Inpaint360GS’s render size appear to have been tuned on the benchmark scenes, with no separate validation split. The ablations in Table 3 are single stochastic runs without variance.
  1. The two headline advantages over 2D-based pipelines, multi-view consistency and speed, are weakly substantiated.
  • Consistency. Multi-view consistency is the central motivation, yet no metric or experiment measures it. Every automated metric is per-view, and the user study shows a single static viewpoint per comparison, with about 8 Ours-vs-Inpaint360GS judgments per scene (231 total).
  • Speed. The claim of being “faster than related 2D-based methods” amounts to 177 s vs. 202 s for Inpaint360GS, about 12%, and the authors could not reproduce Inpaint360GS’s reported ~180 s. Decoder adaptation alone takes 103 s per scene. The faster XS variant (75 s) falls behind Inpaint360GS in FID (57.14 vs. 51.17).
Questions for authors
  1. Can you report scene-averaged metrics alongside dataset means, with confidence intervals or significance tests against Inpaint360GS? How were the masking layers and timesteps, the context ratio, and the baseline render size selected, and was a held-out validation set used? Can you report multiple seeds for Table 3?

  2. Can you evaluate multi-view consistency directly, for example with a video-based user study or a cross-view consistency metric on rendered trajectories? How do you justify the speed claim given the ~12% runtime difference and the 103 s per-scene decoder adaptation?

4 comments
AuthorsOpenReview ↗

Part 1/3 - Response to Reviewer kVRA

We sincerely thank Reviewer kVRA for their constructive feedback.

Q1.1: Can you report scene-averaged metrics alongside dataset means, with confidence intervals or significance tests against Inpaint360GS?

Following this suggestion, we changed the aggregate metrics in Table 2 to be averaged over all scenes instead of per dataset. As the reviewer highlighted, this preserves our method's advantage over Inpaint360GS on LPIPS, FID, Local-FID and human preference. Although it does turn KID in favor of Inpaint360GS, we recall that, as Table 6 shows, KID is only weakly correlated with human preference (Spearman ρ\rho = 0.13), while LPIPS, FID and Local-FID respectively have Spearman correlation coefficients ρ\rho = 0.48, 0.40 and 0.40.

We additionally conducted further experiments to further confirm our comparative advantage over the strongest baseline, Inpaint360GS. We evaluated WINGS and Inpaint360GS over 10 different random seeds and report the full scene-averaged results in Tables (a) and (b), with 95% bootstrapped confidence intervals across random seeds. WINGS outperforms Inpaint360GS on LPIPS, FID, Local-FID and median KID, and is outperformed on mean KID. WINGS exhibits reproducible comparative advantage over the strongest baseline on LPIPS and Local-FID, the two metrics most correlated with human preference. Qualitatively, the perceptual appearance of inpaintings is also robust across seeds. While we focused here on the strongest related work, Inpaint360GS, the camera-ready paper will also report multi-seed performance for the other baselines: AuraFusion360, Gaussian Grouping and InFusion.

Table (a): Average metrics with equal-scene weighting, Ours vs. Inpaint360GS

MethodLPIPS (↓)FID (↓)Local-FID (↓)KID (×10⁻³) (↓)
Inpaint360GS0.230 [-0.33e-4 / +0.36e-4]50.295 [-0.116 / +0.128]51.559 [-0.218 / +0.195]6.603 [-0.0550 / +0.0574]
Ours0.227 [-2.02e-4 / +1.99e-4]49.667 [-0.616 / +0.604]47.242 [-0.624 / +0.623]8.107 [-0.714 / +0.689]

Table (b): Median metrics with equal-scene weighting, Ours vs. Inpaint360GS

MethodLPIPS (↓)FID (↓)Local-FID (↓)KID (×10⁻³) (↓)
Inpaint360GS0.207 [-0.64e-4 / +0.61e-4]40.653 [-0.199 / +0.284]44.245 [-0.281 / +0.147]3.582 [-0.076 / +0.216]
Ours0.207 [-4.03e-4 / +4.60e-4]36.864 [-0.091 / +0.148]33.684 [-0.086 / +0.608]2.356 [-0.293 / +0.402]

Q1.2: How were the masking layers and timesteps, the context ratio, and the baseline render size selected, and was a held-out validation set used?

  • The grid search for attention masking layers, timesteps and the encoding bounding box context ratio was conducted on the COR-NeRF dataset, using randomly cropped bounding boxes unrelated to the true 3D masks used for evaluation.
  • The render size for the Inpaint360GS baseline was selected by following the original paper's implementation details, in which they use quarter-resolution on the 4K Inpaint360 dataset. On the lower-resolution COR-NeRF and 360-USID datasets, following the authors' recommendation yields poor inpainting results, due to aggressive downsampling of render size. For a fair comparison, we thus decided to fix the size of the inpainting resolution to 512 pixels, leading to higher-quality results on COR-NeRF and 360-USID. Implementation details for Inpaint360GS are given in Appendix B.2.
AuthorsOpenReview ↗

Part 2/3 - Response to Reviewer kVRA

Q1.3: Can you report multiple seeds for Table 3?

We report in Table (c) below the ablation study for our method (Table 3 in the paper), averaged over five random seeds, and will revise the manuscript with these figures. The effectiveness of each component of WINGS is demonstrated, with a reproducible metric increase for ablated parts of our method.

Table (c): Ablation study for WINGS, 5 random seeds, equal-scene weighted average

MethodLPIPS (↓)FID (↓)Local-FID (↓)KID (×10⁻³) (↓)
Ours0.227 [-0.00028 / +0.00028]49.948 [-0.646 / +0.583]48.184 [-0.532 / +0.532]7.553 [-0.621 / +0.548]
Ours (w/o attention masking)0.231 [-0.00045 / +0.00050]53.406 [-1.873 / +2.837]52.347 [-2.586 / +4.079]11.783 [-2.290 / +3.606]
Ours (w/o axis alignment)0.233 [-0.00041 / +0.00050]52.147 [-0.784 / +0.799]51.080 [-0.585 / +0.648]9.031 [-0.803 / +0.867]
Ours (w/o structure prediction model)0.233 [-0.00135 / +0.00121]61.646 [-4.450 / +5.834]61.624 [-5.037 / +6.833]19.192 [-6.206 / +7.449]
Ours (w/o decoder adaptation)0.241 [-0.00024 / +0.00023]64.359 [-0.701 / +0.814]61.126 [-0.620 / +0.780]22.917 [-1.060 / +1.152]

Q2.1: Can you evaluate multi-view consistency directly, for example with a video-based user study or a cross-view consistency metric on rendered trajectories?

In order to confirm the multi-view capabilities of our method, we evaluated multi-view consistency of our method through a video-based user study as well as through automated metrics.

For every scene and method, we rendered videos from 8 different viewpoint trajectories, each orbiting 45° around the center of the inpainted content. You may consult the videos through the following anonymized link: https://user-study.wings-iclr27.workers.dev/user_study_videos.html.

We conducted a new version of our user study based on these videos. In order to obtain more direct comparisons within the limited rebuttal time window, WINGS was always drawn, and the competing method was uniformly drawn among InFusion, Gaussian Grouping, AuraFusion360 and Inpaint360GS. Users were asked to select the method that appears the most consistent over the course of the video. A total of 17 users provided 1059 votes, with an average of 62.3 votes per participant. Users were volunteers with diverse levels of expertise in computer vision.

Table (d) reports the tie-adjusted winrate of our method against all other baselines. WINGS is largely preferred by users, thus confirming the strong multi-view consistency of our method.

Table (d): Tie-adjusted winrate, Ours vs. baselines

MethodInpaint360GSGaussian GroupingInFusionAuraFusion360
WINGS winrate (%)89.0299.4397.0891.34

In addition to this human metric, we evaluated the multi-view consistency using the MEt3R [1] metric on five sets of consecutive trajectory frames for each scene. WINGS obtains a better MEt3R metric than Inpaint360GS on 27 out of the 29 scenes in the dataset. We note that this automated metric is not as meaningful as the user study in our case, as this feature-based metric doesn't necessarily evaluate the texture-level artifacts noticed by humans in the video-based study. Table (e) reports the MEt3R score for Ours against Inpaint360GS on the three datasets, as well as an average with equal-scene weighting.

Table (e): MEt3R multi-view consistency score, Ours vs. Inpaint360GS

DatasetCOR-NeRFInpaint360360-USIDAverage (equal-scene weighting)
Inpaint360GS0.0996 [-1.11e-4, +1.13e-4]0.143 [-0.86e-4, +0.92e-4]0.136 [-1.03e-4, +1.00e-4]0.125 [-0.70e-4, +0.73e-4]
Ours0.0970 [-1.25e-4, +1.32e-4]0.139 [-1.31e-4, +1.60e-4]0.131 [-1.82e-4, +1.85e-4]0.121 [-0.89e-4, +0.99e-4]
AuthorsOpenReview ↗

Part 3/3 - Response to Reviewer kVRA

Q2.2: How do you justify the speed claim given the ~12% runtime difference and the 103 s per-scene decoder adaptation?

We agree with the reviewer that the runtime speed gains of WINGS against existing methods are modest. Our method's runtime is faster or competitive with existing baselines, while achieving higher perceptual quality and multi-view consistency. We will update the manuscript to de-emphasize our speed claim, presenting it as a secondary benefit of performing inpainting in the learned latent space of a generative prior. In addition, we acknowledge that a significant fraction of the inpainting time (approximately 58%) on the scene-specific decoder adaptation step. Reducing or amortizing this important cost will be the object of future work, either through training a powerful decoder adaptation network deployed at inference time, or through the use of future generative priors requiring less decoder adaptation iterations.

[1] Asim M, Wewer C, Wimmer T, Schiele B, Lenssen JE. MEt3R: Measuring Multi-View Consistency in Generated Images. In: 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) [Internet]. 2025 [cited 2026 Nov 17]. p. 6034–44. Available from: https://ieeexplore.ieee.org/document/11092524 doi:10.1109/CVPR52734.2025.00566

Thank you for giving us your response - all my questions have been resolved. I will maintain my score of 3.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Comment

We sincerely thank the reviewers for their constructive comments and remarks.

Overall, reviewers found our submission "well-writen and clearly motivated" [kVRA], "technically coherent" [NCFv] and "useful" [kVRA, NCFv]. Our contributions were deemed "timely and interesting" [kVRA], and are supported by "strong experimental breadth" [NCFv].

The discussion phase was constructive and allowed us to strenghten the experimental validation of our method. Our reponses to reviewer comments were judged "clear" [NCFv] and "[resolved] all my questions" [kVRA], with several new experiments and clarifications, shown here by order of impact to the revised manuscript:

  1. The multi-view consistency of our method is confirmed experimentally through both user study and an automated metric. [kVRA, Q2.1]

In order to support the claim that WINGS produces multi-view consistent inpaintings, we performed a video-based user study in which participants are asked to evaluate the consistency of WINGS against existing baselines. Humans largely preferred our method, which was judged more multi-view consistent in 91% of evaluations, and particularly in 89% of evaluations against the strongest baseline, Inpaint360GS (Table (d)). The videos shown in the user study are accessible through the following anonymized link: https://user-study.wings-iclr27.workers.dev/user_study_videos.html

The automated multi-view consistency MEt3R also confirmed the multi-view consistency of our method, beating Inpaint360GS on 27 of 29 scenes (Table (e)). These new results have been incorporated to our revised manuscript in the main text (4.2: Multi-view consistency).

  1. Our method shows a reproducible advantage over existing baselines, evaluated across 10 random seeds. [kVRA, Q1.1]

We conducted further experiments to confirm that the reported metrics improvement is reproducible across independent runs over different random seeds. Due to the limited rebuttal time, we focused on comparing WINGS to the strongest existing baseline, Inpaint360GS, and report a consistent comparative advantage. The revised manuscript will include the same study for other existing baselines (Gaussian Grouping, AuraFusion360 and InFusion). We also followed Reviewer kVRA's suggestion to report scene-averaged metrics, which preserves our advantage on LPIPS, FID, Local-FID and human preference. Inpaint360GS, in turn, obtains a better mean KID and a worse median KID.

  1. WINGS outperforms a relevant existing baseline, AuraFusion360. [5YN4, Q3]

Following Reviewer 5YN4's suggestion, we ran tests AuraFusion360 and compared its performances against WINGS. Our method outperforms AuraFusion360 on the LPIPS, FID, Local-FID, KID and MEt3R metrics, and demonstrates improved multi-view consistency, being preferred by human evaluators in 91.34% of cases (Tables (d), (e) and (h)). The revised manuscript will include full results for AuraFusion360, including multi-seed runs and qualitative comparison in image grids.

  1. The components of our method show robust individual contribution to the effectiveness of our method across random seeds [kVRA, Q1.3]

Our ablation study was repeated across five random seeds. Each of the individual components of our method (attention masking, axis-alignment, structure prediction model and scene-specific decoder adaptation) improve the performances of our method, with reproducible improvements observed on LPIPS, FID, Local-FID and KID.

  1. WINGS outperforms existing latent 3D inpainting methods on scene-level 3D Gaussian splatting inpainting tasks. [NCFv, Q1]

Following Reviewer NCFv's suggestion, we clarified the key technical novelty of WINGS against existing latent inpainting methods such as VoxHammer and InpaintSLat. We additionally compared our method with the original TRELLIS+RePaint framework and an adaptation of VoxHammer to 3DGS scene inpainting, and observed substantial perceptual improvement, both qualitatively and quantitatively, with an improvement of 21.1 and 43.7 FID points over TRELLIS+RePaint and VoxHammer, with similar improvement over LPIPS, Local-FID and KID (Table (j)).

  1. We additionally address the following questions:
  • We evaluate and quantify the domain gap between scene-level and object-centric assets at SLAT level, and describe the resulting changes in flow model dynamics, tying the observed differences to our proposed contributions. [5YN4, Q2]
  • We quantify the effectiveness of our method against the strongest existing baseline with respect to the size of the masked region, and highlight a comparative advantage on larger masked regions [5YN4, Q4].
  • We clarify the scope of our method regarding controllable editing tasks [NCFv, Q3], 3D mask quality [NCFv, Q2] and 360° inpainting [5YN4, Q3].

Further methodological detail and quantitative results are provided in the corresponding review threads.

AuthorsOpenReview ↗

General Comment - Relevant tables

We provide here some relevant tables with additional results obtained during the discussion phase. The entirety of our results can be found in the corresponding review threads.

Table (a): Average metrics with equal-scene weighting, Ours vs. Inpaint360GS

MethodLPIPS (↓)FID (↓)Local-FID (↓)KID (×10⁻³) (↓)
Inpaint360GS0.230 [-0.33e-4 / +0.36e-4]50.295 [-0.116 / +0.128]51.559 [-0.218 / +0.195]6.603 [-0.0550 / +0.0574]
Ours0.227 [-2.02e-4 / +1.99e-4]49.667 [-0.616 / +0.604]47.242 [-0.624 / +0.623]8.107 [-0.714 / +0.689]

Table (d): Tie-adjusted winrate, Ours vs. baselines

MethodInpaint360GSGaussian GroupingInFusionAuraFusion360
WINGS winrate (%)89.0299.4397.0891.34

Table (e): MEt3R multi-view consistency score, Ours vs. Inpaint360GS

DatasetCOR-NeRFInpaint360360-USIDAverage (equal-scene weighting)
Inpaint360GS0.0996 [-1.11e-4, +1.13e-4]0.143 [-0.86e-4, +0.92e-4]0.136 [-1.03e-4, +1.00e-4]0.125 [-0.70e-4, +0.73e-4]
Ours0.0970 [-1.25e-4, +1.32e-4]0.139 [-1.31e-4, +1.60e-4]0.131 [-1.82e-4, +1.85e-4]0.121 [-0.89e-4, +0.99e-4]

Table (h): Perceptual metrics for AuraFusion360 and WINGS

MethodLPIPS (↓)FID (↓)Local-FID (↓)KID (×10⁻³) (↓)
AuraFusion3600.24560.44563.14713.06
Ours0.22749.66747.2428.107

Table (j): Average metrics with equal-scene weighting, Ours, TRELLIS, VoxHammer

MethodLPIPS (↓)FID (↓)Local-FID (↓)KID (×10⁻³) (↓)
TRELLIS0.2434 [-0.0012 / +0.0012]70.741 [-2.777 / +2.777]71.781 [-4.925 / +4.925]28.277 [-2.913 / +2.913]
VoxHammer0.2546 [-0.0012 / +0.0012]93.427 [-7.655 / +7.655]96.733 [-7.464 / +7.464]49.401 [-8.676 / +8.676]
Ours0.2273 [-0.0002 / +0.0002]49.667 [-0.616 / +0.604]47.242 [-0.624 / +0.623]8.107 [-0.714 / +0.689]

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.