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

Noé Lallouet; Michael Fischer; Elie Michel

cs.CVSep 29, 2026

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

Authors: Noé Lallouet, Michael Fischer, Elie Michel

Organizations: Adobe MILES Team, LAMSADE Paris Dauphine – PSL University Paris, France · Adobe London, UK · Adobe Paris, France

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.

Figures & tables

Figure 1: Overview of Wings . Given the input 3DGS encoded as SLAT and a mask $\mathcal{M}$ , our structure completion network predicts masked-region occupancy. The SLAT s are inpainted directly in 3D, conditioned on the context latents and decoded to Gaussian splats.

Figure 2: Attention masking. The attention mask acts on the query-key dot product $QK^{\top}$ , effectively zeroing out target-to-target self-attention interactions. We note that, since the mask is not applied to the value tensor $V$ , the output of self-attention is not null for target latents.

Method	Ours	Inpaint360GS	Gaussian Grouping	InFusion	Expected rank $\downarrow$
Ours	–	75.11	84.19	93.01	1.48 $\pm$ 0.07
Inpaint360GS	24.89	–	77.04	85.04	2.13 $\pm$ 0.07
Gaussian Grouping	15.81	22.96	–	67.17	2.94 $\pm$ 0.08
InFusion	6.99	14.96	32.83	–	3.45 $\pm$ 0.08

Table 1: User study. Each cell contains the tie-adjusted win-rate (higher is better) in %, row-vs-column. See App. A.2 for details.

Method	LPIPS $\downarrow$	FID $\downarrow$	Local-FID $\downarrow$	KID $(\times 10^{3})\downarrow$	Human Pref. $(\%)\uparrow$	Runtime (s) $\downarrow$
Gauss. Grouping	$0.255$	$66.98$	$66.53$	$18.02$	$34.8$	$780.0$
InFusion	$0.250$	$88.71$	$87.74$	$40.31$	$18.6$	$212.9$
Inpaint360GS	$\underline{0.236}$	$\underline{51.17}$	$\underline{52.43}$	$\underline{6.16}$	$\underline{63.6}$	$\underline{202.1}$
Ours	$\bm{0.235}$	$\bm{49.96}$	$\bm{49.21}$	$\bm{6.08}$	$\bm{82.9}$	$\bm{176.8}$

Table 2: Averaged metrics across datasets. Bold and underline indicate best- and second-best.

Figure 3: Qualitative comparison on the COR-NeRF, Inpaint360 & 360-USID datasets (top to bottom, two rows each, best viewed zoomed in). See the supplemental video to assess multiview-consistency.

Method	LPIPS $\downarrow$	FID $\downarrow$	Local-FID $\downarrow$	KID ( $\times 10^{3}$ ) $\downarrow$
Ours	$\bm{0.235}$	$\bm{49.961}$	$\bm{49.214}$	$\bm{6.080}$
Ours (w/o attention masking)	$\underline{0.236}$	$\underline{52.806}$	$\underline{51.718}$	$\underline{9.497}$
Ours (w/o axis alignment)	$0.241$	$55.336$	$54.661$	$10.030$
Ours (w/o structure prediction model)	$0.238$	$57.260$	$59.181$	$11.893$
Ours (w/o decoder adaptation)	$0.250$	$68.776$	$63.755$	$24.867$

Table 3: Perceptual metrics for our method and its ablations.

Figure 4: Sparse structure reconstruction metrics against masked-to-context voxel ratio. Our proposed structure completion network strongly outperforms TRELLIS $\mathcal{G}_{S}$ + RePaint for all context ratios.

Figure 5: Ablation visualization. We show the context and mask regions as overlays on the input image. Each column shows the completed structure (inset, top row), the PCA of the inpainted features after the integration process (principal three components in RGB, middle row), and the final decoded result in the bottom row.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Inpainting results on the “Red bull” scene from the Inpaint360 dataset. Even though Inpaint360GS achieves better (lower) FID and KID scores, our method produces a more coherent continuation of the striped surface, and is preferred by 91.7% of human evaluators.

Figure 7: Inpainting results on the “Car” scene from the Inpaint360 dataset. Similar to Figure 6 , Inpaint360GS performs better on the pixel-aligned metrics PSNR and SSIM, even though our method produces a more plausible and notably sharper inpainting result, making it the preferred choice for 100% of human evaluators.

Figure 8: Inpainting results on the “Plant” scene from the 360-USID dataset. Here, Ours performs better on the distributional FID and KID metrics, despite being preferred by only 37.5% of human evaluators compared to Inpaint360GS.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Gaussian Grouping	$20.58$	$0.617$	$0.255$
InFusion	$20.68$	$0.628$	$0.250$
Inpaint360GS	$\underline{21.94}$	$\bm{0.644}$	$\underline{0.236}$
Ours	$\bm{21.95}$	$\underline{0.640}$	$\bm{0.235}$

Appendix

Table 4: Aggregate image-fidelity metrics, averaged across all datasets. Bold and underline indicate the best and second-best result in each column, respectively.

Stage	Ours	Ours (XS)
Context encoding	20.40	20.13
Decoder adaptation	103.02	23.58
Structure prediction	0.56	0.56
Conditional latent sampling	51.20	29.59
Decoding and merging	1.66	1.67
Total runtime (s) $\downarrow$	176.84	75.39

Appendix

Table 5: Per-stage runtime of our method.

Figure 9: Train and validation viewpoint PSNR during scene-conditioned decoder adaptation vs. number of training cameras.

Figure 10: Illustration of the generative prior’s bias towards generating axis-aligned patterns. (a) While self-attention masking neither degrades nor improves perceptual loss during inpainting for these examples, inpainting perceptually worsens between 0 and 90° rotations. (b) The flow velocity mean squared error also increases as the patterned SLAT grids are rotated, compared to an axis-aligned velocity at 0°.

Figure 11: Qualitative visualization of the effect of our axis-alignment heuristic on inpainting results. Without it, the generation of context-consistent intricate patterns, being biased by the flow velocity towards alignment, is degraded.

Figure 12: Quadric extrapolation against our learned structure prediction network for held-out test examples. Our model is able to successfully predict occupancy from complex context structure.

Figure 13: We ablate our proposed structure completion network with the TRELLIS backbone (see main text) and additionally show a comparison against a quadric-extrapolation baseline of the active ground truth voxels, which performs worse in all cases.

			Single-reference			Distributional
Dataset	Scene	Human (%)	PSNR	SSIM	LPIPS	FID	Local-FID	KID	Local-KID
COR-NeRF	aircap	75.0
	backpack	100.0
	campingbox	88.9
	dolls	66.7
	gunnysack	100.0

Appendix

Table 6: Scene-level comparison of Ours and Inpaint360GS according to automated metrics and human preferences. Metric cells display within-metric normalized differences, where positive values take a green colour and favour Ours. The Human column reports the tie-adjusted preference for Ours. Across the 29 scenes, humans prefer Ours on 21, Inpaint360GS on 7 and rate the two as equal on one.

		Reference-based			Distributional			Human
Scene	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	Local-FID $\downarrow$	KID $\downarrow$	Human Pref. $\uparrow$
Aircap	Gaussian Grouping	$20.35$	$0.599$	$0.259$	$44.04$	$41.31$	$4.26$	$48.3\pm 10.7$
	InFusion	$19.99$	$0.598$	$0.260$	$42.70$	$40.44$	$5.17$	$23.0\pm 10.7$
	Inpaint360GS	$\bm{21.41}$	$0.612$	$0.251$	$37.65$	$34.47$	$2.28$	$63.4\pm 11.2$
	Ours	$21.13$	$\bm{0.617}$	$\bm{0.247}$	$\bm{36.55}$	$\bm{32.36}$	$\bm{2.12}$	$\bm{67.8\pm 13.9}$
Backpack	Gaussian Grouping	$21.34$	$0.856$	$0.211$	$60.66$	$57.02$	$13.40$	$16.2\pm 5.4$

Appendix

Table 7: Per-scene metrics on COR-NeRF. KID is reported

\times 10^{3}

. Human Pref. is each method’s expected score against an average-skill opponent under the Davidson model, reported with a participant-bootstrap 95% confidence interval. Bold denotes the best method for each scene and metric.

		Reference-based			Distributional			Human
Scene	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	Local-FID $\downarrow$	KID $\downarrow$	Human Pref. $\uparrow$
Bag	Gaussian Grouping	$24.84$	$0.835$	$0.148$	$35.46$	$35.38$	$-0.55$	$25.9\pm 6.9$
	InFusion	$25.16$	$0.842$	$0.157$	$60.34$	$70.92$	$15.04$	$10.7\pm 4.3$
	Inpaint360GS	$\bm{27.47}$	$\bm{0.856}$	$\bm{0.133}$	$\bm{29.08}$	$\bm{25.85}$	$\bm{-3.55}$	$75.0\pm 7.8$
	Ours	$27.20$	$0.848$	$0.136$	$30.28$	$27.20$	$-2.44$	$\bm{88.8\pm 5.1}$
Car	Gaussian Grouping	$18.91$	$0.759$	$0.184$	$127.19$	$122.27$	$66.05$	$29.9\pm 5.7$

Appendix

Table 8: Per-scene metrics on Inpaint360. KID is reported

\times 10^{3}

		Reference-based			Distributional			Human
Scene	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	Local-FID $\downarrow$	KID $\downarrow$	Human Pref. $\uparrow$
Carton	Gaussian Grouping	$20.46$	$0.459$	$0.324$	$44.84$	$68.15$	$15.06$	$36.1\pm 10.5$
	InFusion	$21.01$	$0.498$	$0.309$	$87.81$	$128.00$	$59.90$	$12.1\pm 5.2$
	Inpaint360GS	$21.38$	$\bm{0.499}$	$0.302$	$37.63$	$63.23$	$11.52$	$72.0\pm 10.3$
	Ours	$\bm{21.44}$	$0.495$	$\bm{0.298}$	$\bm{34.51}$	$\bm{53.13}$	$\bm{8.04}$	$\bm{83.7\pm 8.2}$
Cone	Gaussian Grouping	$16.49$	$0.183$	$\bm{0.319}$	$55.10$	$67.60$	$2.71$	$\bm{77.4\pm 10.0}$

Appendix

Table 9: Per-scene metrics on 360-USID. KID is reported

\times 10^{3}

Figure 14: Qualitative comparison of our method and the implemented baselines on the COR-NeRF dataset. Our Wings framework provides sharp, consistent inpainting across the majority of the scenes in this dataset.

Figure 15: Qualitative comparison of our method and the implemented baselines on the Inpaint360 dataset.

Figure 16: Qualitative comparison of our method and the implemented baselines on the 360-USID dataset.

Explore similar work

Benchmarking Single-Step Inpainting Methods for Multi-Object 3D Gaussian Splatting Scenes
May 29, 2026Finn Dröge, Cecilia Curreli, Abhishek Saroha +1Inpainting Diverse Occlusion-And-Revelation Scenarios
1Technical University of Munich · 2Munich Center for Machine Learning
3D-GIMP: When 3D Gaussian Inpainting Meets PatchMatch
Jul 22, 2026Xuening Tian, Dieter Schmalstieg, Shohei Mori3D Editing 3D Reconstruction
University of Stuttgart Stuttgart, Germany
CoIn: Comprehensive 2D-3D Inpainting with Gaussian Splatting Guidance
Jun 25, 2026Hana Kim, Minje Kim, Tae-Kyun Kim3D Editing Gaussian Splatting
LG Electronics, Seoul, South Korea · KAIST, Daejeon, South Korea

May 29, 2026, cs.CVJ/K move · Enter open · S save

Benchmarking Single-Step Inpainting Methods for Multi-Object 3D Gaussian Splatting Scenes

Finn Dröge, Cecilia Curreli, Abhishek Saroha, Daniel Cremers

1Technical University of Munich · 2Munich Center for Machine Learning

Inpainting Diverse Occlusion-And-Revelation Scenarios Scene Understanding

Open paper AskPDF arXiv

The tasks of object removal and inpainting 3D Gaussian Splatting (3DGS) scenes face challenges such as 3D consistency across camera views. In comparing 2D inpainters and their suitability for the 3D domain, we find that reconstruction-based inpainters outperform generative diffusion models in 3D consistency. Integrating these 2D inpainters into different single-step methods for creating and finetuning 3DGS scenes, our results indicate that initializing the scene from scratch produces higher quality results than finetuning the existing scene. Using a state-of-the-art generative 2D inpainter, we create a straightforward baseline to underline the importance of object removal before inpainting in the 3D setting. Since 360° datasets rarely include real-world ground truths, and challenging occlusion scenarios are equally sparse, we introduce a novel multi-object scene with recorded ground truth data and many views with object occlusions.

Abstract

Explore similar work

Benchmarking Single-Step Inpainting Methods for Multi-Object 3D Gaussian Splatting Scenes

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

CoIn: Comprehensive 2D-3D Inpainting with Gaussian Splatting Guidance