2Xplat: Decoupling Geometry and Appearance Modeling for Feed-Forward 3D Gaussian Splatting

Hwasik Jeong; Seungryong Lee; Gyeongjin Kang; Seungkwon Yang; Xiangyu Sun; Seungtae Nam; Eunbyung Park

cs.CVMar 22, 2026

2Xplat: Decoupling Geometry and Appearance Modeling for Feed-Forward 3D Gaussian Splatting

Authors: Hwasik Jeong, Seungryong Lee, Gyeongjin Kang, Seungkwon Yang, Xiangyu Sun, Seungtae Nam, Eunbyung Park

Organizations: Yonsei University · Sungkyunkwan University

Abstract

Pose-free feed-forward 3D Gaussian Splatting (3DGS) has opened a new frontier for rapid 3D modeling, enabling high-quality Gaussian representations to be generated from uncalibrated multi-view images in a single forward pass. The dominant approach adopts unified monolithic architectures, often built on geometry-centric 3D foundation models, to jointly estimate camera poses and synthesize 3DGS representations within a single network, entangling geometric reasoning and appearance modeling within a shared representation. In this work, we introduce 2Xplat, a pose-free feed-forward 3DGS framework based on a two-experts design that explicitly separates geometry estimation from Gaussian generation: a dedicated geometry expert first predicts camera poses, which are then passed to an appearance expert that synthesizes 3D Gaussians. Despite its conceptual simplicity, and being largely underexplored in prior works, our two-experts pipeline outperforms prior pose-free feed-forward 3DGS approaches in fewer than 5K training iterations, achieving performance on par with state-of-the-art posed methods. These results challenge the prevailing unified paradigm and suggest the potential advantages of modular design for complex 3D geometric estimation and appearance synthesis tasks.

Figures & tables

Figure 2: Two experts vs. one generalist: (a) The prevailing monolithic architecture employs a shared backbone with task-specific heads to jointly predict camera poses and 3DGS representations in a single forward pass. (b) The two-experts architecture explicitly decomposes the process into sequential stages: a geometry expert first estimates camera poses from multi-view input images, and a dedicated appearance (posed feed-forward 3DGS) expert subsequently generates 3D Gaussian representations conditioned on the predicted poses and input images.

Method	$p$	$k$	6v			12v			24v
Method	$p$	$k$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
MVSplat	✓	✓	22.659	0.760	0.173	21.289	0.709	0.224	19.975	0.662	0.269
DepthSplat	✓	✓	23.418	0.797	0.136	21.911	0.753	0.179	20.088	0.690	0.240
YoNoSplat	✓	✓	24.717	0.817	0.139	23.285	0.773	0.177	22.664	0.758	0.192
Ours	✓	✓	25.217	0.827	0.150	24.592	0.813	0.169	24.354	0.811	0.174
Ours+vdo	✓	✓	26.631	0.856	0.121	27.240	0.866	0.115	27.413	0.877	0.109

Table 1: Novel view synthesis results on DL3DV with 6, 12, and 24 input views at low-resolution (

224\times 224

p

and

k

denote the use of ground-truth poses and intrinsics, respectively. † indicates evaluation with EPA. ‘vdo’ denotes the use of view-dependent opacity and regularization in the appearance expert, and this variant is excluded from bold-letter ranking for fair comparison with baselines.

Figure 3: Qualitative comparison on DL3DV with varying numbers of input views.

Method	Optim.	(p, k)	16v			32v			64v			128v
Method	Optim.	(p, k)	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
3D-GS 30k	✓	✓	21.48	0.753	0.252	24.43	0.827	0.191	27.34	0.883	0.146	29.43	0.914	0.123
Long-LRM		✓	21.05	0.708	0.297	23.97	0.778	0.267	23.60	0.789	0.260	21.24	0.739	0.308
iLRM		✓	21.92	0.748	0.316	24.30	0.803	0.256	24.44	0.819	0.240	22.98	0.807	0.249
MVP+vdo		✓	23.76	0.798	0.239	25.96	0.847	0.187	27.73	0.881	0.154	29.02	0.903	0.134
Ours+vdo			22.90	0.754	0.259	24.75	0.800	0.208	26.11	0.830	0.180	27.16	0.853	0.162

Table 2: Quantitative comparison on the DL3DV dataset under varying numbers of input views (16, 32, 64, and 128) for high-resolution (

960\times 540

) novel view synthesis.

Figure 4: Qualitative comparison on high-resolution DL3DV ( $960\times 540$ ).

Figure 5: Qualitative comparison on low-resolution RE10K dataset with 6 context views.

Method	$p$	$k$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
DepthSplat	✓	✓	24.156	0.846	0.145
YoNoSplat	✓	✓	25.037	0.848	0.134
NoPoSplat †		✓	22.175	0.750	0.207
YoNoSplat †		✓	25.395	0.857	0.131
Ours †		✓	27.108	0.877	0.128
YoNoSplat			19.723	0.613	0.229

Table 3: Quantitative comparison on low-resolution RE10K dataset with 6 context views.

Method	32v	64v	128v
AnySplat	14.05	15.98	16.99
YoNoSplat † w/o GT $k$	16.89	17.37	17.64
YoNoSplat † w/ GT $k$	17.94	18.83	19.28
Ours † w/o GT $k$	18.04	19.64	20.74
Ours † w/ GT $k$	18.09	19.74	20.88

Table 4: Cross-dataset generalization from DL3DV to ScanNet++ (PSNR

\uparrow

Figure 6: Qualitative cross-dataset generalization from DL3DV to ScanNet++.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
FLARE	20.07	0.606	0.315
NoPoSplat †	20.36	0.636	0.291
YoNoSplat	18.68	0.526	0.337
YoNoSplat †	20.22	0.617	0.304
Ours	19.56	0.574	0.337
Ours †	20.06	0.606	0.329

Table 5: Sparse-view (two context views) NVS on DL3DV.

Method	size	speed	PSNR $\uparrow$	LPIPS $\downarrow$
DA3-G †	1.1B	0.65s	23.77	0.165
YoNoSplat † ( $\pi^{3}$ )	1B	0.33s	24.53	0.142
Ours † (DA3-L)	0.5B	0.15s	24.93	0.160
Ours † (DA3-G)	1.3B	0.31s	25.23	0.151
Ours † ( $\pi^{3}$ )	1.1B	0.27s	25.08	0.156

Table 6: Efficiency trade-off (RTX 3090 inference speed).

Backbone	Method	6v	12v	24v
DA3-G	DA3-G †	23.77	22.38	21.69
DA3-G	Ours †	25.23	24.36	24.08
DA3-L	DA3-L †	21.87	20.93	20.61
DA3-L	Ours †	24.93	23.87	23.48
$\pi^{3}$	YoNoSplat	22.29	20.38	19.71
	YoNoSplat †	24.53	22.93	22.17

Table 7: Monolithic vs. two-experts architectures (PSNR

\uparrow

	6v		12v		24v
	PSNR $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	LPIPS $\downarrow$
$\times$ †	21.04	0.267	19.96	0.322	19.38	0.360
$\circ$ †	25.23	0.151	24.36	0.177	24.08	0.185

Table 8: Joint training analysis on DL3DV

224\times 224

Method	(p, k)	16v	32v	64v
iLRM	✓	21.92	24.30	24.44
Ours † (iLRM)		19.71	22.84	22.10
Long-LRM	✓	21.05	23.97	23.60
Ours † (Long-LRM)		20.26	22.41	22.06

Table 9: Generality across appearance experts (PSNR

\uparrow

Method	5 ${}^{\circ}\uparrow$	10 ${}^{\circ}\uparrow$	20 ${}^{\circ}\uparrow$	Backbone
MASt3R 518×288	0.609	0.776	0.878	-
VGGT 518×280	0.566	0.753	0.867	-
$\pi^{3}$ 518×280	0.705	0.841	0.916	-
DA3 504×504	0.694	0.826	0.900	-
NoPoSplat 256×256	0.443	0.627	0.755	MASt3R
YoNoSplat 224×224	0.722	0.852	0.923	$\pi^{3}$

Table 10: Pose estimation comparison on RE10K

224\times 224

(a) 4, 8, 16 input views
Method	Tanks & Temples			Mip-NeRF360			ETH3D			Average
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Ours+vdo †	15.52	0.542	0.507	17.38	0.435	0.581	16.62	0.543	0.553	16.51	0.507	0.547
Ours+vdo †	15.75	0.533	0.506	18.35	0.432	0.541	17.49	0.549	0.537	17.20	0.505	0.528
(b) 32, 64, 128 input views
Method	Tanks & Temples			Mip-NeRF360			ETH3D			Average

Table 11: Average NVS generalization results across two input-view-count regimes: (a) 4, 8, and 16 input views, and (b) 32, 64, and 128 input views, evaluated on three datasets in the high-resolution setting (

960\times 540

). The last three columns report the mean PSNR, SSIM, and LPIPS across the three datasets. The gray-colored row denotes 2Xplat results using MVP variant as the appearance backbone, while the non-gray row uses the original MVP. Full results are provided in table 14 .

Loss type	PSNR $\uparrow$	LPIPS $\downarrow$	5 ${}^{\circ}\uparrow$	10 ${}^{\circ}\uparrow$	20 ${}^{\circ}\uparrow$
Rel. loss	26.161	0.131	0.718	0.843	0.912
Abs. loss	25.704	0.137	0.641	0.797	0.888
W/O loss	26.369	0.129	0.686	0.836	0.905

Table 12: Pose supervision ablation on DL3DV

224\times 224

Figure 7: Comparison between fixed and dynamic grouping. (a) Fixed grouping restricts each token to attend only within its assigned group, limiting network capacity even during later global attention. (b) Dynamic grouping relaxes this inductive bias (purple dotted arrows), giving the network greater capacity to attend across the full view during global attention.

	MVP	MVP variant
Model size	240M	559M
Training dataset	DL3DV ( Ling et al. 2024 )	DL3DV ( Ling et al. 2024 ) RE10K ( Zhou et al. 2018 ) WildRGBD ( Xia et al. 2024 ) CO3Dv2 ( Reizenstein et al. 2021 ) MegaSynth ( Jiang et al. 2025b ) Syn4D ( Jiang et al. 2026 ) ScanNet++v2 ( Yeshwanth et al. 2023 ) BlendedMVS ( Yao et al. 2020 ) SWORD ( Solovev et al. 2023 ) UnrealStereo4K ( Tosi et al. 2021 ) MVS-Synth ( Huang et al. 2018 )
Training resolutions	(480, 270) (960, 540)	(512, 512) (512, 384) (512, 336) (512, 288) (384, 512) (336, 512) (288, 512) (688, 512) (768, 512) (912, 512) (512, 688) (512, 768) (512, 912)
Grouping mechanism	Fixed grouping	Dynamic grouping

Table 13: Key differences between the original MVP and the its variant used as the appearance expert.

Method	Views	Tanks & Temples			Mip-NeRF360			ETH3D
		PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Ours+vdo †	4	13.96	0.494	0.604	14.75	0.375	0.682	13.61	0.498	0.650
Ours+vdo †		14.33	0.491	0.600	16.86	0.389	0.628	15.81	0.522	0.623
Ours+vdo †	8	15.12	0.534	0.516	17.48	0.434	0.586	16.75	0.547	0.555
Ours+vdo †		15.50	0.530	0.509	18.34	0.432	0.538	17.44	0.552	0.532
Ours+vdo †	16	17.49	0.597	0.402	19.90	0.497	0.476	19.51	0.585	0.454

Table 14: Cross Dataset Generalization (Tanks & Temples, Mip-NeRF360, ETH3D). The gray-colored row denotes 2Xplat results using MVP variant as the appearance backbone, while the non-gray row uses the original MVP. All context and target views are sampled uniformly from the original dataset resolution. Results for 128 views on ETH3D are omitted due to an insufficient number of available input views.

Figure 8: Visualization of the averaged global attention map in the last layer for (a) MVP and (b) MVP variant. Given a patch in the reference frame (top-left, outlined in red), MVP variant attends to all relevant patches across every input frame, resulting in more 3D-consistent attention, whereas MVP attends to only a few.

Figure 9: Qualitative comparison on DL3DV ( $224\times 224$ resolution, 1-3 rows: 6 context views, 4-6 rows: 12 context views, 7-9 rows: 24 context views).

Figure 10: Qualitative comparison on RE10K dataset with 6 context views ( $224\times 224$ resolutions).

Figure 11: Qualitative cross-dataset generalization results from DL3DV to ScanNet++ ( $224\times 224$ resolution, 1-2 rows: 32 context views, 3-4 rows: 64 context views, 5-6 rows: 128 context views).

Figure 12: Qualitative comparison on high-resolution DL3DV ( $960\times 540$ , 1-2 rows: 16 context views, 3-4 rows: 32 context views, 5-6 rows: 64 context views, 7-8 rows: 128 context views).

Explore similar work

AESplat: Advancing Pose-Free Feed-Forward 3D Gaussian Splatting via Decoupled Appearance Modeling
Sep 29, 2026Shiwei Ren, Zhiang Liu, Yongchun Fang +13D Gaussian Splatting Novel View Synthesis
Nankai University
QuerySplat: Decoupling Geometry and Appearance Representations in 3DGS Prediction
Aug 2, 2026Yinglong Li, Donghui Shen, Xiaoyu Zhang +53D Gaussian Splatting Representation Disentanglement
State Key Laboratory of Virtual Reality Technology and Systems, Beihang University · InSpatio Research · State Key Lab of CAD&CG, Zhejiang University
DeltaSplat: Iterative Gaussian Refinement for Pose-Free Feed-Forward 3D Gaussian Splatting
Oct 7, 2026Chanung Park, Seunghyeon Song, Joo Chan Lee +23D Gaussian Splatting 3D Reconstruction
Department of Electrical and Computer Engineering, Sungkyunkwan University · Department of Artificial Intelligence, Sungkyunkwan University · Department of Artificial Intelligence, Yonsei University

Sep 29, 2026J/K move · Enter open · S save

AESplat: Advancing Pose-Free Feed-Forward 3D Gaussian Splatting via Decoupled Appearance Modeling

Shiwei Ren, Zhiang Liu, Yongchun Fang, Hongwei Chen

Nankai University

3D Gaussian Splatting Novel View Synthesis Feed-Forward 3D Reconstruction

Open paper AskPDF arXiv

Pose-free feed-forward 3D Gaussian Splatting (3DGS) has demonstrated remarkable potential for generalized novel view synthesis. However, existing methods typically predict Gaussian appearance attributes represented by spherical harmonics (SH) in the same manner, overlooking the fundamental distinction between view-independent and view-dependent appearance, which results in suboptimal rendering quality. In this paper, we present AESplat, a novel and general framework for pose-free feed-forward 3DGS that introduces an effective decoupled appearance modeling strategy based on an analysis of SH, enabling higher-quality rendering. Specifically, AESplat directly derives the zeroth-order SH coefficient, which represents the base view-independent appearance component, from the input images without training. The higher-order SH coefficients are subsequently predicted by a shallow multilayer perceptron equipped with two efficient 3D-aware inductive biases to model view-dependent appearance variations. Extensive experiments across multiple datasets demonstrate that our method significantly outperforms state-of-the-art approaches, achieving a

0.8

dB improvement in PSNR over the pose-free method NAS3R and a

1.1

dB improvement over the pose-required method DepthSplat on the RealEstate10K dataset. Project page: https://aesplat.github.io/.

Abstract

Explore similar work

AESplat: Advancing Pose-Free Feed-Forward 3D Gaussian Splatting via Decoupled Appearance Modeling

QuerySplat: Decoupling Geometry and Appearance Representations in 3DGS Prediction

DeltaSplat: Iterative Gaussian Refinement for Pose-Free Feed-Forward 3D Gaussian Splatting