Generating explorable 3D scenes from a single image is a highly challenging problem in 3D vision. Existing methods struggle to support free exploration, often producing severe geometric distortions and noisy artifacts when the viewpoint moves far from the original perspective. We introduce One2Scene, an effective framework that decomposes this ill-posed problem into three tractable sub-tasks to enable immersive explorable scene generation. We first use a panorama generator to produce anchor views from a single input image as initialization. Then, we lift these 2D anchors into an explicit 3D geometric scaffold via a generalizable, feed-forward Gaussian Splatting network. Instead of treating the panorama as a single image for reconstruction, we project it into multiple sparse anchor views and reformulate the reconstruction task as multi-view stereo matching, which allows us to leverage robust geometric priors learned from large-scale multi-view datasets. A bidirectional feature fusion module is used to enforce cross-view consistency, yielding an efficient and geometrically reliable scaffold. Finally, the scaffold serves as a strong prior for a novel view generator to produce photorealistic and geometrically accurate views at arbitrary cameras. By explicitly conditioning on a 3D-consistent scaffold to perform reconstruction, One2Scene works stably under large camera motions, supporting immersive scene exploration. Extensive experiments show that One2Scene substantially outperforms state-of-the-art methods in panorama depth estimation, feed-forward 360° reconstruction, and explorable 3D scene generation. Project page: https://one2scene5406.github.io/
Figures & tables
Figure 1: Comparison on large-viewpoint novel view synthesis. Existing methods such as Wonderjourny ( Yu et al., 2023 ) and Dreamscene360 ( Zhou et al., 2024 ) exhibit clear geometric distortions and artifacts, while our method generates photorealistic and geometrically accurate novel views. The input image is highlighted by a red bounding box. The other images represent the novel views.
Figure 2: Overview of One2Scene. Our method consists of three stages: (a) an anchor view generation stage to establish an initial 360-degree representation, (b) a feed-forward 3D Gaussian Splatting stage to construct an explicit 3D geometric scaffold, and (c) a synthesis stage that leverages the scaffold information to produce high-quality novel views. The pipeline enables geometrically consistent and photorealistic novel view synthesis from a single input image.
Figure 3: Qualitative comparison. Our method retains compelling visual quality and generates plausible continuations of the scene, even under large viewpoint change.
Methods
NIQE ↓
Q-Align ↑
CLIP-I ↑
TransErr ↓
RotErr ↓
CamMC ↓
DreamScene360 ( Zhou et al. (2024) )
8.40
1.91
74.24
-
-
-
WonderJourney ( Yu et al. (2023) )
4.97
3.02
77.92
-
-
-
SEVA ( Zhou et al. (2025) )
4.53
3.20
87.82
0.460
0.165
0.558
SEVA ( Zhou et al. (2025) ) + Anchor
4.45
3.45
88.70
0.422
0.116
0.460
VMem ( Li et al. (2025a) )
6.86
2.95
75.80
0.573
0.569
0.998
VMem ( Li et al. (2025a) ) + Anchor
5.23
3.04
81.33
0.613
0.426
0.887
Table 1: Quantitative comparisons for 3D scene generation.
Methods
NIQE ↓
Q-Align ↑
CLIP-I ↑
TransErr ↓
RotErr ↓
CamMC ↓
AnySplat ( Jiang et al., 2025 )
4.96
3.61
81.96
0.332
0.367
0.616
Ours
4.43
4.13
89.95
0.326
0.107
0.389
Table 2: Comparison on the 3D scene generation performance by replacing our feed-forward 360° reconstruction network with AnySplat.
Methods
Matterport3D
Stanford2D3D
AbsRel ↓
δ1↑
δ2↑
δ3↑
AbsRel ↓
δ1↑
δ2↑
δ3↑
BiFuse ( Wang et al., 2020 )
0.2048
84.52
93.19
96.32
0.1209
86.60
95.80
98.60
UniFuse ( Jiang et al., 2021 )
0.1063
88.97
96.23
98.31
0.1114
87.11
96.64
98.82
HoHoNet ( Sun et al., 2020 ) )
0.1488
87.86
95.19
97.71
0.1014
90.54
96.93
98.86
BiFuse++ ( Wang et al., 2022 )
−
87.90
95.17
97.72
−
87.83
96.49
98.84
ACDNet ( Zhuang et al., 2022 )
0.1010
90.00
96.78
98.76
0.0984
88.72
97.04
98.95
Table 3: Comparison of depth estimation on Matterport3D and Stanford2D3D datasets.
Figure 4: Ablation study on reconstruction performance. We compare the 3D scene generation quality by replacing our feedforward network with AnySplat. Top row: reconstruction results. Bottom row: generation results using our model.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure A1: Qualitative comparison for the ablation study. (a) Render views from our 3D scaffold. (b) Naive concatenation baseline. (c) Ours (Dual-LoRA training only). (d) Ours (Full model with memory condition).
Methods
NIQE ↓
Q-Align ↑
CLIP-I ↑
TransErr ↓
RotErr ↓
CamMC ↓
Naive Concat.
5.04
3.41
85.30
0.481
0.260
0.655
Dual-LoRA Training
4.42
4.10
89.51
0.326
0.119
0.401
+ Memory Condition
4.43
4.13
89.95
0.326
0.107
0.389
Appendix
Table A1: Ablation study on 3D scaffold guided novel view synthesis.
Methods
Matterport3D
Stanford2D3D
AbsRel ↓
δ1↑
δ2↑
δ3↑
AbsRel ↓
δ1↑
δ2↑
δ3↑
Baseline
0.1576
78.82
93.20
96.15
0.1497
81.99
93.53
97.88
w/o BF
0.1204
86.28
95.36
97.45
0.0797
94.31
97.42
98.85
w BF
0.1070
88.97
96.51
98.61
0.0675
95.20
98.53
99.30
Appendix
Table A2: Effectiveness of BF module. Zero-shot quantitative comparison on Matterport3D and Stanford2D3D datasets.
Methods
PSNR ( ↑ )
SSIM ( ↑ )
LPIPS ( ↓ )
DIST ( ↓ )
FID ( ↓ )
PixelSplat
15.32
0.422
0.517
0.374
139.75
MVSplat
15.94
0.441
0.459
0.282
73.91
MVSplat360
16.37
0.453
0.439
0.238
18.32
Ours
17.35
0.506
0.343
0.181
16.84
Appendix
Table A3: The NVS numerical comparison on the DL3DV ( Ling et al., 2023 ) dataset.
Figure A2: Visual comparison with existing SOTA methods on DL3DV.
Figure A3: Qualitative comparison between One2Scene and SOTA methods.
Figure A4: Qualitative comparison between One2Scene and SOTA methods.
Figure A5: Qualitative comparison between One2Scene and SOTA methods.
Figure A6: Qualitative comparison between One2Scene and SOTA methods.
Figure A7: Qualitative comparison between One2Scene and SOTA methods.
We introduce ReconSplat, a feed-forward model for 3D scene reconstruction that aims to address the longstanding trade-off between plausible view generation for unobserved regions and geometric consistency, providing both geometrically aligned novel views and sharp depth estimates. Our approach builds on 3D Gaussian splatting (3DGS) as an intermediate differentiable scene representation and integrates it with a multi-view latent diffusion model (MV-LDM) trained to act simultaneously as a refiner and an inpainter for appearance and scene geometry. We enforce geometric consistency by guiding the diffusion process with variational 3D latent features for appearance and geometry, encoded by the feed-forward 3DGS representation and rasterized to 2D latent space. ReconSplat produces both photorealistic novel views and accurate depth maps on real-world benchmarks, RealEstate10K and DL3DV-10K, outperforming existing methods in challenging extrapolation setups. Notably, ReconSplat allows the extrapolation of unseen and challenging viewpoints jointly with coherent and precise scene geometry.
Giuseppe Stracquadanio, Kevin Raj, Julia Grabinski +1
We present Genie Sim PanoRecon, a feed-forward Gaussian-splatting pipeline that delivers high-fidelity, low-cost 3D scenes for robotic manipulation simulation. The panorama input is decomposed into six non-overlapping cube-map faces, processed in parallel, and seamlessly reassembled. To guarantee geometric consistency across views, we devise a depth-aware fusion strategy coupled with a training-free depth-injection module that steers the monocular feed-forward network to generate coherent 3D Gaussians. The whole system reconstructs photo-realistic scenes in seconds and has been integrated into Genie Sim - a LLM-driven simulation platform for embodied synthetic data generation and evaluation - to provide scalable backgrounds for manipulation tasks. For code details, please refer to: https://github.com/AgibotTech/genie_sim/tree/main/source/geniesim_world.
Current compositional image-to-3D scene generation approaches construct 3D scenes by time-consuming iterative layout optimization or inflexible joint object-layout generation. Moreover, most methods rely on limited field-of-view perspective images, hindering the creation of complete 360-degree environments. To address these limitations, we design Pano3DComposer, an efficient feed-forward framework for panoramic images. To decouple object generation from layout estimation, we propose a plug-and-play Object-World Transformation Predictor. This module converts the 3D objects generated by off-the-shelf image-to-3D models from local to world coordinates. To achieve this, we adapt the VGGT architecture to Alignment-VGGT by using target object crop, multi-view object renderings and camera parameters to predict the transformation. The predictor is trained using pseudo-geometric supervision to address the shape discrepancy between generated and ground-truth objects. For input images from unseen domains, we further introduce a Coarse-to-Fine (C2F) alignment mechanism for Pano3DComposer that iteratively refines geometric consistency with feedback of scene rendering. Our method achieves superior geometric accuracy for image/text-to-3D tasks on synthetic and real-world datasets. It can generate a high-fidelity 3D scene in approximately 20 seconds on an RTX 4090 GPU. Project page: https://qiuzidian.github.io/pano3dcomposer-page/.