While 3D Vision Foundation Models (3DVFMs) have demonstrated remarkable zero-shot capabilities in visual geometry estimation, their direct application to generalizable novel view synthesis (NVS) remains challenging. In this paper, we propose AirSplat, a novel training framework that effectively adapts the robust geometric priors of 3DVFMs into high-fidelity, pose-free NVS. Our approach introduces two key technical contributions: (1) Self-Consistent Pose Alignment (SCPA), a training-time feedback loop that ensures pixel-aligned supervision to resolve pose-geometry discrepancy; and (2) Rating-based Opacity Matching (ROM), which leverages the local 3D geometry consistency knowledge from a sparse-view NVS teacher model to filter out degraded primitives. Experimental results on large-scale benchmarks demonstrate that our method significantly outperforms state-of-the-art pose-free NVS approaches in reconstruction quality. Our AirSplat highlights the potential of adapting 3DVFMs to enable simultaneous visual geometry estimation and high-quality view synthesis.
Figures & tables
Figure 1: (a) Our proposed AirSplat adapts 3D-VS-VFMs using Self-Consistent Pose Alignment (SCPA) and Rating-based Opacity Matching (ROM) to resolve inherent pose-geometry discrepancies and multi-view inconsistencies. Compared to (b) baseline DA3 [ 22 ] which shows floaters (red boxes) and blurred regions (dashed yellow boxes), (c) AirSplat eliminates these artifacts, faithfully recovering the scene structure shown in (d) the ground truth.
Figure 2: Overview of our AirSplat training pipeline. During training of a Gaussian head of a 3DVFM, the encoder and its pose head are frozen. Our SCPA module corrects the predicted target pose to align the rendered target views to the GT target views. Our ROM module gathers the geometric feedback from the teacher model, and based on the feedback, it enhances the multi-view consistency of the predicted 3D primitives.
Figure 3: Comparison of training paradigms in pose-free NVS. (a) Training with the context-only strategy, adopted in [ 16 ] , leads to a lack of direct supervision for novel viewpoints. (b) The context-target strategy, following [ 14 ] , results in spatial misalignment. (c) Our SCPA corrects the inherent spatial drift, enabling the network to learn both structurally consistent 3D geometry and robust novel view synthesis.
Figure 4: Effect of Self-Consistent Pose Alignment (SCPA). We compare rendered target views using the initial predicted pose P^tgt,t(1) , our aligned pose P^tgt,talign against the ground truth, during training. As highlighted by the red arrows, the initial pose prediction results in a noticeable spatial shift, evident in the misaligned structural lines on the ground. Our proposed SCPA corrects the inherent spatial drift and ensures that the model learn structurally consistent 3D geometry and robust NVS. In the error maps, blue indicates small errors, and red indicates large errors.
Methods
12 Views
24 Views
36 Views
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
w/ Pose
MonoSplat [ 25 ] (CVPR25)
18.16
0.663
0.336
16.66
0.593
0.391
15.79
0.551
0.424
MVSplat [ 8 ] (ECCV24)
17.98
0.638
0.357
17.27
0.609
0.379
OOM
OOM
OOM
DepthSplat [ 56 ] (CVPR25)
22.56
0.793
0.200
21.00
0.734
0.248
19.60
0.676
0.296
Pose- free
NoPoSplat [ 59 ] (ICLR25)
17.15
0.571
0.437
17.10
0.570
0.443
17.09
0.570
0.446
AnySplat [ 16 ] (SIGGRAPHAsia25)
18.69
0.591
0.273
19.15
0.610
0.257
19.31
0.615
0.251
Table 1: Quantitative comparison of NVS performance on the RE10K dataset [ 66 ] under various input-view settings. Bold indicates the best performance. OOM indicates out-of-memory inference error.
Figure 5: Qualitative comparison of NVS performance on RE10K dataset [ 66 ] .
Methods
12 Views
24 Views
36 Views
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
w/ Pose
MVSplat [ 8 ] (ECCV24)
21.55
0.729
0.239
OOM
OOM
OOM
OOM
OOM
OOM
DepthSplat [ 56 ] (CVPR25)
22.14
0.725
0.221
19.87
0.695
0.274
18.69
0.643
0.330
Pose- free
NoPoSplat [ 59 ] (ICLR25)
16.13
0.447
0.497
14.73
0.392
0.603
13.82
0.365
0.640
AnySplat [ 16 ] (SIGGRAPHAsia25)
18.72
0.551
0.310
18.40
0.533
0.333
18.36
0.529
0.337
WorldMirror [ 26 ] (arXiv25)
20.44
0.625
0.278
19.78
0.594
0.312
19.67
0.589
0.316
Table 2: Quantitative comparison of NVS performance on the DL3DV dataset [ 23 ] under various input-view settings.
Figure 6: Qualitative comparison of NVS performance on DL3DV dataset [ 23 ] .
Methods
16 Views
20 Views
24 Views
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
w/ Pose
MonoSplat [ 25 ] (CVPR25)
17.81
0.642
0.345
17.48
0.625
0.361
17.23
0.611
0.372
MVSplat [ 8 ] (ECCV24)
18.19
0.502
0.376
18.12
0.500
0.379
18.09
0.500
0.379
DepthSplat [ 56 ] (CVPR25)
20.41
0.713
0.268
19.92
0.694
0.286
19.78
0.688
0.289
Pose- free
NoPoSplat [ 59 ] (ICLR25)
22.30
0.668
0.286
22.24
0.666
0.288
22.25
0.666
0.288
AnySplat [ 16 ] (SIGGRAPHAsia25)
21.89
0.615
0.275
22.05
0.622
0.256
21.96
0.619
0.258
Table 3: Cross-dataset generalization performance under various input-view settings. We evaluate the zero-shot performance on the ACID dataset [ 24 ] .
Figure 10
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Our strategy for partitioning input context sequences to compute the teacher’s geometric errors.
Figure 9: Pose-Geometry Discrepancy. To validate that this structural misalignment is not confined to a specific baseline, we visualize the consistent spatial drift across distinct models and diverse scenes: (a) DepthAnything3 [ 22 ] evaluated on the DL3DV dataset [ 23 ] , and (b) SPFSplat [ 14 ] on the RE10K dataset [ 66 ] .
Methods
PSNR ↑
SSIM ↑
LPIPS ↓
SPFSplat [ 14 ]
21.32
0.694
0.266
SPFSplat +AirSplat [ 14 ]
22.76
0.793
0.206
WorldMirror [ 26 ]
21.08
0.701
0.274
WorldMirror [ 26 ] +AirSplat
23.66
0.811
0.182
DA3 [ 22 ]
21.06
0.710
0.254
DA3 +AirSplat [ 22 ]
23.77
0.814
0.178
Appendix
Table 5: Additional baselines results on RE10K dataset under 24 input views.
Methods
PSNR ↑
SSIM ↑
LPIPS ↓
Baseline
20.74
0.691
0.242
Baseline + Context-only Training
21.39
0.717
0.233
Baseline + Context-Target Training
21.52
0.727
0.221
Baseline + Context-Target w/ SCPA Training
22.29
0.740
0.214
Baseline + ROM
22.11
0.731
0.218
Ours (Full)
22.50
0.747
0.207
Appendix
Table 6: Ablation analysis on DL3DV dataset [ 23 ] under 12 input views.
Method
View
Acc. ↓
Comp. ↓
NC. ↑
Mean
Med.
Mean
Med.
Mean
Med.
VGGT [ 49 ]
sparse
0.044
0.025
0.056
0.033
0.733
0.845
AnySplat [ 16 ]
0.080
0.053
0.120
0.072
0.684
0.785
π3 [ 53 ]
0.047
0.029
0.075
0.049
0.742
0.841
DA3 [ 22 ] /AirSplat
0.049
0.034
0.065
0.046
0.757
0.866
VGGT [ 49 ]
dense
0.022
0.008
0.026
0.012
0.666
0.760
Appendix
Table 7: Pose-free point map estimation on 7-Scenes [ 42 ] dataset.
Methods
Avg. Time (s) / Iter.
Baseline + Context-Target
2.35
Baseline + Context-Target w/ SCPA
3.67
Ours (Full Model)
3.89
Appendix
Table 8: Training time analysis. We report the average time per training iteration. While our proposed modules (SCPA and ROM) introduce additional forward passes and teacher model evaluations, the overall training remains tractable.
Figure 10: Qualitative comparison of NVS performance on DL3DV dataset [ 23 ] .
We introduce SPFSplatV2, an efficient feed-forward framework for 3D Gaussian splatting from sparse multi-view images, requiring no ground-truth poses during training or inference. The framework employs a shared feature extraction backbone to jointly predict 3D Gaussian primitives and camera poses in a canonical space from unposed inputs. To enable efficient and accurate pose estimation, we introduce a masked attention mechanism for target-view pose prediction and a reprojection loss that enforces pixel-aligned Gaussian primitives, providing stronger geometric constraints. We further demonstrate the compatibility of our training framework with different reconstruction architectures, resulting in two model variants. Remarkably, despite the absence of pose supervision, our method achieves state-of-the-art performance in both in-domain and out-of-domain novel view synthesis, even under extreme viewpoint changes and limited image overlap. It also surpasses many methods that rely on geometric supervision in relative pose estimation. By eliminating dependence on ground-truth poses, our method offers the scalability to leverage larger and more diverse datasets. Code and pretrained models will be available on our project page: https://ranrhuang.github.io/spfsplatv2/.
Ranran Huang, Krystian Mikolajczyk
Department of Electrical and Electronic Engineering, Imperial College London, London SW7 2AZ, United Kingdom
Generating photorealistic novel views from unposed images requires both 3D geometric understanding and the ability to synthesize unseen content. A natural strategy combines feed-forward 3DGS reconstruction with multi-view diffusion. Yet prior pipelines extract at most one signal from the reconstruction, either pixel rendering or learned features, while none exploits per-Gaussian visibility for occlusion-aware reference selection. This information disconnect leaves renderable geometry, visibility cues, and learned features unused. SplatGuide closes this disconnect by reusing a single 3DGS scene across three complementary roles. Rendered images provide pixel-aligned geometric conditioning. Per-Gaussian source-view indices are rendered into a target-view voting map for occlusion-aware reference selection. Reconstruction tokens supply feature-level guidance via cross-attention. All three signals derive from the same reconstruction forward pass. Across RealEstate10K, DL3DV, Tanks-and-Temples, and Mip-NeRF 360, SplatGuide achieves state-of-the-art pose-free novel view synthesis. On RealEstate10K, with a moderate number of input views, it surpasses the ground-truth-pose baseline.
Yejun Zhang, Zihan Wang, Xu Ji +8
Aalto University, Finland · Deep Render, UK · ELLIS Institute Finland +3
Single-image novel view synthesis remains challenging because the underlying 3D geometry is highly ambiguous. Recent diffusion-based approaches produce plausible results, but they often struggle to preserve the geometric structure and spatial coherence of foreground objects. We present GenNVS, a framework for geometry-enhanced novel view synthesis via a disentangled 3D prior. Specifically, GenNVS models foreground objects and the background with 3D Gaussian Splatting and aligns them through a coarse-to-fine geometric optimization process to form a unified 3D scene. This scene conditions a video diffusion model through the proposed Dual-Stream Masking mechanism, which guides synthesis by jointly exploiting rendered validity masks and geometry-aware warping. Experimental results show that GenNVS performs favorably against recent methods in both visual quality and geometric accuracy, while naturally supporting flexible scene editing.