Generative novel view synthesis from sparse input images is rarely all reconstruction or all generation: pixels visible in some source view have a unique correct value modulated only by view-dependent shading, while pixels in disocclusions or beyond the captured volume admit a distribution of plausible completions. Existing generative novel-view-synthesis methods conflate these regimes under a single uniform loss, blurring the line between geometric fidelity and creative hallucinations even when scene geometry is injected through warped point clouds or projected depth. We introduce GenRec, a multi-view flow matching model that builds the reconstruction--generation split directly into its architecture, supervision, and gradient flow. Guided by an observation mask derived from the source cameras and a monocular depth estimator, a flow matching backbone jointly denoises RGB and scene-coordinate maps across all target views, while a pixel-space refinement stage restores high-frequency detail on observed pixels; the same mask gates supervision so regression signals do not contaminate the generative prior. Across RealEstate10K, DL3DV-10K, and Mip-NeRF~360, in both single-view extrapolation and two-view interpolation, GenRec attains the best reconstruction fidelity in observed regions while also surpassing purely generative baselines on perceptual quality in unobserved ones, showing the effectiveness of our approach.
Figures & tables
Figure 1: Overview of GenRec (right) vs. prior generative reconstruction methods (left). Prior methods condition multi-view image or video models on warped point clouds or Plücker rays, supervising output pixels with one generative loss regardless of source coverage. GenRec factors the task by an observation mask: a generative backbone predicts every pixel, and a pixel-space reconstruction module refines the observed ones under a dedicated loss. The two pathways share neither parameters nor gradients, preserving the generative prior in unobserved regions.
Figure 2: Overview of GenRec . Given posed input views and target poses, monocular depth and forward warping produce per-target observation masks and warped renders. These condition a (I) multi-view flow matching backbone , which jointly denoises RGB and scene-coordinate latents through cross-view and cross-modal attention. A (II) reconstruction branch then refines only the observed pixels via sparse 3D attention guided by the predicted scene coordinates, while the observation mask gates gradients so the generative prior is preserved in unobserved regions.
Global
Observed
Unobserved
Method
PSNR ↑
SSIM ↑
LPIPS ↓
FID ↓
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
Inf. Time
CameraCtrl [ 13 ]
13.40
0.4424
0.4761
57.42
15.53
0.4568
12.40
0.3935
∼ 12s
ViewCrafter [ 56 ]
13.29
0.4792
0.4293
46.27
15.93
0.5196
11.83
0.3937
∼ 200s
SEVA [ 59 ]
13.50
0.4839
0.4674
41.64
15.97
0.4999
12.40
0.4189
∼ 20s
GF [ 47 ]
13.11
0.3806
0.4945
54.58
15.56
0.3986
11.96
0.3125
∼ 40s
LVSM [ 18 ]
13.81
0.4094
0.5186
57.89
16.32
0.4224
12.61
0.3435
∼ 1s
Table 1: Quantitative results on RealEstate10K in the single-view extrapolation setting. Global metrics are computed over the full target frame; Observed restricts them to pixels that receive source evidence under the shared observation mask, and Unobserved to pixels that do not. Inference time is reported per scene. Best results in bold .
Global
Observed
Unobserved
Method
PSNR ↑
SSIM ↑
LPIPS ↓
FID ↓
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
Inf. Time
CameraCtrl [ 13 ]
13.26
0.3346
0.4893
69.49
13.45
0.3418
12.76
0.2957
∼ 12s
ViewCrafter [ 56 ]
17.57
0.5302
0.3092
46.36
18.75
0.5659
15.36
0.4282
∼ 200s
SEVA [ 59 ]
14.58
0.4300
0.4359
43.51
14.86
0.4401
13.92
0.3819
∼ 20s
GLD [ 17 ]
15.54
0.4001
0.3979
51.85
16.10
0.4184
14.33
0.3368
∼ 42s
Gen3R [ 16 ]
13.74
0.3754
0.5173
130.10
14.02
0.3886
12.95
0.3048
∼ 90s
Table 2: Quantitative Results on DL3DV-10K. We compare our method with previous works on single-view conditioned generation. We follow the same evaluation protocol as in Table 1 .
Global
Observed
Unobserved
Method
PSNR ↑
SSIM ↑
LPIPS ↓
FID ↓
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
Inf. Time
CameraCtrl [ 13 ]
10.89
0.1870
0.6534
214.2
11.31
0.1869
10.67
0.1821
∼ 12s
ViewCrafter [ 56 ]
11.75
0.2197
0.5694
172.5
14.46
0.2510
10.74
0.1913
∼ 200s
SEVA [ 59 ]
11.52
0.2231
0.5857
139.8
12.09
0.2315
11.30
0.2171
∼ 20s
GLD [ 17 ]
12.03
0.1986
0.5631
162.1
13.32
0.2044
11.45
0.1900
∼ 42s
Gen3R [ 16 ]
11.07
0.2056
0.6564
223.6
11.21
0.2072
10.96
0.2015
∼ 90s
Table 3: Quantitative results on Mip-NeRF 360. Out-of-distribution comparison with prior methods in the single-view extrapolation setting. We follow the same evaluation protocol as in Table 1 .
Figure 3: Comparative qualitative results. Single-view extrapolation against our strongest baseline, Gen3C [ 34 ] . Each row shows the input view, ground-truth target, GenRec , and Gen3C.
RE10K
DL3DV-10K
Method
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
CameraCtrl [ 13 ]
12.87
0.4271
0.5022
15.60
0.3456
0.4705
ViewCrafter [ 56 ]
12.61
0.4550
0.4573
17.38
0.5214
0.3087
SEVA [ 59 ]
16.15
0.5833
0.3477
21.38
0.6906
0.1845
LVSM [ 18 ]
16.88
0.5493
0.3983
21.31
0.6801
0.2223
GLD [ 17 ]
17.64
0.5707
0.2900
19.49
0.5575
0.2376
Table 4: Quantitative results on RE10K and DL3DV-10K. Comparison with prior methods in the two-view interpolation setting.
Global
Observed
Method
PSNR ↑
SSIM ↑
LPIPS ↓
FID ↓
PSNR ↑
SSIM ↑
RE10K
Generative Backbone
15.86
0.5483
0.3402
37.82
19.20
0.6139
+ VAE Skip & LoRA
15.86
0.5591
0.3394
35.27
19.23
0.6152
+ Reconstruction Branch
17.05
0.6071
0.3220
33.09
20.85
0.6738
DL3DV
Generative Backbone
18.62
0.5796
0.2556
39.94
19.96
0.6163
+ VAE Skip & LoRA
18.64
0.5828
0.2512
36.90
20.02
0.6215
Table 5: Ablation study on the effectiveness of individual components on RE10K and DL3DV. Starting from a vanilla multi-view flow matching baseline, we progressively add the VAE skip connections + LoRA weights, and the reconstruction branch.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Qualitative comparison on Mip-NeRF 360 [ 2 ] . Each row shows a held-out target view rendered from a single reference, compared against GEN3C. Mip-NeRF 360’s unbounded outdoor scenes and wide-baseline trajectories stress geometric consistency in the unobserved background; our model preserves thin structures and produces sharper far-field detail than the baselines.
Figure 5: Qualitative comparison on RealEstate10K [ 60 ] . Indoor real-estate walkthroughs feature predominantly forward camera motion and narrow baselines. We highlight cases with disocclusions of structured surfaces (walls, doorways, furniture edges); our method recovers plausible texture in unobserved regions while baselines exhibit blur or ghosting along occlusion boundaries.
Figure 6: Qualitative comparison on DL3DV-10K [ 23 ] . DL3DV-10K provides diverse, wide-baseline captures of indoor and outdoor scenes. The larger camera motions amplify the gap between methods that only inpaint warped pixels and methods that synthesize globally consistent geometry; our outputs remain sharp and view-consistent under aggressive target poses.
Global
Observed
Unobserved
Rec.
Gate
BP
PSNR ↑
SSIM ↑
LPIPS ↓
FID ↓
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
(a)
×
✓
–
16.51
0.5779
0.3459
36.84
19.88
0.6535
13.96
0.4667
(b)
✓
×
×
16.97
0.6020
0.3433
40.43
20.55
0.6724
14.28
0.4940
(c)
✓
✓
✓
16.99
0.6067
0.3429
34.92
20.77
0.6732
14.13
0.4738
(d)
✓
✓
×
17.05
0.6071
0.3220
33.09
20.85
0.6738
14.18
0.4743
Appendix
Table 6: Ablation of the generation–reconstruction split on RealEstate10K, single-view extrapolation. Rec. : dedicated pixel-space reconstruction branch; in (a) the branch is removed and the mask-gated regression loss is applied to the backbone directly. Gate : mask-gated supervision, i.e., flow-matching supervision on all pixels with an additional regression supervision on observed ones. BP : reconstruction gradients back-propagated into the backbone through the last K=1 sampling steps, following Clark et al. [6] . Rows (b)–(d) share the same architecture, parameter count, and training procedure. Best results in bold .
#Views
Method
PSNR ↑
SSIM ↑
LPIPS ↓
2
Gen3C [ 34 ]
15.70
0.5632
0.3282
GenRec
19.80
0.7043
0.2205
4
Gen3C [ 34 ]
20.80
0.7596
0.1860
GenRec (OOD)
23.68
0.7988
0.1387
6
Gen3C [ 34 ]
20.86
0.7656
0.1793
GenRec (OOD)
24.07
0.8061
0.1342
Appendix
Table 7: Multi-view interpolation on RE10K. GenRec is trained with at most two input views, so the 4- and 6-view results are zero-shot (marked OOD ), whereas Gen3C’s 3D cache ingests additional views natively. Best result per view count in bold .
RE10K
DL3DV-10K
Method
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
3DGS [ 20 ]
14.95
0.543
0.421
19.76
0.681
0.248
Difix3D+ [ 48 ]
15.06
0.543
0.387
19.27
0.655
0.250
pixelSplat [ 4 ]
16.96
0.614
0.392
19.29
0.625
0.328
MVSplat [ 5 ]
16.64
0.592
0.383
17.53
0.535
0.366
DepthSplat [ 52 ]
19.92
0.723
0.277
19.79
0.673
0.253
Appendix
Table 8: Comparison with reconstruction methods in the two-view interpolation setting on RE10K and DL3DV-10K. The baselines are specialised, deterministic reconstruction models; GenRec is generative. Best results in bold .
RE10K
DL3DV
Method
ATE ↓
RPE t ↓
RPE r ↓
ATE ↓
RPE t ↓
RPE r ↓
Inf. Time
GLD [ 17 ]
0.131
0.158
0.837
0.030
0.041
0.395
∼ 42s
ViewCrafter [ 56 ]
0.134
0.148
1.446
0.045
0.060
0.535
∼ 200s
Gen3R [ 16 ]
0.134
0.162
1.124
0.370
0.526
4.900
∼ 90s
Gen3C [ 34 ]
0.085
0.083
0.809
0.037
0.052
0.430
∼ 1200s
Ours
0.087
0.106
1.048
0.021
0.029
0.252
∼ 11s
Appendix
Table 9: Camera pose accuracy and multi-view consistency on RE10K and DL3DV-10K . ATE and RPE t are reported in the scale of the dataset camera poses; RPE r in degrees. Inference time is reported per scene. Lower is better for all metrics.
Figure 7: PSNR as a function of per-scene observation percentage on (a) RealEstate10K and (b) DL3DV-10K in the single-view conditioned setting. Each point is one test scene; lines are per-method linear fits. GenRec (red) sits above every baseline across the full range, from generation-heavy scenes on the left to near-reconstruction scenes on the right.
RE10K
DL3DV-10K
Method
PSNR ↑
SSIM ↑
LPIPS ↓
FID ↓
PSNR ↑
SSIM ↑
LPIPS ↓
FID ↓
RGB
15.09
0.5384
0.3602
38.22
18.05
0.5681
0.2631
38.53
RGB + SC (Ours)
15.39
0.5457
0.3528
37.13
18.31
0.5749
0.2575
37.41
Appendix
Table 10: Ablation on multi-modal generation. We compare our joint RGB and scene-coordinate (SC) backbone against an RGB-only variant trained from the same initialization for 20K steps, evaluated in the single-view conditioned extrapolation setting.
Global
Observed
Unobserved
Depth
PSNR ↑
SSIM ↑
LPIPS ↓
FID ↓
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
DA3 † [ 22 ]
17.35
0.6169
0.3206
63.69
22.34
0.6736
15.26
0.5242
MapAnything [ 19 ]
17.23
0.6151
0.3236
63.66
22.08
0.6717
15.15
0.5192
VGGT [ 45 ]
16.43
0.5734
0.3478
67.48
20.54
0.6092
14.79
0.4952
Appendix
Table 11: Depth source at inference on RealEstate10K, single-view extrapolation. GenRec is trained on DA3 renders ( † ); the estimator is swapped at test time with no retraining. Evaluated on the same RealEstate10K subset as Table 12 , so values are not comparable to Table 1 . Best results in bold .
Global
Observed
Unobserved
Corruption
Method
PSNR ↑
SSIM ↑
LPIPS ↓
FID ↓
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
Clean
GenRec
17.35
0.6196
0.3201
63.12
22.36
0.6759
15.25
0.5263
Gen3C [ 34 ]
15.44
0.5613
0.3627
67.04
20.04
0.6037
13.81
0.4893
ViewCrafter [ 56 ]
13.01
0.4749
0.4505
85.04
17.27
0.5222
11.43
0.3875
Affine
GenRec
15.08
0.5122
0.4101
67.08
18.55
0.5232
14.09
0.4691
Gen3C [ 34 ]
14.64
0.5240
0.4071
70.91
18.30
0.5404
13.56
0.4736
Appendix
Table 12: Depth-robustness sweep on RealEstate10K (25-scene subset, so clean values are not comparable to Tables 1 – 3 ). At inference, the source depth is corrupted after the Umeyama alignment that fixes scale; every method builds its own conditioning from the corrupted depth, and all runs are scored against a single frozen observation mask. Each entry averages over the severity levels of the corresponding family: three levels for the affine corruptions, one for boundary, and three for noise. Best result per block in bold .
Novel view synthesis from sparse images must reconcile faithful reconstruction of observed regions with plausible completion of unseen content, while maintaining world consistency across viewpoints. Existing geometry-based methods preserve observed scene structure but often struggle to complete unseen regions, whereas video generative models offer rich appearance priors but accumulate inconsistencies during sequential view generation. We propose GeoVerse, a framework that synthesizes world-consistent novel views by performing generation within the geometric latent space of a pretrained 3D foundation model and injecting appearance priors from a video generative model. Specifically, GeoVerse extracts multilevel features from Wan2.2 VACE and injects them into the geometric latent diffusion model via a ControlNet-style adapter, incorporating video-learned appearance priors to enhance structural completion. To enforce cross-view coherence, a global spatial memory continuously aggregates observed and synthesized content, reprojecting target-aligned guidance to anchor subsequent predictions to a shared scene representation. Extensive experiments across diverse datasets demonstrate improved visual quality and geometric consistency, with a 2.23 dB higher PSNR on DL3DV and 32.4% lower ATE on Mip-NeRF360 compared to GLD.
Kerui Ren, Tao Lu, Linning Xu +5
Shanghai Jiao Tong University · Shanghai Artificial Intelligence Laboratory · The Chinese University of Hong Kong +3
Current visual generation models are capable of producing high-quality content, yet they lack a coherent perception of the spatial structure. Existing generative novel view synthesis methods typically introduce explicit geometry priors, which enforce spatial consistency but inherently restrict generalization in large view changes. In contrast, recent interactive generative methods favor implicit scene modeling, offering greater flexibility at the cost of precise camera control and geometry consistency. In this paper, we propose MetaView, a diffusion-based monocular novel view synthesis framework that enables rendering under large view changes from a single image. Our key insight is to combine implicit geometry modeling with minimal yet essential explicit 3D cues: we incorporate implicit geometry priors from a feed-forward geometry perception network to regularize structure without imposing restrictive reconstruction pipelines, while leveraging metric depth to anchor the generation to a metric scale. This design allows MetaView to achieve both geometry consistency and precise controllability. Extensive experiments demonstrate that, under challenging monocular large viewpoint changes, MetaView significantly outperforms existing methods and exhibits superior generalization. Our code is publicly available at https://github.com/KlingAIResearch/MetaView.
Yufei Cai, Xuesong Niu, Hao Lu +3
Nanyang Technological University · Kolors Team, Kuaishou Technology · The Hong Kong University of Science and Technology (Guangzhou)
We propose Ref-GeNVS, a training-free, reflection-aware method for generative novel view synthesis (NVS) in mirror scenes. Existing multi-view diffusion models often fail to recognize the mirror in the scene and cannot exploit reflected content for scene generation. To fix this issue without additional training, our key idea is to treat a mirror image as two complementary views. From input images, we estimate the mirror plane and reflect camera poses to form virtual views. Based on this virtual view setup, we propose a two-stage generation method consisting of Mirror-gated attention and Reflection injection, which enables reflection-consistent NVS by explicitly leveraging reflection relationships in a multi-view diffusion model. Ref-GeNVS inherits the strong generalizability of the multi-view diffusion backbone, while it does not require finetuning. On synthetic and real scenes including mirrors, Ref-GeNVS outperforms recent generative NVS methods by generating reflection-consistent and contextually coherent novel views, revealing scene structure visible only through mirrors. Project page: https://kim-geonu.github.io/Ref-GeNVS/