Generative novel view synthesis (NVS) employs a diffusion model to render a static or dynamic scene from a target camera view. Explicit warp methods complete a reprojection of the source with video diffusion models, and implicit methods encode the camera parameters or 3D cues. Both add inference overhead, and the generated layout often drifts from the ground truth. We present WaveView, which injects the warped source into the phase of the initial noise of a pretrained generator. The Dual-Tree Complex Wavelet Packet Transform (DT-CWPT) allows local and anisotropic phase injection. The generator is thus constrained less in disoccluded regions and along the epipolar line. WaveView keeps the architecture and adds negligible overhead. On the DL3DV benchmark, it outperforms all baselines in every overlap bin. In the hardest bin, it gains 0.84 dB PSNR on the covisible region over MetaView, which shares its generator, while running 2.1 times faster. For dynamic NVS, fine-tuning AnyView with the same noise raises its zero-shot PSNR on AnyViewBench from 11.58 to 12.35 dB.
Figures & tables
Figure 2: Overview of WaveView. The depth warp x of the source (green) and Gaussian noise ϵ (blue) are decomposed with the DT- C WPT. At a visible location q , the positional uncertainty Σ(q) is stretched along the epipolar line, so bands whose frequency lies along this line take their phase from the noise, and bands keeping the warp’s phase (green dots) form an ellipse elongated across it. The magnitude of every complex band comes from the noise; holes ( v=0 ) are pure noise. Lower right of ϵ^ : the part whose phase comes from the warp.
Easy (overlap > 80%)
Medium (50–80%)
Hard (30–50%)
Whole image
Covisible region
Whole image
Covisible region
Whole image
Covisible region
Method
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Voyager [ 16 ]
17.91
0.514
0.425
18.60
0.502
0.412
14.16
0.400
0.521
15.57
0.409
0.483
12.08
0.348
0.596
13.33
0.350
0.556
PE-Field [ 1 ]
15.26
0.384
0.456
15.26
0.350
0.456
13.40
0.344
0.547
13.87
0.323
0.532
12.15
0.318
0.619
12.55
0.299
0.609
GEN3C [ 27 ]
18.90
0.603
0.301
19.64
0.603
0.279
14.45
0.414
0.452
15.68
0.429
0.409
12.61
0.324
0.546
13.60
0.325
0.508
ViewCrafter [ 50 ]
18.89
0.593
0.324
19.66
0.585
0.305
14.66
0.431
0.482
15.88
0.439
0.441
12.88
0.367
0.578
13.95
0.366
0.539
Table 1: Quantitative results on DL3DV . Covisible region: target pixels covisible with the source. Best in bold, second underlined.
GCD [ 34 ]
GEN3C [ 27 ]
CogNVS [ 8 ]
TrajCrafter [ 49 ]
AnyView [ 35 ]
WaveView (Ours)
Dataset
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
AnyViewBench: in-distribution
DROID [ 19 ]
10.18
0.255
0.688
9.62
0.211
0.666
9.44
0.281
0.634
10.45
0.257
0.620
14.42
0.443
0.473
14.92
0.460
0.453
Ego-Exo4D [ 11 ] ∗
12.10
0.255
0.670
11.69
0.222
0.643
10.80
0.241
0.670
11.33
0.195
0.642
17.46
0.493
0.410
16.02
0.429
0.456
LBM [ 35 ] ∗
12.59
0.398
0.694
13.48
0.449
0.581
13.30
0.453
0.548
13.68
0.447
0.537
17.63
0.619
0.386
18.17
0.644
0.371
Kubric-4D [ 34 , 12 ]
17.57
0.477
0.512
13.59
0.341
0.599
12.55
0.329
0.601
14.14
0.294
0.592
19.91
0.541
0.427
19.45
0.538
0.443
Table 2: Dynamic novel view synthesis on AnyViewBench [ 35 ] . Baseline numbers from [ 35 ] . Best in bold, second underlined. ∗ absent from our training data (see Sec. 5.1 ).
Figure 3: Qualitative comparison on DL3DV Hard. All methods use the same source depth. PE-Field and ViewCrafter tend to distort the geometry when the warp covers little of the target. GEN3C hallucinates the ceiling (first) and renders the pillar at a wrong position (fourth). MetaView shifts the table to the left (second), shrinks the round table so that its near edge enters the view (third), and drops the flowers in the lower left that the warp covers (last). WaveView keeps the layout of the warp and completes what the warp misses.
Figure 4: Qualitative comparison on the zero-shot split of AnyViewBench , first and last frame of each target video. GEN3C and TrajCrafter use the same source depth as ours. GEN3C stays at the source viewpoint (first and last example). TrajCrafter renders views that match neither camera, such as a close-up of a person (third). AnyView’s views differ from the ground truth in layout: it misplaces the traffic cones (first) and shifts the robot arm (second). WaveView stays closer to the layout of the ground truth.
Figure 5: Ablation on DL3DV Hard (covisible region). Isotropic: trained with Σ replaced by 21tr(Σ)I . Rotated 90∘ : our model with the depth covariance rotated at inference. Best baseline: best of all baselines per metric.
Figure 6: Variants of the phase injection versus ours on DL3DV Hard. Top: with the same seed and warp, the isotropic variant breaks the frame of the plaque (enlarged boxes). Bottom: the low-pass variant shifts the truss and the column off their ground-truth positions, while ours keeps them.
Projection-conditioned novel view synthesis (NVS) warps an explicit 3D reconstruction of the input view into the target camera and conditions a generator on the warped rendering. This works well for small viewpoint changes but degrades sharply under large orbital motion: the warp becomes sparse around the orbited object, where hidden surfaces dominate the new view and mirror-like artifacts emerge, causing the generator to lose both pixel content and the implicit camera cue carried by the warp. We introduce WarpHammer, a training-free framework that resolves this failure mode by augmenting the warped scene with an explicit 3D reconstruction of the object obtained from a native 3D generative prior (e.g., SAM3D). The reconstructed object adds missing foreground surfaces and occludes background points that should no longer be visible, restoring both appearance and camera cues without fine-tuning the base model. The same explicit object representation further unlocks a capability current NVS pipelines do not support: incorporating auxiliary views of the object from sources outside the target scene, for example, a casual snapshot of a car paired with a manufacturer studio shot of the same model. We process the reference and auxiliary images jointly with a pretrained multi-view geometry foundation model, which predicts a unified point cloud that we fuse into the 3D object reconstruction. This yields substantially more faithful geometry than single-image reconstruction, without requiring user-provided camera poses for the auxiliary views. On five benchmarks, WarpHammer produces stable novel views at viewpoint deviations where strong baselines collapse, and is the first scene-level NVS method that can naturally fuse auxiliary, pose-unknown object views from an external source.
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.
We tackle the problem of sparse novel view synthesis (NVS) using video diffusion models; given K (≈5) multi-view images of a scene and their camera poses, we predict the view from a target camera pose. Many prior approaches leverage generative image priors encoded via diffusion models. However, models trained on single images lack multi-view knowledge. We instead argue that video models already contain implicit multi-view knowledge and so should be easier to adapt for NVS. Our key insight is to formulate sparse NVS as a low frame-rate video completion task. However, one challenge is that sparse NVS is defined over an unordered set of inputs, often too sparse to admit a meaningful order, so the models should be invariant to permutations of that input set. To this end, we present FrameCrafter, which adapts video models (naturally trained with coherent frame orderings) to permutation-invariant NVS through several architectural modifications, including per-frame latent encodings and removal of temporal positional embeddings. Our results suggest that video models can be easily trained to "forget" about time with minimal supervision, producing competitive performance on sparse-view NVS benchmarks. Project page: https://frame-crafter.github.io/