We present SyncFix, a framework that enforces cross-view consistency during the diffusion-based refinement of reconstructed scenes. SyncFix formulates refinement as a joint latent bridge matching problem, synchronizing distorted and clean representations across multiple views to fix the semantic and geometric inconsistencies. This means SyncFix learns a joint conditional over multiple views to enforce consistency throughout the denoising trajectory. Our training is done only on image pairs, but it generalizes naturally to an arbitrary number of views during inference. Moreover, reconstruction quality improves with additional views, with diminishing returns at higher view counts. Qualitative and quantitative results demonstrate that SyncFix consistently generates high-quality reconstructions and surpasses current state-of-the-art baselines, even in the absence of clean reference images. SyncFix achieves even higher fidelity when sparse references are available.
Figures & tables
Figure 2: SyncFix overview. Distorted renderings from multiple viewpoints xD are encoded into latent representations and transported toward clean targets xGT using latent bridge matching. SyncFix learns a joint latent bridge over multiple views, coupling latent trajectories through cross-view attention to enforce multi-view consistency during refinement. The model is trained using view pairs ( N=2 ) but generalizes to an arbitrary number of views at inference. Optional reference images can be provided to guide refinement. Here zD denotes distorted latents, zGT clean target latents, and zt=(1−t)zD+tzGT the bridge interpolation between them.
DL3DV
NeRFBusters
Method
PSNR ↑
LPIPS ↓
DSIM ↓
FID ↓
CVSC ↑
PSNR ↑
LPIPS ↓
DSIM ↓
FID ↓
CVSC ↑
3DGS [ 10 ]
15.94
0.454
0.297
80.8
0.875
11.85
0.575
0.336
139.4
0.951
Fixer [ 23 ]
15.73
0.466
0.257
45.45
0.822
12.32
0.588
0.296
126.7
0.817
Difix3D+ † [ 37 ]
16.05
0.368
0.171
26.2
0.819
11.94
0.528
0.202
97.4
0.925
Difix3D+ [ 37 ]
16.17
0.343
0.135
16.7
0.850
12.05
0.496
0.174
81.1
0.931
SyncFix †
16.45
0.334
0.129
22.5
0.862
14.07
0.517
0.183
86.2
0.937
Table 1: Quantitative comparison on the DL3DV and NeRFBusters test sets for refining sparse-view 3D Gaussian Splatting reconstruction renderings. SyncFix improves reconstruction quality and cross-view semantic consistency over prior generative refinement methods. † indicates no reference views during training. ↑ indicates higher-is-better and ↓ indicates lower-is-better. Bold denotes best results and underline denotes second best.
Figure 3: Multi-view consistency under extreme sparse-view degradation . The baseline 3DGS model (left) displays severe structural artifacts and floaters. While the marginal refinement method Difix3D+ (middle) removes these base artifacts, its independently generated views hallucinate inconsistent geometry across camera poses. As highlighted by the red boxes, Difix3D+ distorts the architectural structure of the outdoor bus shelter in the top scene. In the middle scene, it causes the object on the indoor pedestal to undergo shape changes between View 1 and View 2. Similarly, in the bottom scene, a door visible in one view is absorbed into the wall in another. In contrast, SyncFix (Ours) leverages joint latent synchronization to ensure that hallucinated high-frequency details and structural repairs remain geometrically and semantically consistent across views. Ours † is a model trained without clean reference images.
Figure 4: Performance at different number of 3DGS training views on the DL3DV dataset. SyncFix consistently achieves higher PSNR and lower DreamSim than Difix3D+ at different numbers of views, and the benefit of SyncFix is more pronounced at sparser settings.
Figure 5: Comparison on the NeRFBusters evaluation dataset. 3DGS reconstructions contain severe artifacts and geometric distortions. Difix3D+ improves the appearance of individual renderings but struggles to extrapolate to unseen regions and often produces unstable structures across views. In contrast, SyncFix generates plausible reconstructions that remain consistent across viewpoints, recovering coherent object geometry and scene structure. † are models trained without clean reference images.
Figure 6: Comparison of multi-view refinement with feedforward renderings . SyncFix provides coherent multi-view refinement that reduces the artifacts and maintains consistency. Notice the region highlighted in the boxes.
MipNeRF 360
Method
PSNR ↑
LPIPS ↓
DSIM ↓
CVSC ↑
AnySplat [ 9 ]
13.58
0.454
0.232
0.649
Difix3D+ [ 37 ]
13.41
0.448
0.196
0.589
SyncFix (ours)
13.51
0.430
0.189
0.686
Table 2: Quantitative comparison on the MipNeRF 360 dataset for refining feedforward Gaussian Splatting renderings. SyncFix improves perceptual quality and cross-view semantic consistency over prior generative refinement methods. We gray out the PSNR as pixel shifts make the pixel-aligned metric unreliable. ↑ indicates higher-is-better and ↓ indicates lower-is-better. Bold denotes best results and underline denotes second best.
Figure 8
Method
PSNR ↑
LPIPS ↓
DSSIM ↓
FID ↓
CVSC ↑
3DGS [ 10 ]
15.94
0.454
0.297
80.8
0.875
Difix3D [ 37 ]
16.14
0.445
0.285
76.6
0.853
Difix3D+ [ 37 ]
16.17
0.343
0.135
16.7
0.850
Difix3D+ (multi-view inference)
16.08
0.370
0.163
29.28
0.827
Ours (single-view, w/o reference)
16.26
0.347
0.142
22.2
0.821
Ours (multi-view, w/o reference)
16.45
0.334
0.129
22.5
0.862
Table 3: Ablation study on the DL3DV test set evaluating the effect of joint multi-view refinement. We also ablate with and without reference views. Simply applying a multi-view input to Difix3D+ provides limited benefit, while our joint training substantially improves both image quality and cross-view consistency. ↑ indicates higher-is-better and ↓ indicates lower-is-better.
Figure 8: Cross-view geometric consistency. Keypoint correspondences between two refined views. Difix3D+ produces inconsistent refinements, resulting in irregular correspondences across views. SyncFix preserves the underlying scene geometry, producing coherent cross-view matches. For visualization clarity, we display 20 sampled correspondences. Matches computed using RaCo + LightGlue.
Figure 9: Limitation . SyncFix may synthesize content that mismatches the ground-truth information. The resulting pixel-wise metrics are worse for this scene despite being multi-view consistent. Multi-view consistency can break down when views are refined in isolation. As seen in our SyncFix results, the windowsills in the bottom views become inconsistent because the independent network passes fail to synchronize geometric details across different perspectives. Note that the level of corruption for the top and bottom panels is different in the 3DGS views.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 10: CVSC scores as the number of degraded views and reference views changes . Joint refinement of more views leads to better multi-view consistency.
Figure 11: FID as the number of degraded views and reference views changes . Increasing the total number of input images increases the FID as attention is distributed between views.
Figure 12: PSNR as the number of degraded views and reference views changes . Adding more reference views helps with the local details, thus higher PSNR, and plateaus with more views.
Figure 13: Visual comparison on scene f102f93e... , Pair 1. The baseline 3DGS renderings exhibit strong blur and missing structure in the chair and surrounding objects. Fixer improves stability but suppresses fine details, while Difix3D+ produces sharper results with inconsistent textures across views. Our method reconstructs clearer chair geometry and preserves the object boundaries more faithfully, while maintaining consistent appearance across viewpoints.
Figure 14: Visual comparison on scene f102f93e... , Pair 2. The cropped regions emphasize the chair structure, where baseline methods either blur the geometry or introduce inconsistent textures. Our method restores sharper edges and produces a more coherent object appearance across the two views. Note that the corruption rates are different from Fig 6.
Figure 15: Visual comparison on scene e0a5470d... , Pair 1. In the doorway region, the baseline renderings show severe blur and inconsistent geometry. Fixer produces smoother results but removes structural details, while Difix3D+ partially restores textures with view-dependent artifacts. Our method recovers cleaner structural edges and maintains a consistent appearance across both viewpoints.
Figure 16: Visual comparison on scene e0a5470d... , Pair 2. The boxed crops highlight brick and railing structures that are difficult to reconstruct from sparse observations. Our method restores sharper geometric detail and more consistent textures compared with Fixer and Difix3D+. Note that the level of corruption is different in the 3DGS views compared to Fig 9.
Figure 17: Visual comparison on scene eba3bdcb... , Pair 1. The baseline 3DGS renderings contain strong blur and missing geometry around the wall fixtures. Fixer produces smoother results but loses structural detail, whereas Difix3D+ introduces artifacts around object boundaries. Our method reconstructs clearer object shapes and preserves the layout of switches and outlets more faithfully.
Figure 18: Visual comparison on scene eba3bdcb... , Pair 2. The selected crops illustrate improvements in structural clarity and artifact removal. Our method produces sharper boundaries and a more accurate layout of the wall fixtures while maintaining a consistent appearance across views. Note that the level of corruption is different in the 3DGS views compared to Fig 10.
Video diffusion models generate high-quality and diverse worlds; however, individual frames often lack 3D consistency across the output sequence, which makes the reconstruction of 3D worlds difficult. To this end, we propose a new method that handles these inconsistencies by non-rigidly aligning the video frames into a globally-consistent coordinate frame that produces sharp and detailed pointcloud reconstructions. First, a geometric foundation model lifts each frame into a pixel-wise 3D pointcloud, which contains unaligned surfaces due to these inconsistencies. We then propose a tailored non-rigid iterative frame-to-model ICP to obtain an initial alignment across all frames, followed by a global optimization that further sharpens the pointcloud. Finally, we leverage this pointcloud as initialization for 3D reconstruction and propose a novel inverse deformation rendering loss to create high quality and explorable 3D environments from inconsistent views. We demonstrate that our 3D scenes achieve higher quality than baselines, effectively turning video models into 3D-consistent world generators.
Per-scene optimization methods such as 3D Gaussian Splatting provide state-of-the-art novel view synthesis quality but extrapolate poorly to under-observed areas. Methods that leverage generative priors to correct artifacts in these areas hold promise but currently suffer from two shortcomings. The first is scalability, as existing methods use image diffusion models or bidirectional video models that are limited in the number of views they can generate in a single pass (and thus require a costly iterative distillation process for consistency). The second is quality itself, as generators used in prior work tend to produce outputs that are inconsistent with existing scene content and fail entirely in completely unobserved regions. To solve these, we propose a two-stage pipeline that leverages two key insights. First, we train a powerful bidirectional generative model with a novel opacity mixing strategy that encourages consistency with existing observations while retaining the model's ability to extrapolate novel content in unseen areas. Second, we distill it into a causal auto-regressive model that generates hundreds of frames in a single pass. This model can directly produce novel views or serve as pseudo-supervision to improve the underlying 3D representation in a simple and highly efficient manner. We evaluate our method extensively and demonstrate that it can generate plausible reconstructions in scenarios where existing approaches fail completely. When measured on commonly benchmarked datasets, we outperform all existing baselines by a wide margin, exceeding prior state-of-the-art methods by 1-3 dB PSNR.
Riccardo de Lutio, Tobias Fischer, Yen-Yu Chang +7
NVIDIA, USA · NVIDIA, Switzerland · ETHZ, Switzerland +4
Rendering views using 3D scene representations such as Gaussian Splatting (3DGS), Neural Radiance Fields (NeRF), meshes, or even point clouds produces artifacts when input views are sparse or target views lie far from the input. Recent work mitigates these artifacts using diffusion-based generative priors, but is specialized to individual representations and require custom architectures or extensive retraining. We present FixAnything, a single model for fixing a wide range of rendering artifacts. It does so by repurposing a pretrained video generative model, leveraging its implicit multi-view priors with only minimal modification and lightweight finetuning. Our key insight is that even noisily-rendered sequences preserve camera motion and coarse scene structure, allowing cleanup to be formulated as video-to-video translation. To control what scene structure should be preserved, we introduce a binary mask denoting the clean pixels, enabling the model to anchor its output to high-quality inputs (e.g. training views) while refining the rest. To encourage FixAnything to produce 3D-consistent renderings that support downstream reconstruction, we use camera pose accuracy (recovered via structure-from-motion) as a reward signal for direct preference optimization (DPO). Across four distinct 3D representations, FixAnything consistently improves rendering quality with lightweight finetuning, demonstrating that a single generalist video prior can replace multiple specialist refinement pipelines. The simplicity of the framework enables immediate adoption of stronger future video models without architectural redesign.