Organizations: Shanghai Research Institute for Intelligent Autonomous Systems, Tongji University · Zaofu Intelligent Technology Co., Ltd. · College of Automotive and Energy Engineering, Tongji University · Institute of Automation, Chinese Academy of Sciences · SAI, Shanghai Jiao Tong University
Driving-scene reconstruction via feed-forward Gaussian splatting from sparse multi-view images remains challenging due to two factors. First, cross-camera photometric inconsistency, caused by variations in exposure, ISP, and illumination, can be incorrectly absorbed into Gaussian geometry. Second, unreliable local geometric supervision arises from normal-based constraints near depth discontinuities. To address these issues, we present MCSplat, a multi-view consistent feed-forward Gaussian splatting framework for driving scenes. MCSplat constructs a canonical appearance space from depth-guided multi-camera correspondences to decouple geometry learning from camera-specific appearance. A multi-scale Bilateral Grid Transformer restores raw camera appearance from canonical renderings, enabling photometric supervision without contaminating Gaussian geometry. To improve local geometric reliability, we estimate Fisher-information-based uncertainty from photometric sensitivity and use it to adaptively weight surface-oriented self-supervision. Extensive experiments demonstrate improved reconstruction quality and generalization across driving datasets. Geometry-oriented evaluations further reveal enhanced cross-view depth consistency and local surface-normal agreement, while qualitative results show fewer geometric artifacts and more coherent structures across overlapping views.
Figures & tables
Figure 2: Overview of MCSplat. (a) Depth-guided canonical appearance construction aligns inconsistent multi-view images. (b) A multi-scale bilateral-grid transformer restores camera-specific appearance from canonical renderings. (c) Fisher information-based uncertainty guides geometric self-supervision in under-constrained regions.
Figure 3: Visualization of photometric-loss uncertainty maps for Gaussian color, scale, and center parameters. The center-based map captures local structure around the white truck boundary and the jeep grille more distinctly than the scale map and the color map.
Method
Views
Waymo
nuScenes
Reconstruction
Novel View Synthesis
Recon&NVS
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
EvolSplat ∗ ( Miao et al., 2025 )
Front
23.35
0.70
0.29
–
–
–
–
–
–
UniSplat ∗ ( Shi et al., 2026 )
Front
28.93
0.86
0.18
27.34
0.80
0.20
–
–
–
Ours
Front
31.05
0.89
0.12
27.64
0.84
0.16
–
–
–
PixelSplat ∗ ( Charatan et al., 2024 )
Multi
–
–
–
–
–
–
21.51
0.616
0.372
Table 1: Quantitative comparison of image quality on the Waymo and nuScenes datasets. The best , second-best , and third-best results are highlighted. ∗ : Reported by UniSplat. 3 : Evaluated on the front 3 views only.
Table 4
Figure 4: Qualitative comparisons on the Waymo dataset. Compared with existing methods, MCSplat produces finer and cross-camera-consistent results, avoiding texture-induced depth ambiguity within the yellow road sign.
Figure 5: Qualitative results under cross-camera appearance inconsistency. Zoom in for best view. Compared with UniSplat (c), our canonical rendering removes camera-boundary artifacts (b), while achieving effective appearance recovery (d).
Method
Eval.
Depth Reproj. (m)
Reconstruction
Pseudo-Normal Consistency
Mean ↓
RMSE ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Mean ( ∘ ) ↓
<5∘ (%) ↑
<11.25∘ (%) ↑
<30∘ (%) ↑
Baseline
Raw
2.378
6.427
29.58
0.86
0.17
115.05
0.02
0.10
1.18
Baseline + CAC
Can.
0.967
4.051
31.06
0.88
0.13
97.25
0.17
0.74
4.08
Baseline + CAC + BGT
Raw
0.967
4.051
30.76
0.88
0.13
97.25
0.17
0.74
4.08
Baseline + CAC + Geo
Can.
0.910
3.920
31.09
0.88
0.14
92.82
4.24
10.08
17.38
Baseline + CAC + Geo
Raw
0.910
3.920
29.35
0.86
0.14
92.82
4.24
10.08
17.38
Table 4: Component ablation on the Waymo dataset. “Can.” and “Raw” denote canonicalized and raw GT images used for evaluation, respectively. Depth reprojection error is evaluated in adjacent-camera overlap regions, while normal consistency is measured against MoGe-2 ( Wang et al., 2025e ) pseudo-normals.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Intuitive pipeline comparison between our method and existing GFM-based feed-forward driving Gaussian frameworks.
Table 6: Quantitative reconstruction results on the Waymo Appearance-Inconsistent Subset.
Figure 7: Additional qualitative comparisons under severe cross-camera exposure inconsistency. Our method reduces exposure-induced cross-camera artifacts and alleviates texture-induced depth errors. Around the pickup truck, it avoids confusing the window depth with the background depth, yielding a more plausible depth estimate.
Figure 8: Additional qualitative comparisons. Our method effectively removes glass-surface artifacts in camera-boundary regions. Compared with LiDAR depth, it also yields improved depth estimation in pedestrian regions. On the planar poster surface, our method avoids texture-induced depth ambiguity, producing a more reasonable depth prediction.
Figure 9: Additional qualitative comparisons on the Waymo dataset. Our method reduces cross-camera stitching artifacts and produces more reasonable depth estimates, especially around road boundaries and tree or shrub boundaries.
Figure 10: Qualitative comparison of single-frame Gaussians under a novel view. Compared with the baseline, our method reconstructs cleaner road boundaries and sharper scene geometry. The highlighted regions show that the baseline suffers from geometry–appearance ambiguity caused by multi-view photometric inconsistency, leading to broken road edges, blurred boundary structures, and incomplete ground reconstruction.
Figure 11: Qualitative comparison of single-frame Gaussians under a novel view in urban scenes. The highlighted regions show that our method recovers dynamic pedestrians more faithfully than the baseline, while preserving clearer wall textures and more complete ground reconstruction.
Figure 12: Qualitative comparison of single-frame Gaussians under a novel view. UniSplat tends to confuse texture-rich regions with geometric variations, leading to blurred and distorted wall surfaces. In contrast, our method preserves sharper texture details and more consistent wall geometry in the highlighted regions.
Figure 13: Qualitative cross-lane novel-view synthesis under a 1.8 m leftward camera-trajectory shift. (a) Original trajectory; (b) shifted view rendered by UniSplat; (c) shifted view rendered by MCSplat. The highlighted regions show improved preservation of the left-foreground car and the parked motorcycles along the right sidewalk.
Figure 14: Qualitative cross-lane novel-view synthesis under a 2.0 m leftward camera-trajectory shift. (a) Original trajectory; (b) shifted view rendered by UniSplat; (c) shifted view rendered by MCSplat. Compared with UniSplat, MCSplat better preserves the highlighted pickup-truck rear structures and the red foreground car, although off-trajectory artifacts remain.
Feedforward 3D Gaussian Splatting (3DGS) often struggles in trajectory-based sparse-view driving scenes. Existing Gaussian repair methods mainly target optimization-based 3DGS, while diffusion-based repair is typically restricted to iterative refinement near observed viewpoints, leaving feedforward 3DGS repair underexplored. We propose ConFixGS, a plug-and-play method that learns to fix feedforward 3DGS with confidence-aware diffusion priors. Starting from a pretrained feedforward model, ConFixGS generates diffusion-enhanced local pseudo-targets and validates them through reprojection-based cross-checking against support views. The resulting dense confidence maps guide refinement, enhancing reliable details while suppressing hallucinated or inconsistent evidence. On Waymo, nuScenes, and KITTI, ConFixGS improves challenging novel view synthesis, with PSNR gains of up to 3.68 dB and FID reduced by nearly half. Our results highlight confidence-aware fusion of generative priors and support-view consistency as a key principle for robust feedforward 3D driving scene reconstruction.
Rui Song, Tianhui Cai, Markus Gross +5
University of California, Los Angeles · Technical University of Munich · University of Cambridge
Feed-forward 3D Gaussian Splatting methods have achieved impressive reconstruction quality for autonomous driving scenes, yet they entangle scene geometry with transient appearance properties such as lighting, weather, and time of day. This coupling prevents relighting, appearance transfer, and consistent rendering across multi-traversal data captured under varying environmental conditions. We present SpectralSplat, a method that disentangles appearance from geometry within a feed-forward Gaussian Splatting framework. Our key insight is to factor color prediction into an appearance-agnostic base stream and and appearance-conditioned adapted stream, both produced by a shared MLP conditioned on a global appearance embedding derived from DINOv2 features. To enforce disentanglement, we train with paired observations generated by a hybrid relighting pipeline that combines physics-based intrinsic decomposition with diffusion based generative refinement, and supervise with complementary consistency, reconstruction, cross-appearance, and base color losses. We further introduce an appearance-adaptable temporal history that stores appearance-agnostic features, enabling accumulated Gaussians to be re-rendered under arbitrary target appearances. Experiments demonstrate that SpectralSplat preserves the reconstruction quality of the underlying backbone while enabling controllable appearance transfer and temporally consistent relighting across driving sequences.
Quentin Herau, Tianshuo Xu, Depu Meng +5
Applied Intuition · The Hong Kong University of Science and Technology · University of California, Berkeley
High-fidelity reconstruction of driving scenes is crucial for autonomous driving. While recent feedforward 3D Gaussian Splatting (3DGS) methods enable fast reconstruction, their per-pixel Gaussian prediction paradigm often suffers from multi-view inconsistency and layering artifacts. Moreover, existing methods often model dynamic instances via dense flow prediction, which lacks explicit cross-view correspondence and instance-level consistency. In this paper, we propose PointForward, a feedforward driving reconstruction framework through point-aligned representations. Unlike pixel-aligned methods, we initialize sparse 3D queries in world space and aggregate multi-view image information via spatial-temporal fusion onto these queries, enforcing explicit cross-view consistency in a single feedforward pass. To handle scene dynamics, we introduce scene graphs that explicitly organize moving instances during reconstruction. By leveraging 3D bounding boxes, our method enables instance-level motion propagation and temporally consistent dynamic representations. Extensive experiments demonstrate that PointForward achieves state-of-the-art performance on large-scale driving benchmarks. The code will be available upon the publication of the paper.
Cheng Chi, Xianqi Wang, Hongcheng Luo +9
Xiaomi EV · Huazhong University of Science and Technology · Zhejiang University