Organizations: The University of Hong Kong · Tuojing Intelligence · King’s College London · The University of Sydney · Mohamed bin Zayed University of Artificial Intelligence
High-fidelity visual reconstruction and novel-view synthesis are essential for realistic closed-loop evaluation in autonomous driving. While 4D Gaussian Splatting (4DGS) offers a promising balance of accuracy and efficiency, existing per-scene optimization methods require costly iterative refinement, rendering them unscalable for extensive urban environments. Conversely, current feed-forward approaches often suffer from degraded photometric quality. To address these limitations, we propose ReconDrive, a feed-forward framework that leverages and extends the 3D foundation model VGGT for rapid, high-fidelity 4DGS generation. Our architecture introduces two core adaptations to tailor the foundation model to dynamic driving scenes: (1) Hybrid Gaussian Prediction Heads, which decouple the regression of spatial coordinates and appearance attributes to overcome the photometric deficiencies inherent in generalized foundation features; and (2) a Static-Dynamic 4D Composition strategy that explicitly captures temporal motion via velocity modeling to represent complex dynamic environments. Benchmarked on nuScenes, ReconDrive significantly outperforms existing feed-forward baselines in reconstruction, novel-view synthesis, and 3D perception. It achieves performance competitive with per-scene optimization while being orders of magnitude faster, providing a scalable and practical solution for realistic driving simulation.
Figures & tables
Figure 1: ReconDrive Inference Framework. ReconDrive is a powerful feed-forward 4D Gaussian splatting generation framework tailored for urban scene reconstruction and novel-view synthesis. In this framework, we select two context frames from each segment of an urban scene as input, and adopt a static-dynamic composition strategy to represent 4D Gaussians. To further enhance geometric precision, we design a dedicated Gaussian prediction head, consisting of a Gaussian Parameter Prediction Head (GPPH) and a Gaussian Center Prediction Head (GCPH), which enables the generation of Gaussians with more accurate geometric attributes.
Figure 2: Mapping Among Images, Masks, Dense Features, and Gaussian Indices. (a) Input image. (b) Dynamic object mask. (c) Dense features. (d) Gaussian kernels. Each pixel in the dense feature map generates a corresponding Gaussian kernel, enabling consistent pixel-wise mapping across all these components.
Figure 3: Visual Comparisons of Scene Reconstruction and Novel-View Synthesis. Compared with per-scene optimization methods (Street Gaussians, PVG, DeformableGS, OminiRe), our ReconDrive maintains high-quality visual rendering in both scene reconstruction (Original View) and novel-view synthesis (1–3 meters of lateral movement leftward and rightward). Compared with the feed-forward method DrivingForward, ReconDrive preserves accurate geometric consistency—evident in the tree and the surroundings—and the performance improvement becomes more pronounced as the moving distance increases. Additionally, it exhibits less image distortion and blurriness, particularly in image boundary regions. Extended visual comparisons are provided in the Appendix.
Method
Visual Scene Reconstruction
Novel-View Synthesis
Inference Speed ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
(Per Scene)
Per-Scene Optimization Methods
Street Gaussians Yan et al. (2024)
29.18
0.8824
0.1658
22.98
0.6959
0.2948
31min
PVG Chen et al. (2023)
29.58
0.8839
0.2200
23.48
0.6919
0.2897
23min
DeformableGS Yang et al. (2024)
28.93
0.8832
0.1610
23.73
0.6919
0.2342
46min
OminiRe Chen et al. (2025c)
29.42
0.8853
0.1577
23.01
0.6885
0.2762
35min
Table 1: Scene Reconstruction and Novel-View Synthesis Evaluation Results on the nuScenes Dataset.
Method
Detection
Tracking
mAP (%) ↑
AMOTA (%) ↑
Per-Scene Optimization Methods
Street Gaussians Yan et al. (2024)
14.6
11.9
PVG Chen et al. (2023)
18.5
14.4
DeformableGS Yang et al. (2024)
16.4
13.4
OminiRe Chen et al. (2025c)
16.1
12.9
Table 2: 3D Perception Results on Novel-View Rendered Scenes with 0 m, ±1 m, ±2 m, and ±3 m View Offsets.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Project Loss Computing Pipeline. Note that Kt and Ks are identical since the camera is fixed.
Figure 5: Distribution of Spatial Distances between the Point Maps Generated by the Original VGGT and Ground Truth LiDAR Point Clouds , highlighting significant metric misalignment.
Figure 6: Qualitative Comparison of VGGT-Generated Point Maps Across Various Scale Factors , demonstrating the structural consistency despite metric scale ambiguity.
Method
Novel-View Synthesis
PSNR ↑
SSIM ↑
LPIPS ↓
ReconDrive-S
23.54
0.6940
0.3177
ReconDrive
23.99 ( +0.45 )
0.7234 ( +0.0294 )
0.2591 ( -0.058 )
Appendix
Table 3: Evaluation Results with Different Frame Input.
Figure 7: Dynamic Gaussian Rendering Visualization. The rendered dynamic vehicles exhibit accurate motion and align closely with their counterparts in the ground truth images. Note that we have enhanced the exposure of the rendered images for visualization purposes to facilitate comparative observation.
Figure 8: Visual Comparisons of Scene Reconstruction and Novel-View Synthesis on Scene-0098 (from 14 Selected Validation Scenes). Consistent with prior results, our ReconDrive maintains high-quality rendering in both scene reconstruction (Original View) and novel-view synthesis (1–3m lateral movement) compared to per-scene optimization methods (Street Gaussians, PVG, DeformableGS, OminiRe).
Figure 9: Visual Comparisons of Scene Reconstruction and Novel-View Synthesis on Scene-0968 (from 14 Selected Validation Scenes). Consistent with prior results, our ReconDrive maintains high-quality rendering in both scene reconstruction (Original View) and novel-view synthesis (1–3m lateral movement) compared to per-scene optimization methods (Street Gaussians, PVG, DeformableGS, OminiRe). Against the feed-forward DrivingForward, it preserves precise visual consistency—clear in tree near the car —with more notable performance gains as movement increases, plus less distortion/blurriness, especially at image boundaries.
Figure 10: Visual Comparisons of Scene Reconstruction and Novel-View Synthesis on Scene-1065 (from 14 Selected Validation Scenes). Consistent with prior results, our ReconDrive maintains high-quality rendering in both scene reconstruction (Original View) and novel-view synthesis (1–3m lateral movement) compared to per-scene optimization methods (Street Gaussians, PVG, DeformableGS, OminiRe). Against the feed-forward DrivingForward (MF/SF modes), it preserves precise visual consistency. This visualization further demonstrates that ReconDrive performs exceptionally well in critical scenarios like nighttime environments.
Novel view synthesis (NVS) of static and dynamic urban scenes is essential for autonomous driving simulation, yet existing methods often struggle to balance reconstruction time with quality. While state-of-the-art neural radiance fields and 3D Gaussian Splatting approaches achieve photorealism, they often rely on time-consuming per-scene optimization. Conversely, emerging feed-forward methods frequently adopt per-pixel Gaussian representations, which lead to 3D inconsistencies when aggregating multi-view predictions in complex, dynamic environments. We propose EvolSplat4D, a feed-forward framework that moves beyond existing per-pixel paradigms by unifying volume-based and pixel-based Gaussian prediction across three specialized branches. For close-range static regions, we predict consistent geometry of 3D Gaussians over multiple frames directly from a 3D feature volume, complemented by a semantically-enhanced image-based rendering module for predicting their appearance. For dynamic actors, we utilize object-centric canonical spaces and a motion-adjusted rendering module to aggregate temporal features, ensuring stable 4D reconstruction despite noisy motion priors. Far-Field scenery is handled by an efficient per-pixel Gaussian branch to ensure full-scene coverage. Experimental results on the KITTI-360, KITTI, Waymo, and PandaSet datasets show that EvolSplat4D reconstructs both static and dynamic environments with superior accuracy and consistency, outperforming both per-scene optimization and state-of-the-art feed-forward baselines.
Sheng Miao, Sijin Li, Pan Wang +5
Zhejiang University, China. · Huawei, China. · University of Tübingen, Germany.
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
Dynamic scene reconstruction in autonomous driving remains a fundamental challenge due to significant temporal variations, moving objects, and complex scene dynamics. Existing feed-forward 3D models have demonstrated strong performance in static reconstruction but still struggle to capture dynamic motion. To address these limitations, we propose DynamicVGGT, a unified feed-forward framework that extends VGGT from static 3D perception to dynamic 4D reconstruction. Our goal is to model point motion within feed-forward 3D models in a dynamic and temporally coherent manner. To this end, we jointly predict the current and future point maps within a shared reference coordinate system, allowing the model to implicitly learn dynamic point representations through temporal correspondence. To efficiently capture temporal dependencies, we introduce a Motion-aware Temporal Attention (MTA) module that learns motion continuity. Furthermore, we design a Dynamic 3D Gaussian Splatting Head that explicitly models point motion by predicting Gaussian velocities using learnable motion tokens under scene flow supervision. It refines dynamic geometry through continuous 3D Gaussian optimization. Extensive experiments on autonomous driving datasets demonstrate that DynamicVGGT significantly outperforms existing methods in reconstruction accuracy, achieving robust feed-forward 4D dynamic scene reconstruction under complex driving scenarios.
Zhuolin He, Jing Li, Guanghao Li +9
Fudan University · Yinwang Intelligent Technology · Huawei +2