Pith. sign in

REVIEW 3 major objections 5 minor 76 references

A single feed-forward pass can turn a short multi-camera driving log into a fully simulatable 3D Gaussian Splatting world — static background, dynamic actors, and sky — in about 1.5 seconds.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 02:49 UTC pith:HSL73QGX

load-bearing objection A real feed-forward reconstruction system with a genuine Waymo win, but the dynamic layer only translates Gaussians — orientation and normals stay frozen — so turning vehicles break the 'fully simulatable' claim. the 3 major comments →

arxiv 2607.14203 v1 pith:HSL73QGX submitted 2026-07-15 cs.GR cs.AIcs.CV

Instant NuRec: Feed-Forward 3D Gaussian Reconstruction for Driving Scene Simulation

NVIDIA: Jiahui Huang , Jiawei Ren , Michal Tyszkiewicz , Bjoern Haefner , Michael Shelley , Xin Kang , Seung Wook Kim , Ning Xu
show 7 more authors
Qi Wu Janick Martinez Esturo Shengyu Huang Nick Schneider Laura Leal-Taixe Zan Gojcic Sanja Fidler
This is my paper
classification cs.GR cs.AIcs.CV
keywords 3D Gaussian Splattingfeed-forward reconstructiondriving scene simulationclosed-loop policy evaluationlayered scene decompositionnovel-view synthesismulti-camera sequencepiecewise-linear trajectories
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to prove that per-scene optimization is not required to make a real driving log usable for closed-loop simulation: one feed-forward network pass produces a layered 3D Gaussian Splatting world with a static background, individually editable dynamic actors, a sky cubemap, and per-camera color corrections. The intended payoff is scale — a 10–20 second multi-camera clip becomes a navigable, re-timable scene in roughly 1.5 seconds instead of the roughly 75 minutes of per-scene tuning today, and on a public driving dataset the reconstruction beats the strongest feed-forward baseline by 2.01 dB PSNR. The second load-bearing claim is that closed-loop policy rankings are identical whether the scene comes from the fast feed-forward pass or from the expensive per-scene optimizer, so quick reconstructions can replace slow ones for policy evaluation. A sympathetic reader would care because this is what makes reconstruction-based simulation viable at the volume of data modern driving fleets generate.

Core claim

The central claim is that a complete, simulation-ready 3D world can be predicted directly from a short calibrated multi-view clip without any iterative per-scene optimization. The network outputs static Gaussians with surface normals and semantic labels, dynamic Gaussians whose positions are three-keyframe piecewise-linear trajectories, a distant-sky cubemap, and per-camera affine color transforms; the same representation plugs straight into a closed-loop simulator. The reported evidence includes a PSNR of 28.26 dB on a public driving benchmark (2.01 dB above the strongest evaluated feed-forward baseline), reconstruction in ~1.5 s versus ~75 min for per-scene optimization, and a 140-scene cl

What carries the argument

The load-bearing object is the layered 3DGS output. Gaussians are not predicted one per pixel; instead, query points lifted from a predicted depth map cross-attend to the shared encoder features, and a motion head outputs the two displacement endpoints that complete each dynamic actor's three-knot linear trajectory. The same query set feeds both the Gaussian head and the motion head, so static and dynamic layers share geometry while semantics decides which points are movable. Chunk merging with a frustum-ownership test then welds per-clip predictions into a single long scene without per-scene tuning.

Load-bearing premise

The dynamic layer models every moving object as moving linearly between three keyframes roughly half a second apart, with a hard fade at the edges; if a pedestrian or vehicle changes speed or direction within that interval — let alone articulates — the re-timed rollout is wrong, and the claimed policy-ranking match with per-scene optimization could break.

What would settle it

Take a set of driving clips in which pedestrians or cyclists accelerate, brake, or change direction within 0.5 s, reconstruct each with the feed-forward model and with the per-scene optimizer, and compare closed-loop policy rankings on those scenes; divergence in ranking, or a large drop in dynamic-region PSNR, would show the linear-trajectory assumption is the limiting factor.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Reconstruction of driving logs becomes a batch operation: any short multi-camera clip can be turned into a sim-ready world in seconds, making fleet-scale ingestion practical.
  • Closed-loop policy evaluation can use fast feed-forward reconstructions without changing policy rankings, at least for the evaluated policy set and scenes.
  • Non-pinhole camera models are supported natively, so fisheye and distorted lenses can be simulated without rectifying images.
  • A Selective query strategy cuts the Gaussian budget by roughly 3× with only marginal quality loss, offering a memory-versus-fidelity trade-off for downstream rendering or simulation.
  • Long clips (tens of seconds) are handled by chunk merging with frustum-ownership pruning, preventing far-field floaters without manual tuning.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the policy-ranking equivalence generalizes beyond the tested 140 scenes and five policies, fast reconstruction could be used as a cheap first-pass filter to select clips for expensive per-scene reconstruction, or as the primary reconstruction layer for large-scale data curation.
  • The three-keyframe linear-trajectory assumption suggests a concrete extension: predicting denser keyframes or a second-order motion model for articulated agents would extend the dynamic layer's fidelity; the paper lists this as future work.
  • The clean split between geometry heads trained with dense supervision and rendering heads trained with rendering loss is a training template that could transfer to other scene types — indoor, aerial, or off-road — where per-scene optimization currently dominates.
  • Because the motion supervision is derived from auto-labeled cuboid tracks, an untested assumption is that track errors do not systematically degrade the dynamic layer; quantifying this would strengthen the policy-ranking claim.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes Instant NuRec, a feed-forward transformer-based model that takes multi-view posed images from a driving rig and outputs a layered 3D Gaussian scene: a static background layer, a dynamic actor layer with piecewise-linear trajectories, a sky cubemap, and per-camera ISP transforms. The central claims are that a single forward pass turns a short multi-view driving log into a fully simulatable 3DGS world in ~1.5 seconds, that the Waymo Open Dataset PSNR is 2.01 dB above the strongest evaluated baseline, and that closed-loop policy ranking under Instant NuRec is identical to that under the per-scene-optimized NuRec. Training uses ~40K internal clips with LiDAR depth and cuboid tracks, with a three-stage curriculum; an extension to LiDAR reconstruction is also presented.

Significance. If the claims hold, this is an important step toward fleet-scale neural simulation, amortizing per-scene optimization into inference while preserving a simulation-ready layered representation. The paper's concrete strengths are the external Waymo evaluation with a standard protocol, an internally consistent ablation table, downstream detection and closed-loop policy experiments, and a LiDAR branch with a scale-floor ablation. The code release and integration with NuRec/AlpaSim are additional assets. The significance is tempered by the substantial quality gap to per-scene NuRec and by the limited dynamic model, both of which affect the strength of the 'fully simulatable' and 'comparable' claims.

major comments (3)
  1. [3.1, 6] The dynamic layer is defined by replacing only the position μ_i with a three-keyframe piecewise-linear trajectory; the rotation quaternion q and world-space normal n are static attributes. For a rotating rigid actor (e.g., a turning vehicle), this translates Gaussians along a chordal path while keeping their source-time orientation and normals. Over a 0.5 s interval, a vehicle turning at 30°/s is misoriented by ~15°, corrupting both rendering and any physics query using normals. This is a more fundamental limitation than the acknowledged sub-second non-rigid motion in Section 6, and it conflicts with the 'fully simulatable' claim in the abstract. Please either animate orientation/normals along the trajectory or explicitly scope the claim to translational kinematics.
  2. [5.2, Fig. 7] The text states that Instant NuRec 'trail[s] only modestly' in appearance quality to NuRec. However, Fig. 7 reports PSNR 29.93 (Dense) vs 34.38 (NuRec), a 4.45 dB gap that corresponds to roughly 2× the mean squared error. That is not 'modest' by standard novel-view synthesis measures. Please provide additional metrics (e.g., SSIM/LPIPS) and reframe the claim; as written, the discrepancy between the text and the figure undermines the credibility of the 'closes a large fraction of the quality gap' statement.
  3. [5.2, Fig. 8] The sentence 'the policy ranking under Instant NuRec is identical to that under NuRec' requires explicit numerical support. The bar chart appears to show large absolute increases in collision rates for Instant NuRec (e.g., 14% vs 37% for the A-1.5 4-camera policy), and the ordering by collision rate alone does not look identical across the two conditions. If the ranking is based on a composite score (e.g., collision + offroad), please specify the aggregation and provide the rank table. As presented, the evidence is insufficient to support a core claim of the closed-loop evaluation.
minor comments (5)
  1. [3.3] The values of the selective query stride s and cluster size p are only given as an example (s=2, p=8). Please report the actual hyperparameters used in the experiments.
  2. [4.2] Motion supervision is derived from auto-labeled cuboid tracks. Please discuss the sensitivity of trajectory quality to label noise, or cite evidence of the auto-labeler's accuracy, since the dynamic layer's fidelity inherits track errors.
  3. [5.4, Table 3] The LiDAR benchmark shows a 40% higher Chamfer distance and notably higher coverage error for Instant NuRec LiDAR. Calling the result 'qualitatively comparable' is a stretch; please quantify the coverage-precision trade-off or state the gap more precisely.
  4. [5.2] The closed-loop policy comparison reports 140 scenes × 6 trials but no confidence intervals or significance tests. Please report error bars or statistical tests to support the ranking claim.
  5. [1] Typo: 'Instant NuRecemits' is missing a space. Minor formatting issues also appear in §4.1 equation alignment.

Circularity Check

0 steps flagged

No circularity: the central claims are empirical measurements on held-out data against external baselines.

full rationale

Instant NuRec's headline results are empirical rather than derived: Table 1 reports PSNR/SSIM/depth on the Waymo Open Dataset validation split against external baselines with released checkpoints (DepthSplat, STORM, Depth-Anything-3, DGGT), and the reported 2.01 dB margin is the difference between the paper's reported PSNR (28.26) and the strongest baseline DGGT (26.25). This is self-contained evidence on held-out target frames, not a quantity defined by the model's own training objective. The closed-loop claim that 'the policy ranking under Instant NuRec is identical to that under NuRec' is also a measured outcome of running five policies through AlpaSim over 140 scenes with six trials each; NuRec and AlpaSim are NVIDIA self-citations, but they function as the experimental apparatus/baseline, not as an authoritative theorem that forces the outcome. The dynamic-layer simplification (three-keyframe piecewise-linear positions) is an explicit modeling choice, and Section 6 acknowledges it 'cannot capture sub-second non-rigid motion, such as pedestrian articulation.' The skeptic's concern that rotating rigid actors inherit static orientations/normals is a legitimate limitation bearing on the word 'fully simulatable,' but it is not a circular reduction: the representation is not defined in terms of the simulation outcome, nor is the claim equivalent to its inputs. The only citation oddity is Section 4.2 attributing the semantic segmentation auto-labeler to reference [1] (NuRec), which is a reference-quality issue rather than a load-bearing circular step. No fitted parameter is renamed as a prediction, and no uniqueness or ansatz is imported through self-citation to force the central result. Therefore no significant circularity is present; the honest finding is score 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

No physics is derived; the central claims are empirical. The listed free parameters are hand-set constants in the method. The trained network weights are fitted to data but are not enumerated individually. No new physical entities are postulated; the layered static/dynamic/sky/ISP outputs are algorithmic representations, not unobserved ontology.

free parameters (5)
  • Opacity pruning threshold = 0.01
    Section 3.4: Gaussians with alpha < 1e-2 are dropped; hand-set constant controlling primitive count and quality.
  • Frustum-ownership tolerance delta = not specified ('small')
    Section 3.4, Eq. (1): boundary tolerance in chunk-merging; no value or sensitivity study given.
  • Dynamic fade exponent = 10
    Section 3.1: fade f(x)=exp(-x^10) gates opacity near the first and third trajectory knots.
  • Selective query stride s and cluster size p = s=2, p=8 in example
    Section 3.3: strided query sampling controls the Gaussian budget (~4x reduction from Dense); trade-off characterized empirically but s and p chosen by hand.
  • Trajectory knot count = 3
    Section 3.1: dynamic positions are three knots with linear interpolation; fixed representation choice that limits sub-second motion fidelity.
axioms (5)
  • domain assumption Input camera poses and intrinsics are accurate enough to lift depth to world space.
    Section 3.1 takes 6-DoF pose T and intrinsics kappa as input; if these are noisy, all world-space Gaussians and trajectories inherit the error.
  • domain assumption LiDAR-derived depth maps and auto-labeled cuboid tracks are reliable supervision.
    Section 4.2: depth ground truth is accumulated from LiDAR and dynamic actors are supervised via an internal cuboid auto-labeler; label errors propagate to geometry and motion.
  • domain assumption The off-the-shelf semantic segmentation model correctly separates movable objects, sky, and road.
    Section 4.2: semantic labels are generated by applying a segmentation model; these labels determine which Gaussians become dynamic.
  • domain assumption A finite set of 3D Gaussians can represent the scene well enough for closed-loop simulation.
    Section 3.1 output representation assumes 3DGS expressivity; the paper's own limitations say a small budget under-samples thin structures.
  • domain assumption Policy rankings under NuRec approximate rankings under real deployment.
    Section 5.2 validates Instant NuRec by matching NuRec's rankings; this assumes NuRec itself is a faithful simulator for policy selection.

pith-pipeline@v1.3.0-alltime-deepseek · 17460 in / 16381 out tokens · 143219 ms · 2026-08-02T02:49:33.734818+00:00 · methodology

0 comments
read the original abstract

3D simulation platforms are critical for autonomous driving because they enable end-to-end policy evaluation, thereby reducing development costs and improving safety. In recent years, neural simulation has become predominant, with methods such as NuRec playing a central role; however, these methods remain relatively slow and typically require per-scene tuning. In this work, we present Instant NuRec, a feed-forward neural reconstruction model that turns a short multi-view driving log into a fully simulatable 3D Gaussian Splatting (3DGS) world in a single forward pass. The model accepts multi-view input from a calibrated camera rig and emits a layered output consisting of static and dynamic 3DGS layers, a sky cubemap, and per-camera ISP corrections, while providing native support for non-pinhole camera models via 3DGUT. It reconstructs a 10-20-second multi-camera scene in roughly 1.5 seconds and achieves a PSNR on the Waymo Open Dataset that is 2.01 dB above the strongest evaluated baseline. Instant NuRec is deeply integrated into NuRec and is compatible with AlpaSim for closed-loop simulation.

Figures

Figures reproduced from arXiv: 2607.14203 by Bjoern Haefner, Janick Martinez Esturo, Jiawei Ren, Laura Leal-Taixe, Michael Shelley, Michal Tyszkiewicz, Nick Schneider, Ning Xu, NVIDIA: Jiahui Huang, Qi Wu, Sanja Fidler, Seung Wook Kim, Shengyu Huang, Xin Kang, Zan Gojcic.

Figure 1
Figure 1. Figure 1: Instant NuRec ingests a multi-camera driving sequence and, in a single feed-forward pass, produces a layered 3DGS world with static, dynamic, and sky components that the user can freely navigate spatially and temporally. For each scene (top and bottom), the location markers indicate the vehicle positions used to produce the renderings in the right columns. Abstract 3D simulation platforms are critical for … view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline overview. Multi-view driving images are tokenized into patches and processed by an alternating-attention ViT encoder. Several decoder heads share the resulting latent features and produce depth maps, semantic labels, motion estimates, a sky cubemap, and 3DGS attributes. Optionally, the output can be further optimized on a per-scene basis and used for downstream simulation tasks. 3.1. Problem Formu… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on the Waymo Open Dataset. Novel-view synthesis results for held-out target frames from the validation set. Compared with recent feed-forward baselines, Instant NuRec produces sharper imagery and better-preserved thin structures [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Single-camera reconstruction. From a single front-camera input, Instant NuRec reconstructs a navigable 3DGS scene in one forward pass. The re-posed and top-down renderings remain geometrically consistent and preserve thin structures such as poles and traffic lights. Closed-loop simulation viability. A key question for any feed-forward reconstruction method is whether the reconstruction produced by a single… view at source ↗
Figure 5
Figure 5. Figure 5: Multi-camera surround-view reconstruction. From a five-camera input rig, Instant NuRec jointly reconstructs the full surround view. Each row shows a different scene, and the rendered views remain consistent in appearance and geometry across overlapping cameras. Query Point: Dense (#GS = 8.25M) Query Point: Dense (#GS = 4.45M) Query Point: Selective (#GS = 3.07M) Query Point: Selective (#GS = 1.52M) [PITH_… view at source ↗
Figure 6
Figure 6. Figure 6: Query-point selection. Comparison of the Dense and Selective query-point strategies. The Selective strategy produces far fewer Gaussians (#GS) while maintaining comparable reconstruction quality. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Quantitative reconstruction comparison on internal data. Instant NuRec (Dense and Selective) versus the per-scene-optimized NuRec baseline on PSNR, downstream 3D detection precision and recall, and reconstruction time. Instant NuRec reaches comparable appearance and detection quality while reconstructing a scene in ∼1.5 s instead of ∼75 min. A-1.5 (4 cam) A-R1 A-1.5 (2 Cam) VaVAM A-1.5 (1 Cam) 1 2 3 4 5 Po… view at source ↗
Figure 8
Figure 8. Figure 8: Closed-loop comparison between NuRec and Instant NuRec across five policy configurations. Each bar pair compares the same policy under NuRec (blue) versus Instant NuRec (green), averaged over 140 scenes × 6 trials. Despite requiring only a single forward pass, Instant NuRec reproduces the same policy ranking as NuRec, validating it as a reliable simulator for policy comparison. 5.3. Ablation Study We ablat… view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative ablation. Effects of the sky cubemap branch, LPIPS loss, and frustum-ownership merging strategy. In each comparison, the top row omits the named component, while the bottom row shows the full model. geometry, as reflected in the quantitative depth metrics. Primitive merging strategy. We replace frustum-ownership merging with a naive concatenation of Gaussians without pruning. Without this pruni… view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparison of LiDAR reconstruction on internal data for recorded and right-shifted [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

76 extracted references · 36 linked inside Pith

  1. [1]

    NVIDIA. NuRec. Website, 2026. URLhttps://research.nvidia.com/labs/sil/nurec/. Accessed 2026-03-08. 1, 3, 8, 9, 14

    2026

  2. [2]

    3DGUT: Enabling distorted cameras and secondary rays in gaussian splatting.Conference on Computer Vision and Pattern Recognition (CVPR),

    Qi Wu, Janick Martinez Esturo, Ashkan Mirzaei, Nicolas Moenne-Loccoz, and Zan Gojcic. 3DGUT: Enabling distorted cameras and secondary rays in gaussian splatting.Conference on Computer Vision and Pattern Recognition (CVPR),

  3. [3]

    AlpaSim: A modular, lightweight, and data-driven research simulator for autonomous driving

    NVIDIA, Yulong Cao, Riccardo de Lutio, Sanja Fidler, Guillermo Garcia Cobo, Zan Gojcic, Maximilian Igl, Boris Ivanovic, Peter Karkus, Janick Martinez Esturo, Marco Pavone, Aaron Smith, Ellie Tanimura, Michal Tyszkiewicz, Michael Watson, Qi Wu, and Le Zhang. AlpaSim: A modular, lightweight, and data-driven research simulator for autonomous driving. Softwar...

    2025

  4. [4]

    3D Gaussian Splatting for real-time radiance field rendering.ACM Transactions on Graphics (SIGGRAPH), 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3D Gaussian Splatting for real-time radiance field rendering.ACM Transactions on Graphics (SIGGRAPH), 2023. 2

    2023

  5. [5]

    OmniRe: Omni urban scene reconstruction

    Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, Li Song, and Yue Wang. OmniRe: Omni urban scene reconstruction. InThe Thirteenth International Conference on Learning Representations, 2025. URLhttps://openreview.net/forum?id=11xgiMEI5o. 2, 3

    2025

  6. [6]

    An image is worth 16x16 words: Transformers for image recognition at scale.ICLR, 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale.ICLR, 2021. 2

    2021

  7. [7]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. InEuropean Conference on Computer Vision (ECCV),

  8. [8]

    Srinivasan, Jonathan T

    Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2

    2022

  9. [9]

    Srinivasan, Jonathan T

    Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, and Henrik Kretzschmar. Block-NeRF: Scalable large scene neural view synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2

    2022

  10. [10]

    Neural scene graphs for dynamic scenes

    Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2

    2021

  11. [11]

    EmerNeRF: Emergent spatial-temporal scene decomposition via self-supervision

    Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, and Yue Wang. EmerNeRF: Emergent spatial-temporal scene decomposition via self-supervision. In International Conference on Learning Representations (ICLR), 2024. 2

    2024

  12. [12]

    NeuRAD: Neural rendering for autonomous driving

    Adam Tonderski, Carl Lindström, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. NeuRAD: Neural rendering for autonomous driving. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2

    2024

  13. [13]

    DrivingGaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes

    Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. DrivingGaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3

    2024

  14. [14]

    Street Gaussians: Modeling dynamic urban scenes with gaussian splatting

    Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street Gaussians: Modeling dynamic urban scenes with gaussian splatting. InEuropean Conference on Computer Vision (ECCV), 2024. 3

    2024

  15. [15]

    Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering.arXiv preprint arXiv:2311.18561, 2023

    Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering.arXiv preprint arXiv:2311.18561, 2023. 3 16 Instant NuRec: Feed-Forward 3D Gaussian Reconstruction for Driving Scene Simulation

    Pith/arXiv arXiv 2023

  16. [16]

    3D Gaussian Ray Tracing: Fast tracing of particle scenes.ACM Transactions on Graphics (SIGGRAPH Asia), 2024

    Nicolas Moenne-Loccoz, Ashkan Mirzaei, Or Perel, Riccardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3D Gaussian Ray Tracing: Fast tracing of particle scenes.ACM Transactions on Graphics (SIGGRAPH Asia), 2024. 3

    2024

  17. [17]

    CARLA documentation: NVIDIA NuRec

    CARLA Team. CARLA documentation: NVIDIA NuRec. Website, 2026. URLhttps://carla.readthedocs.io/en/ latest/nvidia_nurec/. Accessed 2026-03-09. 3, 9

    2026

  18. [18]

    Zero-1-to-3: Zero-shot one image to 3D object.arXiv preprint arXiv:2303.11328, 2023

    Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3D object.arXiv preprint arXiv:2303.11328, 2023. 3

    Pith/arXiv arXiv 2023

  19. [19]

    MVDream: Multi-view diffusion for 3D generation.arXiv preprint arXiv:2308.16512, 2023

    Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3D generation.arXiv preprint arXiv:2308.16512, 2023. 3

    Pith/arXiv arXiv 2023

  20. [20]

    ImageDream: Image-prompt multi-view diffusion for 3D generation.arXiv preprint arXiv:2312.02201, 2023

    Peng Wang and Yichun Shi. ImageDream: Image-prompt multi-view diffusion for 3D generation.arXiv preprint arXiv:2312.02201, 2023. 3

    Pith/arXiv arXiv 2023

  21. [21]

    LGM: Large multi-view gaussian model for high-resolution 3D content creation.arXiv preprint arXiv:2402.05054, 2024

    Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large multi-view gaussian model for high-resolution 3D content creation.arXiv preprint arXiv:2402.05054, 2024. 3

    Pith/arXiv arXiv 2024

  22. [22]

    Structured 3D latents for scalable and versatile 3D generation.arXiv preprint arXiv:2412.01506, 2024

    Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3D latents for scalable and versatile 3D generation.arXiv preprint arXiv:2412.01506, 2024. 3

    Pith/arXiv arXiv 2024

  23. [23]

    Native and compact structured latents for 3D generation.arXiv preprint arXiv:2512.14692, 2025

    Jianfeng Xiang, Xiaoxue Chen, Sicheng Xu, Ruicheng Wang, Zelong Lv, Yu Deng, Hongyuan Zhu, Yue Dong, Hao Zhao, Nicholas Jing Yuan, and Jiaolong Yang. Native and compact structured latents for 3D generation.arXiv preprint arXiv:2512.14692, 2025. 3

    Pith/arXiv arXiv 2025

  24. [24]

    Hunyuan3D 2.0: Scaling diffusion models for high resolution textured 3D assets generation.arXiv preprint arXiv:2501.12202, 2025

    Zibo Zhao, Zeqiang Lai, Qingxiang Lin, et al. Hunyuan3D 2.0: Scaling diffusion models for high resolution textured 3D assets generation.arXiv preprint arXiv:2501.12202, 2025. 3

    Pith/arXiv arXiv 2025

  25. [25]

    SAM 3D: 3Dfy anything in images, 2025

    SAM 3D Team, Xingyu Chen, Fu-Jen Chu, Pierre Gleize, Kevin J Liang, Alexander Sax, Hao Tang, Weiyao Wang, Michelle Guo, Thibaut Hardin, Xiang Li, Aohan Lin, Jiawei Liu, Ziqi Ma, Anushka Sagar, Bowen Song, Xiaodong Wang, Jianing Yang, Bowen Zhang, Piotr Dollár, Georgia Gkioxari, Matt Feiszli, and Jitendra Malik. SAM 3D: 3Dfy anything in images, 2025. URLht...

    Pith/arXiv arXiv 2025

  26. [26]

    pixelSplat: 3D gaussian splats from image pairs for scalable generalizable 3D reconstruction

    David Charatan, Sizhe Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelSplat: 3D gaussian splats from image pairs for scalable generalizable 3D reconstruction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3

    2024

  27. [27]

    MVSplat: Efficient 3D gaussian splatting from sparse multi-view images

    Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. MVSplat: Efficient 3D gaussian splatting from sparse multi-view images. InEuropean Conference on Computer Vision (ECCV), 2024. 3

    2024

  28. [28]

    GS-LRM: Large reconstruction model for 3D gaussian splatting

    Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: Large reconstruction model for 3D gaussian splatting. InEuropean Conference on Computer Vision (ECCV), 2024. 3

    2024

  29. [29]

    DepthSplat: Connecting gaussian splatting and depth

    Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. DepthSplat: Connecting gaussian splatting and depth. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3, 8, 9

    2025

  30. [30]

    DUSt3R: Geometric 3D vision made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3D vision made easy. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3

    2024

  31. [31]

    No pose, no problem: Surprisingly simple 3D gaussian splats from sparse unposed images.arXiv preprint arXiv:2410.24207, 2024

    Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3D gaussian splats from sparse unposed images.arXiv preprint arXiv:2410.24207, 2024. 3

    Pith/arXiv arXiv 2024

  32. [32]

    AnySplat: Feed-forward 3D gaussian splatting from unconstrained views.arXiv preprint arXiv:2505.23716, 2025

    Lihan Jiang, Yucheng Mao, Linning Xu, Tao Lu, Kerui Ren, Yichen Jin, Xudong Xu, Mulin Yu, Jiangmiao Pang, Feng Zhao, Dahua Lin, and Bo Dai. AnySplat: Feed-forward 3D gaussian splatting from unconstrained views.arXiv preprint arXiv:2505.23716, 2025. 3 17 Instant NuRec: Feed-Forward 3D Gaussian Reconstruction for Driving Scene Simulation

    arXiv 2025

  33. [33]

    STORM: Spatio-temporal reconstruction model for large-scale outdoor scenes

    Jiawei Yang, Jiahui Huang, Boris Ivanovic, Yuxiao Chen, Yan Wang, Boyi Li, Yurong You, Apoorva Sharma, Maximilian Igl, Peter Karkus, et al. STORM: Spatio-temporal reconstruction model for large-scale outdoor scenes. InInternational Conference on Learning Representations, volume 2025, pages 50446–50465, 2025. 3, 5, 8, 9, 12

    2025

  34. [34]

    DrivingForward: Feed-forward 3D gaussian splatting for driving scene reconstruction from flexible surround-view input

    Qijian Tian, Xin Tan, Yuan Xie, and Lizhuang Ma. DrivingForward: Feed-forward 3D gaussian splatting for driving scene reconstruction from flexible surround-view input. InProceedings of the AAAI Conference on Artificial Intelligence,

  35. [35]

    DGGT: Feedforward 4D reconstruction of dynamic driving scenes using unposed images.arXiv preprint arXiv:2512.03004, 2025

    Xiaoxue Chen, Ziyi Xiong, Yuantao Chen, Gen Li, Nan Wang, Hongcheng Luo, Long Chen, Haiyang Sun, Bing Wang, Guang Chen, Hangjun Ye, Hongyang Li, Ya-Qin Zhang, and Hao Zhao. DGGT: Feedforward 4D reconstruction of dynamic driving scenes using unposed images.arXiv preprint arXiv:2512.03004, 2025. 3, 8, 9

    arXiv 2025

  36. [36]

    ReconDrive: Fast feed-forward 4D gaussian splatting for autonomous driving scene reconstruction

    Haibao Yu, Kuntao Xiao, Jiahang Wang, Ruiyang Hao, Yuxin Huang, Guoran Hu, Haifang Qin, Bowen Jing, Yuntian Bo, and Ping Luo. ReconDrive: Fast feed-forward 4D gaussian splatting for autonomous driving scene reconstruction. arXiv preprint arXiv:2603.07552, 2026. 3

    arXiv 2026

  37. [37]

    StreetForward: Perceiving dynamic street with feedforward causal attention.arXiv preprint arXiv:2603.19552, 2026

    Zhongrui Yu, Zhao Wang, Yijia Xie, Yida Wang, Xueyang Zhang, Yifei Zhan, and Kun Zhan. StreetForward: Perceiving dynamic street with feedforward causal attention.arXiv preprint arXiv:2603.19552, 2026. 3, 8

    arXiv 2026

  38. [38]

    Ground4D: Spatially-grounded feedforward 4D reconstruction for unstructured off-road scenes.arXiv preprint arXiv:2605.04435, 2026

    Shuo Wang, Jilin Mei, Fuyang Liu, Wenfei Guan, Fanjie Kong, Zhihua Zhao, Shuai Wang, Chen Min, and Yu Hu. Ground4D: Spatially-grounded feedforward 4D reconstruction for unstructured off-road scenes.arXiv preprint arXiv:2605.04435, 2026. 3

    Pith/arXiv arXiv 2026

  39. [39]

    TokenGS: Decoupling 3D gaussian prediction from pixels with learnable tokens

    Jiawei Ren, Michal Tyszkiewicz, Jiahui Huang, and Zan Gojcic. TokenGS: Decoupling 3D gaussian prediction from pixels with learnable tokens. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026. 3

    2026

  40. [40]

    UniQueR: Unified query-based feedforward 3D reconstruction.arXiv preprint arXiv:2603.22851, 2026

    Chensheng Peng, Quentin Herau, Jiezhi Yang, Yichen Xie, Yihan Hu, Wenzhao Zheng, Matthew Strong, Masayoshi Tomizuka, and Wei Zhan. UniQueR: Unified query-based feedforward 3D reconstruction.arXiv preprint arXiv:2603.22851, 2026. 3

    arXiv 2026

  41. [41]

    GlobalSplat: Efficient feed-forward 3D gaussian splatting via global scene tokens.arXiv preprint arXiv:2604.15284, 2026

    Roni Itkin, Noam Issachar, Yehonatan Keypur, Xingyu Chen, Anpei Chen, and Sagie Benaim. GlobalSplat: Efficient feed-forward 3D gaussian splatting via global scene tokens.arXiv preprint arXiv:2604.15284, 2026. 3

    Pith/arXiv arXiv 2026

  42. [42]

    GAIA-1: A generative world model for autonomous driving.arXiv preprint arXiv:2309.17080, 2023

    Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. GAIA-1: A generative world model for autonomous driving.arXiv preprint arXiv:2309.17080, 2023. 3

    Pith/arXiv arXiv 2023

  43. [43]

    DriveDreamer: Towards real-world-driven world models for autonomous driving.arXiv preprint arXiv:2309.09777, 2023

    Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. DriveDreamer: Towards real-world-driven world models for autonomous driving.arXiv preprint arXiv:2309.09777, 2023. 3

    Pith/arXiv arXiv 2023

  44. [44]

    MagicDrive: Street view generation with diverse 3D geometry control.arXiv preprint arXiv:2310.02601, 2023

    Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. MagicDrive: Street view generation with diverse 3D geometry control.arXiv preprint arXiv:2310.02601, 2023. 3

    Pith/arXiv arXiv 2023

  45. [45]

    Panacea: Panoramic and controllable video generation for autonomous driving.arXiv preprint arXiv:2311.16813, 2023

    Yuqing Wen, Yucheng Zhao, Yingfei Liu, Fan Jia, Yanhui Wang, Chong Luo, Chi Zhang, Tiancai Wang, Xiaoyan Sun, and Xiangyu Zhang. Panacea: Panoramic and controllable video generation for autonomous driving.arXiv preprint arXiv:2311.16813, 2023. 3

    Pith/arXiv arXiv 2023

  46. [46]

    Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving, 2023

    Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving, 2023. URLhttps://arxiv.org/abs/ 2311.17918. 3

    Pith/arXiv arXiv 2023

  47. [47]

    Vista: A generalizable driving world model with high fidelity and versatile controllability.Advances in Neural Information Processing Systems (NeurIPS), 2024

    Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability.Advances in Neural Information Processing Systems (NeurIPS), 2024. 3

    2024

  48. [48]

    DriveDreamer- 2: LLM-enhanced world models for diverse driving video generation.arXiv preprint arXiv:2403.06845, 2024

    Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Xinze Chen, Guan Huang, Xiaoyi Bao, and Xingang Wang. DriveDreamer- 2: LLM-enhanced world models for diverse driving video generation.arXiv preprint arXiv:2403.06845, 2024. 3 18 Instant NuRec: Feed-Forward 3D Gaussian Reconstruction for Driving Scene Simulation

    Pith/arXiv arXiv 2024

  49. [49]

    GAIA-2: A controllable multi-view generative world model for autonomous driving, 2025

    Lloyd Russell, Anthony Hu, Lorenzo Bertoni, George Fedoseev, Jamie Shotton, Elahe Arani, and Gianluca Corrado. GAIA-2: A controllable multi-view generative world model for autonomous driving, 2025. URLhttps://arxiv.org/ abs/2503.20523. 3

    Pith/arXiv arXiv 2025

  50. [50]

    GEM: A generalizable ego-vision multimodal world model for fine-grained ego-motion, object dynamics, and scene composition control

    Mariam Hassan, Sebastian Stapf, Ahmad Rahimi, Pedro M B Rezende, Yasaman Haghighi, David Brüggemann, Isinsu Katircioglu, Lin Zhang, Xiaoran Chen, Suman Saha, Marco Cannici, Elie Aljalbout, Botao Ye, Xi Wang, Aram Davtyan, Mathieu Salzmann, Davide Scaramuzza, Marc Pollefeys, Paolo Favaro, and Alexandre Alahi. GEM: A generalizable ego-vision multimodal worl...

    Pith/arXiv arXiv 2024

  51. [51]

    DrivingWorld: Constructing world model for autonomous driving via video GPT.arXiv preprint arXiv:2412.19505,

    Xiaotao Hu, Wei Yin, Mingkai Jia, Junyuan Deng, Xiaoyang Guo, Qian Zhang, Xiaoxiao Long, and Ping Tan. DrivingWorld: Constructing world model for autonomous driving via video GPT.arXiv preprint arXiv:2412.19505,

    Pith/arXiv arXiv

  52. [52]

    Epona: Autoregressive diffusion world model for autonomous driving

    Kaiwen Zhang, Zhenyu Tang, Xiaotao Hu, Xingang Pan, Xiaoyang Guo, Yuan Liu, Jingwei Huang, Li Yuan, Qian Zhang, Xiao-Xiao Long, Xun Cao, and Wei Yin. Epona: Autoregressive diffusion world model for autonomous driving. arXiv preprint arXiv:2506.24113, 2025. 3

    Pith/arXiv arXiv 2025

  53. [53]

    Cosmos world foundation model platform for physical AI.arXiv preprint arXiv:2501.03575, 2025

    NVIDIA, Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, et al. Cosmos world foundation model platform for physical AI.arXiv preprint arXiv:2501.03575, 2025. 3

    Pith/arXiv arXiv 2025

  54. [54]

    Cosmos-Drive-Dreams: Scalable synthetic driving data generation with world foundation models.arXiv preprint arXiv:2506.09042, 2025

    Xuanchi Ren, Yifan Lu, Tianshi Cao, Ruiyuan Gao, Shengyu Huang, Amirmojtaba Sabour, Tianchang Shen, Tobias Pfaff, Jay Zhangjie Wu, Runjian Chen, Seung Wook Kim, Jun Gao, Laura Leal-Taixe, Mike Chen, Sanja Fidler, and Huan Ling. Cosmos-Drive-Dreams: Scalable synthetic driving data generation with world foundation models.arXiv preprint arXiv:2506.09042, 2025. 3

    Pith/arXiv arXiv 2025

  55. [55]

    DriveDreamer4D: World models are effective data machines for 4D driving scene representation.arXiv preprint arXiv:2410.13571, 2024

    Guosheng Zhao, Chaojun Ni, Xiaofeng Wang, Zheng Zhu, Xueyang Zhang, Yida Wang, Guan Huang, Xinze Chen, Boyuan Wang, Youyi Zhang, Wenjun Mei, and Xingang Wang. DriveDreamer4D: World models are effective data machines for 4D driving scene representation.arXiv preprint arXiv:2410.13571, 2024. 3

    Pith/arXiv arXiv 2024

  56. [56]

    DreamDrive: Generative 4D scene modeling from street view images.arXiv preprint arXiv:2501.00601,

    Jiageng Mao, Boyi Li, Boris Ivanovic, Yuxiao Chen, Yan Wang, Yurong You, Chaowei Xiao, Danfei Xu, Marco Pavone, and Yue Wang. DreamDrive: Generative 4D scene modeling from street view images.arXiv preprint arXiv:2501.00601,

    Pith/arXiv arXiv

  57. [57]

    ReconDreamer: Crafting world models for driving scene reconstruction via online restoration.arXiv preprint arXiv:2411.19548, 2024

    Chaojun Ni, Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Wenkang Qin, Guan Huang, Chen Liu, Yuyin Chen, Yida Wang, Xueyang Zhang, Yifei Zhan, Kun Zhan, Peng Jia, Xianpeng Lang, Xingang Wang, and Wenjun Mei. ReconDreamer: Crafting world models for driving scene reconstruction via online restoration.arXiv preprint arXiv:2411.19548, 2024. 3

    Pith/arXiv arXiv 2024

  58. [58]

    Difix3D+: Improving 3D reconstructions with single-step diffusion models.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    Jay Zhangjie Wu, Yuxuan Zhang, Haithem Turki, Xuanchi Ren, Jun Gao, Mike Zheng Shou, Sanja Fidler, Zan Gojcic, and Huan Ling. Difix3D+: Improving 3D reconstructions with single-step diffusion models.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3

    2025

  59. [59]

    DiffusionHarmonizer: Bridging neural reconstruction and photorealistic simulation with online diffusion enhancer.arXiv preprint arXiv:2602.24096, 2026

    YuxuanZhang, KatarínaTóthová, ZianWang, KangxueYin, HaithemTurki, RiccardodeLutio, Yen-YuChang, OrLitany, Sanja Fidler, and Zan Gojcic. DiffusionHarmonizer: Bridging neural reconstruction and photorealistic simulation with online diffusion enhancer.arXiv preprint arXiv:2602.24096, 2026. URLhttps://arxiv.org/abs/2602.24096. 3

    arXiv 2026

  60. [60]

    StreetCrafter: Street view synthesis with controllable video diffusion models.arXiv preprint arXiv:2412.13188, 2024

    Yunzhi Yan, Zhen Xu, Haotong Lin, Haian Jin, Haoyu Guo, Yida Wang, Kun Zhan, Xianpeng Lang, Hujun Bao, Xiaowei Zhou, and Sida Peng. StreetCrafter: Street view synthesis with controllable video diffusion models.arXiv preprint arXiv:2412.13188, 2024. 3

    Pith/arXiv arXiv 2024

  61. [61]

    FreeVS: Generative view synthesis on free driving trajectory.arXiv preprint arXiv:2410.18079, 2024

    Qitai Wang, Lue Fan, Yuqi Wang, Yuntao Chen, and Zhaoxiang Zhang. FreeVS: Generative view synthesis on free driving trajectory.arXiv preprint arXiv:2410.18079, 2024. 3

    Pith/arXiv arXiv 2024

  62. [62]

    Xiaomi EV World Model: A joint world model integrating reconstruction and generation for autonomous driving.arXiv preprint arXiv:2605.18137, 2026

    Lijun Zhou, Hongcheng Luo, Zhenxin Zhu, Cheng Chi, Mingfei Tu, Kaixin Xiong, Lei Gong, Zhanqian Wu, Zehan Zhang, Fangzhen Li, Hao Li, Yingying Shen, Jiale He, Haohui Zhu, Shan Zhao, Kai Wang, Zhiwei Zhan, Yuechuan Pu, Kaiyuan Tan, Ruiling Yang, Xianqi Wang, Tianyi Yan, Jiawei Zhou, Lei Zhang, Jingyang Zhao, Xi Zhou, Chitian Sun, Chenming Wu, Jiong Deng, H...

    Pith/arXiv arXiv 2026

  63. [63]

    Depth Anything 3: Recovering the visual space from any views.arXiv preprint arXiv:2511.10647, 2025

    Haotong Lin, Sili Chen, Junhao Liew, Donny Y Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang. Depth Anything 3: Recovering the visual space from any views.arXiv preprint arXiv:2511.10647, 2025. 4, 7, 8, 9

    Pith/arXiv arXiv 2025

  64. [64]

    DINOv2: Learning robust visual features without supervision

    MaximeOquab, TimothéeDarcet, ThéoMoutakanni, HuyVo, MarcSzafraniec, VasilKhalidov, PierreFernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4

    Pith/arXiv arXiv 2023

  65. [65]

    Vision transformers for dense prediction

    René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. InProceedings of the IEEE/CVF international conference on computer vision, pages 12179–12188, 2021. 5

    2021

  66. [66]

    V-DPM: 4D video reconstruction with dynamic point maps.arXiv preprint arXiv:2601.09499, 2026

    Edgar Sucar, Eldar Insafutdinov, Zihang Lai, and Andrea Vedaldi. V-DPM: 4D video reconstruction with dynamic point maps.arXiv preprint arXiv:2601.09499, 2026. 5

    arXiv 2026

  67. [67]

    4RC: 4D reconstruction via conditional querying anytime and anywhere.arXiv preprint arXiv:2602.10094, 2026

    Yihang Luo, Shangchen Zhou, Yushi Lan, Xingang Pan, and Chen Change Loy. 4RC: 4D reconstruction via conditional querying anytime and anywhere.arXiv preprint arXiv:2602.10094, 2026. 5

    Pith/arXiv arXiv 2026

  68. [68]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 5

    2023

  69. [69]

    Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the wild: Neural radiance fields for unconstrained photo collections. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6

    2021

  70. [70]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586–595, June 2018. doi: 10.1109/CVPR.2018.00068. 7

    arXiv 2018

  71. [71]

    Déjà View: Looping transformers for multi-view 3D reconstruction

    Alessandro Burzio, Tobias Fischer, Sven Elflein, Qunjie Zhou, Riccardo de Lutio, Jiawei Ren, Jiahui Huang, Shengyu Huang, Marc Pollefeys, Laura Leal-Taixé, et al. Déjà View: Looping transformers for multi-view 3D reconstruction. arXiv preprint arXiv:2605.30215, 2026. 7

    Pith/arXiv arXiv 2026

  72. [72]

    Scalability in perception for autonomous driving: Waymo Open Dataset

    Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception...

    2020

  73. [73]

    VaViM and VaVAM: Autonomous driving through video generative modeling.arXiv preprint arXiv:2502.15672, 2025

    Florent Bartoccioni, Elias Ramzi, Victor Besnier, Shashanka Venkataramanan, Tuan-Hung Vu, Yihong Xu, Loick Chambon, Spyros Gidaris, Serkan Odabas, David Hurych, Renaud Marlet, Alexandre Boulch, Mickael Chen, Éloi Zablocki, Andrei Bursuc, Eduardo Valle, and Matthieu Cord. VaViM and VaVAM: Autonomous driving through video generative modeling.arXiv preprint ...

    Pith/arXiv arXiv 2025

  74. [74]

    Alpamayo-R1: Bridging reasoning and action prediction for generalizable autonomous driving in the long tail.arXiv preprint arXiv:2511.00088, 2025

    NVIDIA, Yan Wang, Wenjie Luo, Junjie Bai, Yulong Cao, Tong Che, Ke Chen, Yuxiao Chen, Jenna Diamond, Yifan Ding, Wenhao Ding, Liang Feng, Greg Heinrich, Jack Huang, Peter Karkus, Boyi Li, Pinyi Li, Tsung-Yi Lin, Dongran Liu, Ming-Yu Liu, Langechuan Liu, Zhijian Liu, Jason Lu, Yunxiang Mao, Pavlo Molchanov, Lindsey Pavao, Zhenghao Peng, Mike Ranzinger, Ed ...

    Pith/arXiv arXiv 2025

  75. [75]

    Mip-Splatting: Alias-free 3D gaussian splatting

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-Splatting: Alias-free 3D gaussian splatting. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19447–19456,

  76. [2024]

    Contributors Research.Jiahui Huang, Jiawei Ren, Michal Tyszkiewicz, Xin Kang, Seung Wook Kim, Shengyu Huang, Laura Leal-Taixe‡, Zan Gojcic‡, Sanja Fidler‡

    13 20 Instant NuRec: Feed-Forward 3D Gaussian Reconstruction for Driving Scene Simulation A. Contributors Research.Jiahui Huang, Jiawei Ren, Michal Tyszkiewicz, Xin Kang, Seung Wook Kim, Shengyu Huang, Laura Leal-Taixe‡, Zan Gojcic‡, Sanja Fidler‡. Engineering.Bjoern Haefner, Michael Shelley, Ning Xu, Qi Wu, Janick Martinez Esturo, Nick Schneider‡. B. Ack...