Pith. sign in

REVIEW 2 major objections 4 minor 72 references

muSync-GS: Physics-Synchronized Driving Video Synthesis for Weather and Geometric Road Hazards

T0 review · 2 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read muSync-GS couples weather and road edits with tire–road friction and vehicle response in driving-video synthesis.

desk verdict A well-built, unusually transparent physics-synchronized driving-video synthesis pipeline whose weather-to-friction link is honestly labeled as unvalidated; the CarSim evaluation validates the vehicle model given a shared friction input, not the precipitation-to-friction mapping. read the letter →

arxiv 2608.04412 v1 pith:AC6HHCTV submitted 2026-08-05 cs.CV

Yang Chen , Yicheng Zhu , Tao Li , Zilin Bian This is my paper
classification cs.CV
keywords drivingvideosynthesiscounterfactualdata3DGaussianrenderingadverseweathereditingtire-roadfrictionvehicledynamicsego-cameratrajectoryroadelevationhazards
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Driving videos edited to add rain, snow, speed humps, or depressions usually keep the original ego trajectory, even though real physics would change braking, wheel slip, load transfer, and camera motion. muSync-GS claims to close that gap by making the same scenario controls—precipitation rate, brake command, and a road-elevation profile—drive both the visible scene edit and a calibrated vehicle model. The model predicts speed, pitch, slip ratio, and per-wheel normal load, and those states build the rendered camera trajectory and frame-wise physical annotations. On held-out simulator cases the frozen model matches speed to 0.0273 m/s, pitch to 0.0590 degrees, slip to 0.0101, and normal load to 26.61 N, and it reproduces the early brake-dive trend on real-vehicle braking events. The paper therefore argues that physically synchronized counterfactual driving clips are feasible without collecting the rare events themselves.

What carries the argument

The load-bearing mechanism is a shared scenario state threaded through three interfaces. First, the precipitation-derived surface state $z_{\text{surf}} = (h_t, h_s, c_s)$ branches into a rendering control and a reference friction coefficient $\mu_c = \phi_\mu(w, z_{\text{surf}})$, so the same wetness that changes the picture also lowers tire–road friction. Second, a raised-cosine road-elevation profile $H_g(s)$ is used both to edit the visible 3D road geometry and to supply the height and slope at the front and rear tire contacts, coupling geometry to axle excitation. Third, the simulated path position $s(t)$ from the coupled longitudinal and half-car model resamples the reference route to produce the corrected camera pose $P(t) = P_0(s(t)) \oplus (z_s(t), \theta(t))$. The calibrated vehicle model is the component that turns visual edits into motion: it converts surface coefficient and elevation inputs into the forces and moments that set braking, slip, load transfer, and pitch.

What would settle it

Measure actual tire–road friction or stopping distance on instrumented asphalt under controlled rainfall and snowfall rates, and compare with the paper's friction coefficients and stopping-distance envelopes; deviations outside the declared parameter ranges, or a break in the monotonic severity ordering, would falsify the precipitation-to-friction branch while leaving the vehicle model itself intact.

Watch

Extended reading notes

Core claim

The central claim is that appearance, tire–road interaction, vehicle response, and ego-camera motion should be consequences of one scenario specification rather than separately prescribed. A precipitation-derived road-surface state determines road appearance and a reference friction coefficient; the same signed road-elevation profile that visibly raises or lowers the road also excites the axles; and a coupled longitudinal and four-degree-of-freedom half-car model converts these into speed, wheel slip, dynamic normal loads, heave, and pitch. The simulated path position then resamples the reference route, so the camera moves as the simulated vehicle moves. With all calibration parameters frozen, the framework reproduces speed, pitch, slip, and normal-load trajectories on held-out weather, brake, and road-profile cases, and after vehicle-specific calibration it tracks the measured pitch transient of real brake events. The discovery, stated as the author would state it, is that a shared surface state and a shared elevation profile are enough to keep what you see in a driving video synchronized with how the vehicle actually responds.

Load-bearing premise

The load-bearing premise is that the formula mapping precipitation rate to tire–road friction coefficient is correct for real roads; the held-out validation only checks that, once that friction value is supplied to both the simulator and the model, the vehicle response matches, so a wrong friction mapping would give wrong synchronized motion even with perfect simulator agreement.

Editorial extensions

If this is right

  • Generated counterfactual clips come with frame-wise physical telemetry (speed, slip, normal loads, pitch, friction, and brake pressure), so a training or validation pipeline receives synchronized labels instead of reconstructed ones.
  • Because the same controls set appearance and dynamics, weather severity can be swept continuously and stopping distance changes smoothly and monotonically, rather than switching between fixed weather effects.
  • Road-profile edits are no longer cosmetic: a hump or depression inserted into the scene produces the corresponding camera-pitch and load-transient response at the right time.
  • The frozen-parameter agreement on held-out controls suggests the calibrated model transfers to new brake pressures, onset triggers, and profile dimensions without refitting.
  • A physics-synchronized trajectory means the framework can generate brake-dive-accurate clips for rare events such as emergency braking on wet or snow-covered roads.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the precipitation-to-friction mapping is independently validated against instrumented wet- and snow-road measurements, the same shared-surface design could become a data generator for policy training, since the annotations are exact by construction.
  • The paired particle-residual trick used to isolate falling precipitation suggests a general pattern: separating stochastic visual effects from pose-corrected static renders makes it possible to add motion-dependent appearance without breaking mechanical consistency.
  • The evaluation design implies that for current video generators the main deficit under weather and geometry edits is not visual fidelity but the absence of implied vehicle dynamics; a test that recovers normalized braking progress from output clips could serve as a cheap screening metric for any driving-video editor.
  • Extending the shared state to lateral dynamics and measured tire data, which the paper lists as future work, would move the framework from braking-and-elevation counterfactuals toward full accident-scenario synthesis.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper presents muSync-GS, a framework for counterfactual driving-video synthesis that couples 3D Gaussian weather and road-geometry editing with a calibrated vehicle dynamics model. A precipitation-derived surface state z_surf = (h_t, h_s, c_s) is designed to control both road appearance and tire–road friction mu_c; a shared raised-cosine road-elevation profile drives both visible geometry edits and axle excitation; and a coupled longitudinal–vertical half-car model predicts speed, slip ratio, pitch, normal loads, and the resulting ego-camera pose. The evaluation includes a frozen 40-scalar calibration on 19 CarSim development cases, a 12-case held-out CarSim suite, an A2D2 cross-drive brake-dive check, and video-level VGGT motion comparisons against Cosmos, LTX-Video, and V ACE. The reported mean case-wise CarSim RMSEs are 0.0273 m/s for speed, 0.0590 degrees for pitch, 0.0101 for slip ratio, and 26.61 N for per-wheel normal load.

Significance. The paper's evaluation discipline is a real strength: 40 calibration scalars are frozen before a 12-case controlled holdout, all predefined cases are retained, per-case errors are reported, no temporal alignment or pitch-offset removal is used, and a CarSim-pose video reference provides an empirical rendering and monocular-recovery floor. The A2D2 leave-one-drive-out brake-dive check, although based on a small retained-event set, is a genuine external consistency test. If the central weather-to-friction claim were independently validated, muSync-GS would be a useful tool for generating physically synchronized counterfactual driving data. As it stands, the paper is unusually transparent about its main limitation, but that limitation touches the paper's central claim.

major comments (2)
  1. [Sec. 3.2, Eq. (5), Appendix E.1] The precipitation-to-friction mapping mu_c = phi_mu(w, z_surf) is not independently validated. In the 12-case CarSim holdout, Table S7 passes the same frozen mu_c to both CarSim and muSync-GS, so the reported RMSEs validate the vehicle model conditional on mu_c; they do not validate the claim that precipitation rate determines friction. Appendix E.1 explicitly states that the surface-parameter study is 'a bounded model-sensitivity analysis rather than independent validation against measured wet- or snow-road friction.' Because the abstract and Sec. 3 claim that a 'precipitation-derived road-surface state' jointly controls appearance and tire friction, this gap is load-bearing. The paper should either add external validation against measured friction-versus-water-depth or snow-depth data, or explicitly reframe the central claim as synchronization conditional on a user-specified mu_c.
  2. [Sec. 4.1, Table S7] The weather-severity holdout cases are presented as spanning new precipitation levels, but the mu_c values for these cases are generated by the same frozen mapping used during calibration, and the same values are supplied to CarSim. The holdout therefore tests interpolation of the calibrated vehicle model over a smooth mu_c schedule and cannot detect systematic error in Eq. (5)'s transition scales (h_c, h_cover) or anchor values. The text should state this limitation wherever the 12-case results are summarized, including the abstract's 'spanning precipitation levels' phrasing, or accompany the holdout with an independent check of Eq. (5).
minor comments (4)
  1. [Appendix A.5] The sentence 'They differ from muSync-GS by at most 0.010' is inconsistent with the reported numbers: |0.354 - 0.316| = 0.038 and |0.406 - 0.431| = 0.025. Please correct the claim or the reported values.
  2. [Sec. 3.3 and Appendix B.4] The statement 'The loads feed back into the longitudinal tire forces' is not literally true for flat-road braking, where Eq. (S18) substitutes analytic load transfer for the half-car loads. Please clarify that the feedback path is case-dependent.
  3. [Table 3] The muSync-GS row reports '–' for the Abs. pitch corr. column; the caption should explain that motion preservation is not applicable to muSync-GS because it does not consume a motion-preservation input video, to avoid implying a missing result.
  4. [Appendix A.4] For rain, the visual wet-road anchors are specified directly by precipitation rate (2/10/25/50 mm/h), while Eq. (4) says the rendering branch depends on z_surf. Please reconcile how h_t from Eq. (3) enters the rain rendering anchors rather than the rate alone.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the vehicle-response claims are genuine held-out predictions; the weather-to-friction mapping is an unvalidated external assumption, not a circular reduction.

full rationale

The paper's derivation chain is not circular. The shared surface state z_surf (Eq. 2) and friction branch mu_c (Eq. 5) are defined from weather controls via literature formulas with fixed anchors and scales, and the same mu_c is passed to both CarSim and muSync-GS in the weather holdouts. This means the reported 12-case CarSim agreement validates the reduced-order vehicle model conditional on mu_c, but it does not validate the precipitation-to-friction mapping itself. The paper explicitly discloses this limitation in Appendix E.1: 'This is a bounded model-sensitivity analysis rather than independent validation against measured wet- or snow-road friction.' That is an external-validation gap, not a case where a predicted quantity is equivalent to its input by construction. The vehicle model parameters are calibrated on 19 development cases and frozen before the 12-case holdout, so the speed, pitch, slip, and normal-load outputs on held-out precipitation levels, brake commands, and road profiles are genuine out-of-sample predictions. The A2D2 brake-dive evaluation likewise calibrates on two drives and evaluates on a held-out drive, using only brake pressure and event-start speed at test time. The self-citations in the reference list (e.g., SceneFactory [48] and related work [4,6,7]) are contextual and do not carry any load-bearing premise; no uniqueness theorem is imported, and the rendering components are adopted from independent prior work (WeatherEdit, RainyGS, Weather-Magician). The visual-physical coupling is an architectural construction that honestly states that both appearance and friction are functions of the same z_surf; this is a design choice, not a hidden equivalence between an input and an output.

Assumptions & free parameters 18 free parameters · 8 assumptions · 1 invented entities

The central claim rests on 18 free parameters (40 individual scalar entries plus hand-chosen surface anchors and transition scales) and a set of modeling assumptions from vehicle dynamics, tire mechanics, pavement drainage, and monocular video reconstruction. Most parameters are calibrated on a 19-case CarSim development set and frozen for the holdout; the precipitation-friction anchors are hand-selected from literature ranges with sensitivity bounds. The paper is transparent about these choices, but they limit the claim's scope to the calibrated regime.

free parameters (18)
  • Dry friction anchor mu_dry = 0.82 (sensitivity 0.78-0.86)
    Frozen anchor from literature range for dry bare pavement; hand-selected nominal, not measured for the reconstructed scenes (Table S11).
  • Flooded friction anchor mu_flooded = 0.34 (sensitivity 0.306-0.374)
    Frozen anchor for rain-flooded limit; value within literature wet-pavement range but hand-selected (Table S11).
  • Bare-wet transition anchor mu_bw = 0.60 (sensitivity 0.54-0.66)
    Anchors the snow coverage interpolation; hand-selected within field wet-pavement measurements (Table S11).
  • Snow friction anchor mu_snow = 0.24 (sensitivity 0.216-0.264)
    New-snow skid-resistance anchor from controlled tests; hand-selected nominal (Table S11).
  • Fresh-snow density rho_s = 100 kg/m3 (80-120)
    Conventional 10:1 water-to-snow density; site observations vary (Table S11).
  • Texture depth MTD = 0.50 mm (0.40-0.60)
    Dense-asphalt macrotexture nominal; route-dependent input, not measured for each scene (Table S11).
  • Drainage path length L_d = 3.75 m (3.0-4.5)
    Scenario flow-path input to the Gallaway relation (Table S11).
  • Crossfall S = 2.0% (1.6-2.4)
    Road cross-slope input to the drainage relation (Table S11).
  • Rain film-decay scale h_c = 0.30 mm (0.24-0.36)
    Model-specific exponential decay scale for friction reduction with water depth; not a physical constant (Eq. 5; Table S11).
  • Snow full-coverage threshold h_cover = 20 mm (16-24)
    Model-specific coverage saturation threshold for snow; not a road-design standard (Eq. 3; Table S11).
  • Tire-force efficiency eta_tire per mu_c knot = 8 entries, e.g., dry 1.0100; rain 0.9278-1.0000; snow 0.9588-0.9932
    Calibrated on 19 development flat-braking cases against CarSim speed and stopping distance (Tables S3, S5).
  • Slip-stiffness scale beta_kappa per knot = 8 entries, e.g., dry 1.65; rain/snow 1.00-1.63
    Calibrated against per-wheel slip-ratio RMSE on development cases (Tables S3, S5).
  • Brake-pitch scale gamma_theta per knot = 8 entries, approximately 1.10-1.14
    Calibrated against body-pitch RMSE during braking (Tables S3, S5).
  • Low-speed pitch hold ratio rho_h and hold/release durations tau_h, tau_r = rho_h=0.8; tau_h=0.05-0.20 s; tau_r=0.04-0.08 s
    Ad hoc stop-tail pitch rule fit to CarSim release behavior (Sec. B.4, Table S5).
  • Coupled brake-pitch scale (shared) = 0.98
    Calibrated on ten road-feature development cases to match pitch RMSE (Table S3).
  • Static total-pitch offset = 0.154150 deg
    Read from matched CarSim development traces (Table S5).
  • Coast-drag deceleration a_coast = 0.004 m/s^2
    Calibrated to shared development coasting response (Tables S3, S4).
  • A2D2 pressure-to-wheel-torque gain and effective pitch damping = Fold-specific estimates, not tabulated
    Estimated on two training drives (pressure gain from deceleration residual; damping from pitch/rate), frozen per held-out drive (Sec. C.4).
assumptions (8)
  • domain assumption Burckhardt tire-curve peak normalization (Eq. S7) represents tire-road force-slip behavior across dry, wet, and snow surfaces.
    The model assumes the fixed (c1,c2,c3) tuples and peak normalization to mu_eff hold for all cases; deviations would change slip and braking predictions.
  • domain assumption Precipitation-to-surface state equations (Eq. 3, Gallaway drainage and snow accumulation) map weather controls to water-film depth, snow depth, and coverage.
    The mapping is taken from the literature but is not validated against measured wet/snow road conditions for the test scenes; Appendix E calls it a bounded sensitivity analysis, not independent validation.
  • domain assumption 4-DOF half-car model with linear suspension, no lateral dynamics, and decoupled pitch/heave (Eq. 10, S15) captures the vertical and pitch response.
    The model omits roll, lateral load transfer, and nonlinear suspension; suitable only for straight-line, moderate-excitation scenarios.
  • domain assumption Raised-cosine road-elevation profile (Eq. 6) adequately represents speed humps and depressions at both the visual and tire-contact level.
    Real hazards have varied shapes; the paper evaluates only this parametric family.
  • domain assumption Unilateral tire-contact clamp (Eq. S14) stays inactive in all evaluated cases, so linear superposition of road and braking responses is valid.
    The paper asserts all cases remain in the active-contact regime (Sec. B.4); valid only for the tested conditions.
  • ad hoc to paper Low-speed pitch hold/release rule (Sec. B.4) correctly models stop-tail pitch after full stop.
    Introduced specifically to match CarSim stop-tail behavior; no physical basis beyond the calibration fit.
  • domain assumption VGGT monocular estimation recovers normalized forward progress and camera pitch well enough for video-level response comparison.
    The CarSim-pose oracle check (NRMSE 0.354/0.406) bounds this error for geometry; the weather-progress recovery is not independently bounded.
  • domain assumption CarSim 2022.1 trajectories are the ground truth for vehicle-response validation.
    The central numerical claims compare against CarSim, not instrumented real vehicles; real-vehicle evidence is limited to 14 A2D2 braking events.
invented entities (1)
  • Precipitation-derived surface state z_surf = (h_t, h_s, c_s)
    purpose: Shared internal variable that jointly controls road appearance and tire-road friction, linking weather controls to both visual and dynamic outputs.
    No direct measurement of this state is provided; in the CarSim holdouts the derived mu_c is passed to the reference simulator (Table S7), so the mapping's external validity is not established. It is a modeling abstraction, not a measured quantity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of muSync-GS: Physics-Synchronized Driving Video Synthesis for Weather and Geometric Road Hazards." pith.science (2026). https://pith.science/paper/AC6HHCTV

@misc{pith2026260804412,
  author       = {Pith},
  title        = {Pith review of: muSync-GS: Physics-Synchronized Driving Video Synthesis for Weather and Geometric Road Hazards},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AC6HHCTV}},
  note         = {Machine review of arXiv:2608.04412}
}
read the original abstract

High-quality driving data are essential for autonomous-driving systems and generative world models. However, rare and safety-critical scenarios involving adverse weather, braking under low tire--road friction, and uneven road geometry are costly and risky to collect at scale. Existing video-generation and 3D Gaussian editing methods can modify weather appearance or road geometry, but typically do not couple these edits with tire--road interaction and vehicle dynamics. As a result, an edited video may retain its original trajectory even when the modified road condition should alter braking, wheel slip, load transfer, and ego-camera motion. We present muSync-GS, a physics-synchronized framework for driving video synthesis under adverse-weather and road-elevation hazards. A precipitation-derived road-surface condition jointly controls road appearance and tire friction, while a shared road-elevation profile drives both visible road-geometry editing and axle excitation. A calibrated vehicle model predicts speed, slip ratio, normal loads, and pitch for constructing the ego-camera trajectory and synchronized physical annotations. On 12 held-out CarSim cases spanning precipitation levels, brake inputs, and road-profile parameters, the model achieves mean case-wise RMSEs of 0.0273 m/s for speed, 0.0590 degrees for pitch, 0.0101 for slip ratio, and 26.61 N for per-wheel normal load. Together with the reconstructed-scene experiments, these results show that muSync-GS accurately reproduces vehicle responses under held-out controls while synchronizing them with controllable scene edits and ego-camera motion.

Figures

Figures reproduced from arXiv: 2608.04412 by the authors.

Figure 1
Figure 1. Conceptual comparison between appearance-only editing and muSync-GS. Appearance-only [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of muSync-GS. Weather, road-elevation, and vehicle controls are converted into [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. muSync-GS and CarSim trajectories for a held-out 5 [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Peak-magnitude-normalized measured and predicted brake-dive responses from each [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Five-level weather controllability at frame 40 and seed 2025: sunny, light rain (10 [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Road editing under the same dry-weather input (seed 2025) before contact, at front-axle [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Coupled-physics ablation: (a) stopping distance across rain and snow severity; (b) pitch [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

72 extracted references · 40 canonical work pages

  1. [1]

    Gaia-1: A generative world model for autonomous driving, 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, 2023

    2023

  2. [3]

    Drivedreamer-2: Llm-enhanced world models for diverse driving video generation

    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. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 10412–10420,

    work page

  3. [4]

    Tao Li, Zilin Bian, Haozhe Lei, Fan Zuo, Ya-Ting Yang, Quanyan Zhu, Zhenning Li, Zhibin Chen, and Kaan Ozbay. Digital twin-based driver risk-aware predictive mobility analytics for real-time situational awareness through cooperative sensing.IEEE Transactions on Intelligent Transportation Systems, 26(11):20071–20090, 2025. doi: 10.1109/TITS.2025.3604569

    work page arXiv 2025

  4. [5]

    When language and vision meet road safety: Leveraging multimodal large language models for video-based traffic accident analysis.Accident Analysis & Prevention, 219:108077, 2025

    Ruixuan Zhang, Beichen Wang, Juexiao Zhang, Zilin Bian, Chen Feng, and Kaan Ozbay. When language and vision meet road safety: Leveraging multimodal large language models for video-based traffic accident analysis.Accident Analysis & Prevention, 219:108077, 2025. doi: 10.1016/j.aap.2025.108077

    work page arXiv 2025

  5. [6]

    Multi-level traffic-responsive tilt camera surveillance through predictive correlated online learning.Transportation Research Part C: Emerging Technologies, 167:104804, 2024

    Tao Li, Zilin Bian, Haozhe Lei, Fan Zuo, Ya-Ting Yang, Quanyan Zhu, Zhenning Li, and Kaan Ozbay. Multi-level traffic-responsive tilt camera surveillance through predictive correlated online learning.Transportation Research Part C: Emerging Technologies, 167:104804, 2024. doi: 10.1016/j.trc.2024.104804

    work page arXiv 2024

  6. [8]

    Weatheredit: Controllable 13 weather editing with 4D gaussian field

    Chenghao Qian, Wenjing Li, Yuhu Guo, and Gustav Markkula. Weatheredit: Controllable 13 weather editing with 4D gaussian field. InProceedings of the AAAI Conference on Artificial Intelligence, 2026. doi: 10.1609/aaai.v40i10.37802. arXiv:2505.20471

    arXiv 2026

  7. [10]

    WeatherGS: 3D scene recon- struction in adverse weather conditions via gaussian splatting

    Chenghao Qian, Yuhu Guo, Wenjing Li, and Gustav Markkula. WeatherGS: 3D scene recon- struction in adverse weather conditions via gaussian splatting. InIEEE International Conference on Robotics and Automation, 2025. doi: 10.1109/ICRA55743.2025.11128699. arXiv:2412.18862

    arXiv 2025

  8. [12]

    Weather-magician: Reconstruction and rendering framework for 4D weather synthesis in real time, 2025

    Chen Sang, Yeqiang Qian, Jiale Zhang, Chunxiang Wang, and Ming Yang. Weather-magician: Reconstruction and rendering framework for 4D weather synthesis in real time, 2025. URL https://arxiv.org/abs/2505.19919

    arXiv 2025

Show all 72 references
  1. [13]

    Drivingdiffusion: Layout-guided multi-view driving scenarios video generation with latent diffusion model

    Xiaofan Li, Yifu Zhang, and Xiaoqing Ye. Drivingdiffusion: Layout-guided multi-view driving scenarios video generation with latent diffusion model. InEuropean Conference on Computer Vision, 2024. doi: 10.1007/978-3-031-73229-4 27

    2024 doi

  2. [14]

    Magicdrive: Street view generation with diverse 3d geometry control, 2024

    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, 2024. URL https: //arxiv.org/abs/2310.02601

    2024 arXiv

  3. [15]

    Wovogen: World volume-aware diffusion for controllable multi-camera driving scene generation

    Jiachen Lu, Ze Huang, Zeyu Yang, Jiahui Zhang, and Li Zhang. Wovogen: World volume-aware diffusion for controllable multi-camera driving scene generation. InEuropean Conference on Computer Vision, 2024. doi: 10.1007/978-3-031-72989-8 19

    2024 doi

  4. [16]

    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. URL https://arxiv.org/abs/2503.20523

    2025 arXiv

  5. [18]

    SafeMVDrive: Multi-view safety-critical driving video synthesis in the real world domain, 2025

    Jiawei Zhou, Linye Lyu, Zhuotao Tian, Cheng Zhuo, and Yu Li. SafeMVDrive: Multi-view safety-critical driving video synthesis in the real world domain, 2025. URL https://arxiv.org/ abs/2505.17727

    2025 arXiv

  6. [19]

    Challenger: Affordable adversarial driving video generation,

    Zhiyuan Xu, Bohan Li, Huan-ang Gao, Mingju Gao, Yong Chen, Ming Liu, Chenxu Yan, Hang Zhao, Shuo Feng, and Hao Zhao. Challenger: Affordable adversarial driving video generation,

  7. [20]

    Cosmos-transfer1: Conditional world generation with adaptive multimodal control,

    NVIDIA. Cosmos-transfer1: Conditional world generation with adaptive multimodal control,

  8. [21]

    URL https://arxiv.org/abs/2505.15880

    arXiv

  9. [23]

    URL https://arxiv.org/abs/2503.14492

    arXiv

  10. [25]

    HUGSIM: A real-time, photo-realistic and closed-loop simulator for autonomous driving, 2024

    Hongyu Zhou, Longzhong Lin, Jiabao Wang, Yichong Lu, Dongfeng Bai, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. HUGSIM: A real-time, photo-realistic and closed-loop simulator for autonomous driving, 2024. URL https://arxiv.org/abs/2412.01718

    2024 arXiv

  11. [26]

    Street gaussians: Modeling dynamic urban scenes with 14 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 14 gaussian splatting. InEuropean Conference on Computer Vision, pages 156–173. Springer,

  12. [27]

    HorizonForge: Driving scene editing with any trajectories and any vehicles, 2026

    Yifan Wang, Francesco Pittaluga, Zaid Tasneem, Chenyu You, Manmohan Chandraker, and Ziyu Jiang. HorizonForge: Driving scene editing with any trajectories and any vehicles, 2026. URL https://arxiv.org/abs/2602.21333

    2026

  13. [29]

    DreamDrive: Generative 4D scene modeling from street view images, 2025

    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, 2025. URL https://arxiv.org/abs/2501.00601

    2025 arXiv

  14. [30]

    SplatAD: Real-time lidar and camera rendering with 3D gaussian splatting for autonomous driving

    Georg Hess, Carl Lindstr¨ om, Maryam Fatemi, Christoffer Petersson, and Lennart Svensson. SplatAD: Real-time lidar and camera rendering with 3D gaussian splatting for autonomous driving. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025....

    2025

  15. [31]

    WorldSplat: Gaussian-centric feed-forward 4D scene generation for autonomous driving, 2025

    Ziyue Zhu, Zhanqian Wu, Zhenxin Zhu, Lijun Zhou, Haiyang Sun, Bing Wan, Kun Ma, Guang Chen, Hangjun Ye, Jin Xie, and Jian Yang. WorldSplat: Gaussian-centric feed-forward 4D scene generation for autonomous driving, 2025. URL https://arxiv.org/abs/2509.23402

    2025

  16. [32]

    3D streetunveiler with semantic-aware 2DGS: A simple baseline

    Jingwei Xu, Yikai Wang, Yiqun Zhao, Yanwei Fu, and Shenghua Gao. 3D streetunveiler with semantic-aware 2DGS: A simple baseline. InInternational Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=G6aJyS0ZV0. arXiv:2405.18416

    2025 arXiv

  17. [33]

    Physgaussian: Physics-integrated 3D gaussians for generative dynamics

    Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3D gaussians for generative dynamics. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4389–4398, 2024. doi: 10.1...

    2024

  18. [34]

    Drive- Dreamer4D: World models are effective data machines for 4D driving scene representation,

    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. Drive- Dreamer4D: World models are effective data machines for 4D driving scene representation,

  19. [35]

    URL https://arxiv.org/abs/2410.13571

    arXiv

  20. [36]

    Physics3d: Learning physical properties of 3D gaussians via video diffusion, 2024

    Fangfu Liu, Hanyang Wang, Shunyu Yao, Shengjun Zhang, Jie Zhou, and Yueqi Duan. Physics3d: Learning physical properties of 3D gaussians via video diffusion, 2024. URL https://arxiv.org/ abs/2406.04338

    2024 arXiv

  21. [37]

    AutoWeather4D: Autonomous driving video weather conversion via G-buffer dual-pass editing,

    Tianyu Liu, Weitao Xiong, Kunming Luo, Manyuan Zhang, Peng Li, Yuan Liu, and Ping Tan. AutoWeather4D: Autonomous driving video weather conversion via G-buffer dual-pass editing,

  22. [38]

    Learning physics-grounded 4D dynamics with neural gaussian force fields, 2026

    Shiqian Li, Ruihong Shen, Junfeng Ni, Chang Pan, Chi Zhang, and Yixin Zhu. Learning physics-grounded 4D dynamics with neural gaussian force fields, 2026. URL https://arxiv.org/ abs/2602.00148

    2026

  23. [39]

    Real2sim: A physics-driven and editable gaussian splatting framework for autonomous driving scenes, 2026

    Kaicong Huang, Talha Azfar, Weisong Shi, and Ruimin Ke. Real2sim: A physics-driven and editable gaussian splatting framework for autonomous driving scenes, 2026. URL https: //arxiv.org/abs/2605.13591

    2026 arXiv

  24. [40]

    Feature splatting: Language-driven physics-based scene synthesis and editing

    Ri-Zhao Qiu, Ge Yang, Weijia Zeng, and Xiaolong Wang. Feature splatting: Language-driven physics-based scene synthesis and editing. InEuropean Conference on Computer Vision, 2024. doi: 10.1007/978-3-031-72940-9 21. arXiv:2404.01223

    2024 arXiv

  25. [41]

    Tianyu Huang, Haoze Zhang, Yihan Zeng, Zhilu Zhang, Hui Li, Wangmeng Zuo, and Rynson W. H. Lau. Dreamphysics: Learning physics-based 3D dynamics with video diffusion priors. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. doi: 10.1609/aaai.v39i4. 32389....

    2025 arXiv

  26. [42]

    Physgen: Rigid-body physics-grounded image-to-video generation

    Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded image-to-video generation. InEuropean Conference on Computer Vision,

  27. [43]

    Efficient physics simulation for 3D scenes via MLLM-guided gaussian splatting

    Haoyu Zhao, Hao Wang, Xingyue Zhao, Hao Fei, Hongqiu Wang, Chengjiang Long, and Hua Zou. Efficient physics simulation for 3D scenes via MLLM-guided gaussian splatting. InProceedings of the IEEE/CVF International Conference on Computer Vision, 2025. doi: 10.1109/ICCV51701.2025....

    2025

  28. [44]

    WISA: World simulator assistant for physics-aware text-to-video generation, 2025

    Jing Wang, Ao Ma, Ke Cao, Jun Zheng, Zhanjie Zhang, Jiasong Feng, Shanyuan Liu, Yuhang Ma, Bo Cheng, Dawei Leng, Yuhui Yin, and Xiaodan Liang. WISA: World simulator assistant for physics-aware text-to-video generation, 2025. URL https://arxiv.org/abs/2503.08153

    2025 arXiv

  29. [45]

    Physvid: Physics-aware local conditioning for generative video models

    Saurabh Pathak, Elahe Arani, Mykola Pechenizkiy, and Bahram Zonooz. Physvid: Physics-aware local conditioning for generative video models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026. URL https://arxiv.org/abs/2603.26285. arXiv:2603.26285

    2026

  30. [46]

    CARLA-GS: Decoupling representation, reasoning, and physics simulation for autonomous driving corner-case synthesis, 2026

    Kaicong Huang, Meng Ma, and Ruimin Ke. CARLA-GS: Decoupling representation, reasoning, and physics simulation for autonomous driving corner-case synthesis, 2026. URL https://arxiv. org/abs/2607.07601

    2026 arXiv

  31. [47]

    Physics-aware 3D gaussian editing for driving scene generation, 2026

    Feng Zhou, Jian Zhang, Yuhang Sun, He Wang, Qiong Wen, Debao Kong, Tieru Wu, and Rui Ma. Physics-aware 3D gaussian editing for driving scene generation, 2026. URL https: //arxiv.org/abs/2605.25373

    2026 arXiv

  32. [48]

    Scenefactory: GPU-accelerated multi-agent driving simulation with physics-based vehicle dynamics, 2026

    Yicheng Zhu, Yang Chen, Tao Li, and Zilin Bian. Scenefactory: GPU-accelerated multi-agent driving simulation with physics-based vehicle dynamics, 2026. URL https://arxiv.org/abs/2605. 08528

    2026

  33. [49]

    arXiv:2409.18964

    URL https://arxiv.org/abs/2409.18964. arXiv:2409.18964

    arXiv

  34. [50]

    VLIPP: Towards physically plausible video generation with vision and language informed physical prior

    Xindi Yang, Baolu Li, Yiming Zhang, Zhenfei Yin, Lei Bai, Liqian Ma, Zhiyong Wang, Jianfei Cai, Tien-Tsin Wong, Huchuan Lu, and Xu Jia. VLIPP: Towards physically plausible video generation with vision and language informed physical prior. InProceedings of the IEEE/CVF Internat...

    2025 arXiv

  35. [51]

    Scalability in perception for autonomous driving: Waymo open dataset

    Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aur´ elien 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 Zhan...

    2020

  36. [52]

    Jakob Geyer, Yohannes Kassahun, Mentar Mahmudi, Xavier Ricou, Rupesh Durgesh, Andrew S. 16 Chung, Lorenz Hauswald, Viet Hoang Pham, Maximilian Muehlegg, Sebastian Dorn, Tiffany Fernandez, Martin Jaenicke, Sudesh Mirashi, Chiragkumar Savani, Martin Sturm, Oleksandr Vorobiov, an...

    2020

  37. [53]

    GenieDrive: Towards physics-aware driving world model with 4D occupancy guided video generation, 2025

    Zhenya Yang, Zhe Liu, Yuxiang Lu, Liping Hou, Chenxuan Miao, Siyi Peng, Bailan Feng, Xiang Bai, and Hengshuang Zhao. GenieDrive: Towards physics-aware driving world model with 4D occupancy guided video generation, 2025. URL https://arxiv.org/abs/2512.12751

    2025

  38. [54]

    Point as skeleton: Accumulated point cloud enhanced autoregressive generation for closed-loop autonomous driving simulation, 2026

    Songbur Wong, Xiaosong Jia, Junqi You, Bo Zhang, Pei Xu, Renqiu Xia, Yuping Qiu, Shaofeng Zhang, Zelin Zhao, Xuechao Yan, Yuchen Zhou, Yurui Chen, Wen Guo, Hang Xu, and Junchi Yan. Point as skeleton: Accumulated point cloud enhanced autoregressive generation for closed-loop au...

    2026 arXiv

  39. [55]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InProceedings of the...

    2021

  40. [56]

    CARLA: An open urban driving simulator

    Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. InProceedings of the 1st Annual Conference on Robot Learning, volume 78 ofProceedings of Machine Learning Research, pages 1–16. PMLR, 2017. URL https://...

    2017

  41. [57]

    CarSim: Vehicle dynamics simulation software

    Mechanical Simulation Corporation. CarSim: Vehicle dynamics simulation software. https: //www.carsim.com/, 2026. Accessed: 2026-07-01

    2026

  42. [58]

    F AST-VQA: Efficient end-to-end video quality assessment with fragment sampling

    Haoning Wu, Chaofeng Chen, Jingwen Hou, Liang Liao, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. F AST-VQA: Efficient end-to-end video quality assessment with fragment sampling. InEuropean Conference on Computer Vision, pages 538–554. Springer, 2022. URL https://arxiv.org...

    2022 arXiv

  43. [59]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The un- reasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 586–595, 2018. URL https://arxiv.org/...

    2018 arXiv

  44. [60]

    Vggt: Visual geometry grounded transformer

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025

    2025

  45. [61]

    Exploring video quality assessment on user generated contents from aesthetic and technical perspectives

    Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. InProceedings of the IEEE/CVF International Conference on ...

    2023 arXiv

  46. [62]

    Friction measurement methods and the correlation between road friction and traffic safety: A literature review

    Carl-Gustaf Wallman and Henrik ˚Astr¨ om. Friction measurement methods and the correlation between road friction and traffic safety: A literature review. Technical Report VTI meddelande 911A, Swedish National Road and Transport Research Institute (VTI), Link¨ oping, Sweden,

  47. [63]

    Ltx-video: Realtime video latent diffusion, 2024

    Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richard- son, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. Ltx-video: Realtime video latent diffusion, 2024

    2024

  48. [64]

    Vogel, W¨ urzburg, Germany, 1993

    Manfred Burckhardt.Fahrwerktechnik: Radschlupf-Regelsysteme. Vogel, W¨ urzburg, Germany, 1993

    1993

  49. [65]

    Astm e965/e965m-15(2019): Standard test method for measuring pavement macrotexture depth using a volumetric technique

    ASTM International. Astm e965/e965m-15(2019): Standard test method for measuring pavement macrotexture depth using a volumetric technique. Technical report, ASTM International, West Conshohocken, PA, 2019. URL https://www.astm.org/e0965-15r19.html

    2019

  50. [66]

    Kilgore, A

    R. Kilgore, A. T. Atayee, and G. R. Herrmann. Urban drainage design. Technical Report Hydraulic Engineering Circular No. 22, Fourth Edition, FHW A-HIF-24-006, Federal Highway Administration, February 2024. URL https://www.fhwa.dot.gov/engineering/hydraulics/pubs/ hif24006.pdf

    2024

  51. [67]

    B. M. Gallaway, R. E. Schiller, Jr., and J. G. Rose. The effects of rainfall intensity, pavement cross slope, surface texture, and drainage length on pavement water depths. Technical Report Research Report 138-5, Texas Transportation Institute, Texas A&M University, May 1971. ...

    1971

  52. [68]

    B. M. Gallaway, D. L. Ivey, G. G. Hayes, W. B. Ledbetter, R. M. Olson, D. L. Woods, and R. F. Schiller. Pavement and geometric design criteria for minimizing hydroplaning. Technical Report FHW A-RD-79-31, Federal Highway Administration, December 1979. URL https://trid.trb.org/...

    1979

  53. [71]

    Ivan, Nalini Ravishanker, Eric Jackson, Brien Aronov, and Sizhen Guo

    John N. Ivan, Nalini Ravishanker, Eric Jackson, Brien Aronov, and Sizhen Guo. Incorporating wet pavement friction into traffic safety analysis. Technical Report JHR 10-324, Connecticut Transportation Institute, University of Connecticut, November 2010. URL https://rosap.ntl. b...

    2010

  54. [72]

    Flintsch, and Edgar de Le´ on Izeppi

    Ilker Boz, Gerardo W. Flintsch, and Edgar de Le´ on Izeppi. Functional characteristics of dense-graded asphalt surface mixtures. Technical Report FHW A/VTRC 23-R15, Virginia Transportation Research Council, April 2023. URL https://vtrc.virginia.gov/media/vtrc/vtrc- pdf/vtrc-pd...

    2023

  55. [75]

    Density of freshly fallen snow in the central rocky mountains.Bulletin of the American Meteorological Society, 81(7):1577–1587, 2000

    Arthur Judson and Nolan Doesken. Density of freshly fallen snow in the central rocky mountains.Bulletin of the American Meteorological Society, 81(7):1577–1587, 2000. doi: 10.1175/1520-0477(2000)081⟨1577:DOFFSI⟩2.3.CO;2. URL https://climate.colostate.edu/pdfs/ SnowDensity BAMS.pdf

    2000 doi

  56. [76]

    Skid resistance of snow- or ice-covered roads

    Kaoru Ichihara and Mitsumasa Mizoguchi. Skid resistance of snow- or ice-covered roads. In Snow Removal and Ice Control Research, number 115 in Highway Research Board Special Report, pages 104–114. Highway Research Board, 1970. URL https://onlinepubs.trb.org/ Onlinepubs/sr/sr11...

    1970

  57. [77]

    evaluatez r,f /rfrom the current path position

  58. [78]

    integrate the common half-car and computeF z,f /r

  59. [79]

    normal-load feedback

    update tire force, wheel rotation, speed, and path position using Eqs. S9–S11. Flat-road braking uses acceleration-dependent longitudinal load transfer in the wheel solver, F long z,f = max m glr −˙vhcg Lw ,0 , F long z,r = max m glf + ˙vhcg Lw ,0 . (S18) The common half-car s...

    1960

  60. [80]

    a clear dry road seen from a dashcam,

    The weather controls are shared, while visible road support follows each scene’s reconstruction and road mask. are frozen. DOVER, F AST-VQA, CLIP severity, and temporal LPIPS use the 108-video weather subset; localization, anchor, and T-IoU use the 20-video road subset. For no...

    2025

  61. [2001]

    URL https://vti.diva-portal.org/smash/get/diva2:673366/FULLTEXT01.pdf

  62. [2024]

    doi: 10.1007/978-3-031-73464-9 10

    doi

  63. [2025]

    doi: 10.1609/aaai.v39i10.33130

    doi

  64. [2026]

    URL https://arxiv.org/abs/2603.26546

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.