Pith. sign in

REVIEW 4 major objections 5 minor 50 references

GSRAIN claims that rainfall synthesis in 3D Gaussian driving scenes can be made physically calibrated and intensity-controllable over 0–13 mm/h by separating high-frequency raindrop Gaussians from low-frequency diffusion-based rainy appeara

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

2026-08-04 12:49 UTC pith:OVACO2KF

load-bearing objection GSRAIN offers a genuinely useful two-branch rain synthesis pipeline for 3DGS driving scenes, but the headline 'physically calibrated 0–13 mm/h control' is not actually specified anywhere in the paper. the 4 major comments →

arxiv 2608.02177 v1 pith:OVACO2KF submitted 2026-08-03 cs.CV

GSRAIN: Physically Calibrated High-/Low-Frequency Rainfall Synthesis for 3D Gaussian Driving Scenes

Fanyu Wang , Longgao Zhang , Junyi Chen This is my paper
classification cs.CV
keywords 3D Gaussian Splattingrainfall synthesisrainfall intensity controldrop size distributiondiffusion modelmulti-view consistencyautonomous driving simulationclosed-loop testing
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 tries to establish that rainfall can be added to a 3D Gaussian Splatting driving scene with two separated mechanisms: a high-frequency branch that places rain-streak and haze Gaussians according to measured raindrop statistics (diameter, velocity, number density), and a low-frequency branch that transfers a global rainy look (wet roads, reflections, dimmed contrast) through a geometry-aware single-step diffusion model. These are fused in one 3DGS scene by depth-ordered alpha compositing. If correct, this gives autonomous-driving developers a repeatable, closed-loop-compatible test environment where rainfall intensity is a numerical parameter over 0–13 mm/h. The paper reports a Fréchet Inception Distance of 149.09 for the rainy appearance, below two prior transfer baselines, and shows that an end-to-end driving system and two object detectors respond differently as intensity rises—including collisions in a cut-in scenario at 8.5 and 12.0 mm/h.

Core claim

GSRAIN claims that measured raindrop statistics can be embedded directly into a 3D Gaussian scene, making rainfall intensity a calibrated parameter. High-frequency rain uses Gaussian ellipsoids sized by a gamma drop-size distribution fit to disdrometer data, with speeds from a fitted velocity–diameter curve; far drops become low-opacity haze via a distance-dependent opacity rule. Low-frequency wetness uses a single-step diffusion model conditioned on surface normals and coordinated across views by mutual self-attention. The branches are fused by depth-ordered alpha compositing. Tests show detection confidence falls with intensity for a distant dark car, and a planner collides in a cut-in sce

What carries the argument

The central mechanism is the high-/low-frequency decoupling inside a single 3D Gaussian scene. High-frequency rain arises from Gaussian ellipsoids whose count, size, and velocity are drawn from a gamma drop-size distribution (shape μ_R and scale θ_R), a fitted quadratic velocity–diameter relation, and a distance-dependent opacity that turns near-field droplets into streaks and far-field droplets into haze. Low-frequency rainy appearance is produced by a single-step diffusion model conditioned on surface normals and coordinated across views by mutual self-attention. The two branches are fused via depth-ordered alpha compositing so that local raindrop disturbances and global wet appearance coe

Load-bearing premise

The load-bearing premise is that the rainfall-intensity-to-drop-distribution mapping is fully specified and reproducible, but the paper lists parameters at only three intensities and gives no interpolation rule, leaving the 0–13 mm/h claim dependent on an unspecified fit.

What would settle it

Render the same scene at an intermediate intensity such as 5 mm/h using any consistent interpolation of the three tabulated parameter sets; if the streak density and haze do not follow the monotone trend implied by 1.2 → 8.5 → 12.0 mm/h, or if the gamma scale at 12.0 mm/h (0.085 mm) cannot be reconciled with the 8.5 mm/h value (0.114 mm) under a smooth physical fit, the claimed calibrated control fails.

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

If this is right

  • Rainfall intensity becomes a numeric test parameter: an evaluator can vary a single scene from clear to 1.2, 8.5, and 12.0 mm/h and observe how perception and planning change.
  • Repeatability follows from the parameterization: the same intensity setting yields the same statistical rain, which is what regression testing of driving stacks requires.
  • Scene-dependent effects are visible: a light close car keeps confidence under rain, while a distant dark car loses confidence and can be misclassified, so rain tests can expose failures a global metric would hide.
  • A planner can be pushed across a behavior threshold: in the cut-in scenario the system collides at 8.5 and 12.0 mm/h but not in clear weather or at 1.2 mm/h.
  • The low-frequency transfer reaches a lower FID than two prior baselines, indicating closer feature statistics to real rainy images, although FID alone does not certify physical accuracy.

Where Pith is reading between the lines

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

  • The 0–13 mm/h control claim would be stronger if the paper gave the closed-form mapping from R to the drop-size and density parameters; it currently tabulates only three intensities and leaves interpolation unspecified.
  • The high/low-frequency split suggests a general recipe for adverse weather: fog, snow, and spray could each be modeled as particle-statistics Gaussians plus a global appearance transfer.
  • Because the low-frequency branch is 2D transfer re-optimized into 3D, intensity currently modulates only the high-frequency part; a direct 3D low-frequency model would let intensity also control wet-road gloss and global contrast.
  • The observed distance-dependent detector degradation hints at a targeted test-generation strategy: place dark or distant objects at intersections and sweep rain intensity to probe perception failure modes.

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

4 major / 5 minor

Summary. GSRAIN is a two-branch rainfall synthesis method for 3D Gaussian Splatting (3DGS) driving scenes. The high-frequency branch models near-field rain streaks and far-field haze as Gaussian primitives, with raindrop characteristics (gamma drop-size distribution, velocity–diameter relation, number density) fitted to measured rainfall data. The low-frequency branch uses a LoRA-fine-tuned single-step diffusion model with a surface-normal channel, cross-view feature matching, and mutual self-attention to produce multi-view-consistent rainy appearance, which is then reconstructed into a rainy 3DGS background. The two branches are fused in one 3DGS renderer. Experiments report FID comparisons with CycleGAN-Turbo and WeatherEdit, ablations for the normal channel and mutual attention, object-detection tests with YOLO variants, and closed-loop UniAD driving evaluations under clear/1.2/8.5/12.0 mm/h rainfall. The central claim is physically calibrated continuous rainfall-intensity control over 0–13 mm/h with multi-view consistency.

Significance. If fully substantiated, GSRAIN addresses a relevant gap: providing controllable, repeatable, multi-view-consistent rainy scenes for closed-loop autonomous-driving evaluation. The explicit modeling of raindrop size distributions, the high-quality velocity–diameter fit (R²=0.997), the geometry-aware diffusion branch, and the downstream UniAD/YOLO evaluations are strengths. However, the central 'physically calibrated continuous 0–13 mm/h control' claim currently rests on an underspecified mapping and an incomplete quantitative evaluation. The manuscript would be significantly strengthened by releasing fitted functional forms or code for the rainfall-intensity mapping, adding statistical rigor to FID and multi-view-consistency metrics, and providing external validation that the synthesized intensities behave like real rainfall of the same intensity.

major comments (4)
  1. [Algorithm 1, §III-C, Table III] The rainfall-intensity-to-parameter mapping is underspecified. Algorithm 1 Stage 1 line 2 says 'Determine N̄(R) and the gamma parameters μ_R, θ_R from the target rainfall intensity R' without giving any functional form. Table III provides only three discrete points (R=1.2, 8.5, 12.0 mm/h); moreover, μ_R (9.705/7.132/10.466) and θ_R (0.074/0.114/0.085) are nonmonotonic in R, so even simple interpolation is not well defined. No fitted equations, lookup tables, raw disdrometer data, or code are provided. Without this mapping, the advertised continuous 0–13 mm/h control cannot be reproduced at unlisted intensities, and the 'physically calibrated' claim is not supportable. Additionally, the fog-Gaussian count M in Algorithm 1 is never defined, leaving the haze density unspecified.
  2. [§IV-A.3, Table II] The headline FID comparison (149.09 vs 155.71 vs 157.94) is reported as point estimates with no error bars, number of test images, random seeds, or statistical significance test. The margins are small relative to typical FID variability, and no protocol details are given. Please report mean±std over multiple runs or confidence intervals, and specify the exact evaluation subset and preprocessing. As written, the claimed FID improvement is not statistically grounded.
  3. [§III-D.2, §IV-A.2, Figs. 6 and 8] The paper claims 'multi-view-consistent rainy appearance' as a key advantage, but the evidence is purely qualitative. No quantitative metric is provided to measure cross-view consistency (e.g., matched-region LPIPS/SSIM, feature-matching error, or downstream 3DGS reconstruction quality). Given that 3DGS reconstruction is sensitive to inter-view inconsistency, please add a quantitative cross-view consistency evaluation.
  4. [§IV-B.4, Eq. (7)] The 'graded rainfall control' demonstration in Fig. 10 is not an independent validation: higher intensity producing denser streaks/haze follows directly from the fitted N̄(R) and the chosen Gaussian placement, so it re-states the model parameters rather than confirming physical calibration. Please add external validation, e.g., comparing synthesized scenes against real rainfall images at matched intensities with a perceptual or statistical metric, or measuring agreement with independent rain-rate observations. The hand-set constants φ=2 and α_m=0.05 in Eq. (7) also need justification or a sensitivity study.
minor comments (5)
  1. [Figures 5–8, 13–17] Several figure panels contain garbled '/uni000...' tokens in the captions or embedded text; these should be cleaned before submission.
  2. [Fig. 1 and Table I] RainSense is referred to both as a disdrometer (Fig. 1) and as a dataset with 22 clear/22 rainy images (Table I). Clarify the data-acquisition setup and how the disdrometer measurements relate to the image dataset.
  3. [Algorithm 1, Eq. (12)] The fog-Gaussian count M, the maximum duplication number n_max, and the value of τ in Eq. (12) are not specified. Please provide these values or explain how they are determined.
  4. [§IV-A.1 and Table II] Gemini 3.1 Flash is excluded from the quantitative FID table because it is API-based. Please state explicitly why FID could not be computed for it (e.g., no fixed training set), and consider reporting a zero-shot FID if feasible.
  5. [§II-B, §IV] RainyGS (ref. [29]) is discussed in related work but is not included in comparisons. If it is not directly comparable, say so; otherwise include it in the FID or qualitative evaluation.

Circularity Check

1 steps flagged

Rainfall-intensity control demonstration is forced by the fitted density-to-R mapping that defines it, so the graded-control result is true by construction.

specific steps
  1. fitted input called prediction [Section IV-B-4 (Rainfall-Intensity Control), Table III, Fig. 10; Algorithm 1, Stage 1 lines 2-3]
    "GSRAIN adjusts rainfall over 0–13 mm/h through calibrated parameter mappings. Table III lists the number density and gamma-distribution parameters for three target intensities. The simulated results in Fig. 10 show that both rain-streak density and far-field haze increase with rainfall intensity, indicating graded control of the visual rainfall strength."

    The intensity-dependence is an input, not a prediction. Algorithm 1 says 'Determine N̄(R) and the gamma parameters μ_R, θ_R from the target rainfall intensity R' and then 'Compute ... the primitive counts N = N̄(R)V_k', so the number of rain-streak Gaussians is directly proportional to the fitted density N̄(R). Table III lists N̄(1.2)=271, N̄(8.5)=1190, N̄(12.0)=1710 m^-3, monotonically increasing with R. Thus the Fig. 10 observation that rain-streak density increases with R is guaranteed by construction; the demonstration re-states the fitted mapping rather than validating it.

full rationale

The paper's central asset is a 3DGS rainfall simulator whose high-frequency branch converts measured raindrop statistics into Gaussian primitives. The fitted inputs are the drop-size distribution parameters, number density, and velocity curve, all taken from RainSense disdrometer data. The strongest circularity is in the 'Rainfall-Intensity Control' experiment: the claimed graded control at 1.2, 8.5, and 12.0 mm/h is shown by rendering scenes whose streak count N is set as N = N̄(R)V_k, i.e., the density N̄(R) was itself fitted to those same R values. The visual 'increase with rainfall intensity' is therefore a restatement of the chosen input values, not an independent check that simulated rain at 8.5 mm/h behaves like real 8.5 mm/h rain. The FID comparison (149.09 vs. 155.71 and 157.94) is an external, non-circular benchmark, and the closed-loop UniAD / YOLO experiments are downstream applications rather than predictions derived from the fitted parameters. No load-bearing self-citation chain or uniqueness argument appears in the paper. Separate from circularity, the mapping R -> (N̄, μ_R, θ_R) is underspecified: only three tabulated points are given, the gamma shape parameter is nonmonotonic (9.705, 7.132, 10.466), and no interpolation or extrapolation rule is supplied for the claimed 0-13 mm/h continuum; this is a reproducibility/validation gap, not itself a circular step. Overall, one central demonstration reduces by construction, warranting a partial circularity score of 6.

Axiom & Free-Parameter Ledger

8 free parameters · 8 axioms · 0 invented entities

The physical calibration rests on measured raindrop statistics (DSD gamma parameters, number density, velocity-diameter relation) which are fit to data and tabulated at three intensities; all other constants (φ, haze opacity, region sizes) are hand-set. The paper provides no independent falsifiable prediction—the intensity-dependent effects are constructed from the fitted parameters, so the 'control' is not a validated forecast.

free parameters (8)
  • Gamma DSD shape parameter μ_R = μ_R = 9.705 @ 1.2 mm/h; 7.132 @ 8.5 mm/h; 10.466 @ 12.0 mm/h
    Fitted to RainSense disdrometer drop-size spectra; tabulated in Table III. No functional relation to R is given.
  • Gamma DSD scale parameter θ_R (mm) = θ_R = 0.074 @ 1.2 mm/h; 0.114 @ 8.5 mm/h; 0.085 @ 12.0 mm/h
    Fitted to RainSense drop-size spectra; tabulated in Table III. No interpolation law is specified.
  • Raindrop number density N̄(R) = 271 m^-3 @ 1.2 mm/h; 1190 m^-3 @ 8.5 mm/h; 1710 m^-3 @ 12.0 mm/h
    Used to set Gaussian primitive counts in Algorithm 1; derived from measured rainfall data but only at three intensities.
  • Velocity-diameter quadratic coefficients = V(D) = -0.342D^2 + 3.379D + 0.340
    Least-squares fit to RainSense mean velocities, R^2 = 0.997 (Eq 9); assumed valid across all scenes and intensities.
  • Exposure time T_exp = 0.05 s
    Set for the simulation; directly controls streak length and opacity via Eqs 5 and 10. It is a physical camera parameter, but the value is a choice made in §IV-B2.
  • Critical haze coefficient φ = 2
    Hand-set; described as depending on scene brightness and camera sensitivity, but no supporting measurement or sensitivity study is given (Eq 7, §IV-B2).
  • Haze opacity α_m = 0.05
    Assigned to far-field haze Gaussians; no physical derivation or ablation (Algorithm 1, §IV-B2).
  • Rain-streak and haze region extents = Streak [-1,5]×[-3,3]×[-1.5,4.5] m; haze [-10,60]×[-10,10]×[-1.5,20.5] m
    Chosen for Waymo scenes; affects which Gaussians are visible; no justification or sensitivity test is reported (§IV-B2).
axioms (8)
  • standard math 3D Gaussian Splatting representation and differentiable rendering (Eqs 1-2)
    Background rendering pipeline; standard in the field.
  • standard math Gamma distribution models the raindrop size distribution (Eq 8)
    Adopted from Thurai & Bringi [39]; standard DSD model.
  • domain assumption Raindrop equilibrium axis ratio is linear in diameter: τ(D)=kD+b, k=-0.07, b=1.07 (Eq 4)
    Taken from Beard & Chuang [26]; assumes all raindrops reach equilibrium shape, which is an approximation for turbulent rain.
  • domain assumption Raindrop opacity and streak length follow α0 = D/(V T_exp) and L = V T_exp (Eqs 5, 10)
    From Garg & Nayar [27]; assumes uniform exposure and ignores motion blur of background.
  • ad hoc to paper Distance-dependent opacity model with φ and α_m (Eq 7)
    Introduced in this paper with hand-set constants; no independent justification.
  • domain assumption Velocity-diameter fit Vμ(D) from RainSense disdrometer extrapolates to all scenes and intensities (Eq 9)
    Fit to a single disdrometer dataset; transfer to other geographic/lighting conditions is assumed.
  • ad hoc to paper Continuous rainfall intensity control over 0-13 mm/h is achievable from the three tabulated parameter sets (Table III)
    The paper gives no interpolation/extrapolation rule between 1.2, 8.5, and 12.0 mm/h.
  • domain assumption Low-frequency rainy appearance transfer via fine-tuned CycleGAN-Turbo with normal maps and mutual self-attention is sufficient to reconstruct a 3DGS background
    The paper evaluates the transferred images qualitatively and via one FID number; no quantitative multi-view consistency or reconstruction metric is provided.

pith-pipeline@v1.3.0-daily-deepseek · 18044 in / 15229 out tokens · 105287 ms · 2026-08-04T12:49:23.649799+00:00 · methodology

0 comments
read the original abstract

Existing rainfall simulation methods for autonomous driving remain limited in physical controllability and multi-view consistency. This paper presents GSRAIN, a high-/low-frequency rainfall synthesis method for 3D Gaussian Splatting (3DGS) driving scenes. GSRAIN constructs a high-frequency raindrop model from measured rainfall data and generates low-frequency rainy appearance using a geometry-aware single-step diffusion model. The two effects are then fused in a unified 3DGS scene, enabling rainfall-intensity control over the range of 0--13~mm/h. The proposed method achieves a Fr\'{e}chet Inception Distance (FID) of 149.09, outperforming CycleGAN-Turbo (155.71) and WeatherEdit (157.94). Object-detection and closed-loop driving experiments further show that the generated scenes expose scene-dependent performance changes of the evaluated algorithms under controllable rainfall. These results indicate that GSRAIN provides an effective approach for constructing physically controllable, repeatable, and closed-loop-compatible rainy-weather test scenes for autonomous driving.

Figures

Figures reproduced from arXiv: 2608.02177 by Fanyu Wang, Junyi Chen, Longgao Zhang.

Figure 1
Figure 1. Figure 1: Overview of the proposed GSRAIN framework. The high-frequency branch builds physically calibrated rain-streak and haze Gaussians, while the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Raindrop imaging regions and distance-dependent attenuation. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Fine-tuning pipeline of the normal-conditioned single-step diffusion model. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Cross-view consistency enhancement using feature matching and mutual self-attention. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Style-transfer comparison on BDD-100K. The columns show three driving scenes, and rows (a)–(e) show the clear inputs, CycleGAN-Turbo [37], [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Multi-view style-transfer comparison on Waymo. The columns correspond to the selected left, center, and right camera views. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ablation study of the normal-map channel. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Ablation study of mutual self-attention. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative comparison of complete rainfall synthesis. Gemini 3.1 [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Rainfall synthesis under three target intensities in two scenes. Higher intensity produces denser near-field rain streaks and stronger far-field haze. [PITH_FULL_IMAGE:figures/full_fig_p009_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Each scenario controls the rainfall level by adjusting [PITH_FULL_IMAGE:figures/full_fig_p009_12.png] view at source ↗
Figure 11
Figure 11. Figure 11: 3DGS-based closed-loop simulation and evaluation framework. [PITH_FULL_IMAGE:figures/full_fig_p010_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Closed-loop test scenarios in which the ego vehicle travels straight. [PITH_FULL_IMAGE:figures/full_fig_p010_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: SUT speed and acceleration in the cut-in scenario. [PITH_FULL_IMAGE:figures/full_fig_p010_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: SUT speed and acceleration in the stationary-leading-vehicle sce [PITH_FULL_IMAGE:figures/full_fig_p011_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Ground-truth annotations of the evaluated objects at two urban [PITH_FULL_IMAGE:figures/full_fig_p011_15.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

50 extracted references · 2 linked inside Pith

  1. [1]

    End-to-end autonomous driving: From classic paradigm to large model empowerment—A comprehensive survey,

    W. Dong, S. Lu, X. Chen, et al., “End-to-end autonomous driving: From classic paradigm to large model empowerment—A comprehensive survey,”IEEE Internet Things J., vol. 13, no. 3, pp. 3870–3898, 2025

    2025

  2. [2]

    A comprehensive review on limitations of autonomous driving and its impact on accidents and collisions,

    A. Chougule, V . Chamola, A. Sam, et al., “A comprehensive review on limitations of autonomous driving and its impact on accidents and collisions,”IEEE Open J. V eh. Technol., vol. 5, pp. 142–161, 2023

    2023

  3. [3]

    Advancing autonomous driving system testing: Demands, challenges, and future directions,

    Y . Liao, J. Zhang, J. Keung, et al., “Advancing autonomous driving system testing: Demands, challenges, and future directions,”Inf. Softw. Technol., vol. 187, Art. no. 107859, 2025

    2025

  4. [4]

    Drive&Gen: Co-evaluating end-to- end driving and video generation models,

    J. Wang, Z. Yang, Y . Bai, et al., “Drive&Gen: Co-evaluating end-to- end driving and video generation models,” inProc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), 2025, pp. 8934–8941

    2025

  5. [5]

    GeoSim: Realistic video simulation via geometry-aware composition for self-driving,

    Y . Chen, F. Rong, S. Duggal, et al., “GeoSim: Realistic video simulation via geometry-aware composition for self-driving,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2021, pp. 7230–7240

    2021

  6. [6]

    UniSim: A neural closed-loop sensor simulator,

    Z. Yang, Y . Chen, J. Wang, et al., “UniSim: A neural closed-loop sensor simulator,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, pp. 1389–1399

    2023

  7. [7]

    CARLA: An open urban driving simulator,

    A. Dosovitskiy, G. Ros, F. Codevilla, et al., “CARLA: An open urban driving simulator,” inProc. Conf. Robot Learn. (CoRL), 2017, pp. 1–16

    2017

  8. [8]

    Vision and rain,

    K. Garg and S. K. Nayar, “Vision and rain,”Int. J. Comput. Vis., vol. 75, no. 1, pp. 3–27, 2007

    2007

  9. [9]

    Rain rendering for evaluating and improving robustness to bad weather,

    M. Tremblay, S. S. Halder, R. de Charette, et al., “Rain rendering for evaluating and improving robustness to bad weather,”Int. J. Comput. Vis., vol. 129, pp. 341–360, 2021

    2021

  10. [10]

    Recent development and applications of SUMO—Simulation of Urban MObility,

    D. Krajzewicz, J. Erdmann, M. Behrisch, et al., “Recent development and applications of SUMO—Simulation of Urban MObility,”Int. J. Adv. Syst. Meas., vol. 5, nos. 3–4, pp. 128–138, 2012

    2012

  11. [11]

    MetaDrive: Composing diverse driving scenarios for generalizable reinforcement learning,

    Q. Li, Z. Peng, L. Feng, et al., “MetaDrive: Composing diverse driving scenarios for generalizable reinforcement learning,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 45, no. 3, pp. 3461–3475, 2023

    2023

  12. [12]

    DriveDreamer: Towards real-world- driven world models for autonomous driving,

    X. Wang, Z. Zhu, G. Huang, et al., “DriveDreamer: Towards real-world- driven world models for autonomous driving,” arXiv:2309.09777, 2023

    Pith/arXiv arXiv 2023

  13. [13]

    DriveDreamer-2: LLM-enhanced world models for diverse driving video generation,

    G. Zhao, X. Wang, Z. Zhu, et al., “DriveDreamer-2: LLM-enhanced world models for diverse driving video generation,” inProc. AAAI Conf. Artif. Intell., vol. 39, no. 10, 2025, pp. 10412–10420

    2025

  14. [14]

    Street-view image generation from a bird’s-eye view layout,

    A. Swerdlow, R. Xu, and B. Zhou, “Street-view image generation from a bird’s-eye view layout,”IEEE Robot. Autom. Lett., vol. 9, no. 4, pp. 3578–3585, 2024

    2024

  15. [15]

    PerLDiff: Controllable street-view synthesis using perspective-layout diffusion model,

    J. Zhang, H. Sheng, S. Cai, et al., “PerLDiff: Controllable street-view synthesis using perspective-layout diffusion model,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2025, pp. 26306–26315

    2025

  16. [16]

    Foundation models in au- tonomous driving: A survey on scenario generation and scenario analy- sis,

    Y . Gao, M. Piccinini, Y . Zhang, et al., “Foundation models in au- tonomous driving: A survey on scenario generation and scenario analy- sis,”IEEE Open J. Intell. Transp. Syst., 2026

    2026

  17. [17]

    NeRF: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. Srinivasan, M. Tancik, et al., “NeRF: Representing scenes as neural radiance fields for view synthesis,”Commun. ACM, vol. 65, no. 1, pp. 99–106, 2021

    2021

  18. [18]

    3D Gaussian splatting for real-time radiance field rendering,

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, et al., “3D Gaussian splatting for real-time radiance field rendering,”ACM Trans. Graph., vol. 42, no. 4, Art. no. 139, pp. 1–14, 2023

    2023

  19. [19]

    NeuRAD: Neural rendering for autonomous driving,

    A. Tonderski, C. Lindstr ¨om, G. Hess, et al., “NeuRAD: Neural rendering for autonomous driving,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024, pp. 14895–14904

    2024

  20. [20]

    3D Gaussian splatting: Survey, technologies, challenges, and opportunities,

    Y . Bao, T. Ding, J. Huo, et al., “3D Gaussian splatting: Survey, technologies, challenges, and opportunities,”IEEE Trans. Circuits Syst. Video Technol., vol. 35, no. 7, pp. 6832–6852, 2025

    2025

  21. [21]

    AutoSplat: Constrained Gaussian splatting for autonomous driving scene reconstruction,

    M. Khan, H. Fazlali, D. Sharma, et al., “AutoSplat: Constrained Gaussian splatting for autonomous driving scene reconstruction,” inProc. IEEE Int. Conf. Robot. Autom. (ICRA), 2025, pp. 8315–8321

    2025

  22. [22]

    HO-Gaussian: Hybrid optimization of 3D Gaussian splatting for urban scenes,

    Z. Li, Y . Zhang, C. Wu, et al., “HO-Gaussian: Hybrid optimization of 3D Gaussian splatting for urban scenes,” inProc. Eur . Conf. Comput. Vis. (ECCV), 2024, pp. 19–36

    2024

  23. [23]

    DrivingGaussian: Composite Gaussian splatting for surrounding dynamic autonomous driving scenes,

    X. Zhou, Z. Lin, X. Shan, et al., “DrivingGaussian: Composite Gaussian splatting for surrounding dynamic autonomous driving scenes,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024, pp. 21634–21643

    2024

  24. [24]

    OmniRe: Omni urban scene reconstruction,

    Z. Chen, J. Yang, J. Huang, et al., “OmniRe: Omni urban scene reconstruction,” inProc. Int. Conf. Learn. Represent. (ICLR), 2025

    2025

  25. [25]

    WeatherEdit: Controllable weather editing with 4D Gaussian field,

    C. Qian, W. Li, Y . Guo, et al., “WeatherEdit: Controllable weather editing with 4D Gaussian field,” inProc. AAAI Conf. Artif. Intell., vol. 40, no. 10, 2026, pp. 8511–8519

    2026

  26. [26]

    A new model for the equilibrium shape of raindrops,

    K. V . Beard and C. Chuang, “A new model for the equilibrium shape of raindrops,”J. Atmos. Sci., vol. 44, no. 11, pp. 1509–1524, 1987

    1987

  27. [27]

    Photorealistic rendering of rain streaks,

    K. Garg and S. K. Nayar, “Photorealistic rendering of rain streaks,”ACM Trans. Graph., vol. 25, no. 3, pp. 996–1002, 2006

    2006

  28. [28]

    Physics-based rendering for improving robustness to rain,

    S. S. Halder, J. F. Lalonde, and R. de Charette, “Physics-based rendering for improving robustness to rain,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2019, pp. 10203–10212

    2019

  29. [29]

    RainyGS: Efficient rain synthesis with physically based Gaussian splatting,

    Q. Dai, X. Ni, Q. Shen, et al., “RainyGS: Efficient rain synthesis with physically based Gaussian splatting,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2025, pp. 16153–16162

    2025

  30. [30]

    SGSST: Scaling Gaussian splatting style transfer,

    B. Galerne, J. Wang, L. Raad, et al., “SGSST: Scaling Gaussian splatting style transfer,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2025, pp. 26535–26544

    2025

  31. [31]

    StyleMe3D: Stylization with disentangled priors by multiple encoders on 3D Gaussians,

    C. Zhuang, Y . Hu, X. Zhang, et al., “StyleMe3D: Stylization with disentangled priors by multiple encoders on 3D Gaussians,” arXiv:2504.15281, 2025

    arXiv 2025

  32. [32]

    FantasyStyle: Controllable stylized distillation for 3D Gaussian splatting,

    Y . Yang, Y . Wang, C. Wang, et al., “FantasyStyle: Controllable stylized distillation for 3D Gaussian splatting,” inProc. AAAI Conf. Artif. Intell., vol. 40, no. 14, 2026, pp. 11784–11792

    2026

  33. [33]

    Tune-Your-Style: Intensity-tunable 3D style transfer with Gaussian splatting,

    Y . Zhao, R. Ye, R. Zheng, et al., “Tune-Your-Style: Intensity-tunable 3D style transfer with Gaussian splatting,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2025, pp. 19032–19042

    2025

  34. [34]

    Closing the loop: Joint rain generation and removal via disentangled image translation,

    Y . Ye, Y . Chang, H. Zhou, et al., “Closing the loop: Joint rain generation and removal via disentangled image translation,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2021, pp. 2053–2062

    2021

  35. [35]

    From rain generation to rain removal,

    H. Wang, Z. Yue, Q. Xie, et al., “From rain generation to rain removal,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2021, pp. 14791–14801

    2021

  36. [36]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, et al., “High-resolution image synthesis with latent diffusion models,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022, pp. 10684–10695

    2022

  37. [37]

    One-step image translation with text-to-image models,

    G. Parmar, T. Park, S. Narasimhan, et al., “One-step image translation with text-to-image models,” arXiv:2403.12036, 2024

    Pith/arXiv arXiv 2024

  38. [38]

    LoRA: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, et al., “LoRA: Low-rank adaptation of large language models,” inProc. Int. Conf. Learn. Represent. (ICLR), 2022

    2022

  39. [39]

    Application of the generalized gamma model to represent the full rain drop size distribution spectra,

    M. Thurai and V . N. Bringi, “Application of the generalized gamma model to represent the full rain drop size distribution spectra,”J. Appl. Meteorol. Climatol., vol. 57, no. 5, pp. 1197–1210, 2018

    2018

  40. [40]

    Estimating and exploiting the aleatoric uncertainty in surface normal estimation,

    G. Bae, I. Budvytis, and R. Cipolla, “Estimating and exploiting the aleatoric uncertainty in surface normal estimation,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2021, pp. 13137–13146

    2021

  41. [41]

    MasaCtrl: Tuning-free mutual self- attention control for consistent image synthesis and editing,

    M. Cao, X. Wang, Z. Qi, et al., “MasaCtrl: Tuning-free mutual self- attention control for consistent image synthesis and editing,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2023, pp. 22560–22570. 13

    2023

  42. [42]

    BDD100K: A diverse driving dataset for heterogeneous multitask learning,

    F. Yu, H. Chen, X. Wang, et al., “BDD100K: A diverse driving dataset for heterogeneous multitask learning,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2020, pp. 2636–2645

    2020

  43. [43]

    Gemini 3.1 Flash Image,

    Google, “Gemini 3.1 Flash Image,” accessed Apr. 18, 2026. [Online]. Available: https://ai.google.dev/gemini-api/

    2026

  44. [44]

    Waymo Open Dataset: Panoramic video panoptic segmentation,

    J. Mei, A. Zhu, X. Yan, et al., “Waymo Open Dataset: Panoramic video panoptic segmentation,” inProc. Eur . Conf. Comput. Vis. (ECCV), 2022, pp. 53–72

    2022

  45. [45]

    GANs trained by a two time-scale update rule converge to a local Nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, et al., “GANs trained by a two time-scale update rule converge to a local Nash equilibrium,”Adv. Neural Inf. Process. Syst., vol. 30, 2017

    2017

  46. [46]

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

    H. Zhou, L. Lin, J. Wang, et al., “HUGSIM: A real-time, photo-realistic and closed-loop simulator for autonomous driving,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 48, no. 4, pp. 4673–4691, 2025

    2025

  47. [47]

    Planning-oriented autonomous driving,

    Y . Hu, J. Yang, L. Chen, et al., “Planning-oriented autonomous driving,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, pp. 17853–17862

    2023

  48. [48]

    A comparison of headway and time to collision as safety indicators,

    K. V ogel, “A comparison of headway and time to collision as safety indicators,”Accid. Anal. Prev., vol. 35, no. 3, pp. 427–433, 2003

    2003

  49. [49]

    YOLOv10: Real-time end-to-end object detection,

    A. Wang, H. Chen, L. Liu, et al., “YOLOv10: Real-time end-to-end object detection,”Adv. Neural Inf. Process. Syst., vol. 37, pp. 107984– 108011, 2024

    2024

  50. [50]

    YOLO26: Key architec- tural enhancements and performance benchmarking for real-time object detection,

    R. Sapkota, R. H. Cheppally, A. Sharda, et al., “YOLO26: Key architec- tural enhancements and performance benchmarking for real-time object detection,” arXiv:2509.25164, 2025

    arXiv 2025