Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

SafeMVDrive: Multi-view Safety-Critical Driving Video Synthesis in the Real World Domain

T0 review · 5 major / 6 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 SafeMVDrive generates realistic, multi-view, safety-critical driving videos from real-world scenes and reports that the UniAD planner's sample-level collision rate jumps from 0.004 on original clips to 0.202 on its generated clips.

desk verdict A genuinely new multi-view safety-critical video pipeline with a smart evasion-trajectory trick, but the empirical claims need a matched-set and artifact-control pass before I'd trust the numbers. read the letter →

arxiv 2505.17727 v1 pith:7ZG7UOFL submitted 2025-05-23 cs.CV

Jiawei Zhou , Linye Lyu , Zhuotao Tian , Cheng Zhuo , Yu Li This is my paper
classification cs.CV
keywords safety-criticalscenariogenerationmulti-viewdrivingvideosynthesisend-to-endautonomousadversarialvehicleselectionvision-languagemodelGRPOfine-tuningdiffusiontrafficsimulationcollision-evasiontrajectory
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

The paper tries to establish that safety-critical, multi-view driving videos can be synthesized from real-world scenes rather than from simulators, and that these videos are realistic enough and adversarial enough to stress-test end-to-end autonomous-driving planners. It combines a trajectory generator with a multi-view video generator, and it solves the integration problem by adding visual context to the adversarial-vehicle selection and by converting collision trajectories into collision-evasion trajectories the video model can render. The authors report that the UniAD planner's sample-level collision rate rises from 0.004 on original NuScenes clips to 0.202 on SafeMVDrive clips, while human raters judge the generated videos nearly as realistic as the originals. If correct, this gives the community a scalable way to test and improve vision-based planners on rare, dangerous events without collecting such events in the real world.

What carries the argument

The load-bearing object is a two-stage evasion trajectory generator built on a controllable diffusion-based traffic simulator. Stage one guides a chosen adversarial vehicle into a collision with the ego vehicle using a time-decayed adversarial loss that is set to zero after impact; stage two re-runs the same initial scene but updates only the ego vehicle's trajectory, so the ego evades while other vehicles keep their aggressive collision-stage trajectories. This converts a would-be collision into a safety-critical near-miss that video models can render. A second load-bearing component is the VLM-based adversarial vehicle selector: a vision-language model fine-tuned with GRPO on automatically generated collision annotations, which chooses which vehicle to make aggressive based on multi-view images rather than non-visual heuristics. A third component is the multi-view video generator that converts the final trajectories into control signals (3D boxes, HD maps, camera conditions) and rolls out 9-second videos autoregressively.

What would settle it

Run an end-to-end planner on two video sets built from the same 41 initial scenes—one with original NuScenes trajectories and one with SafeMVDrive's evasion trajectories—using a planner architecture different from UniAD; if the collision rate does not rise significantly on the SafeMVDrive clips, the claimed stress-testing effect is specific to UniAD rather than a property of the generated data.

Watch

Extended reading notes

Core claim

The central claim is that SafeMVDrive is the first framework to produce high-quality, multi-view, safety-critical driving videos in the real-world domain, and that these videos actually challenge an end-to-end planner. The key result is the planner collision-rate jump: sample-level collision rate goes from 0.004 on original NuScenes videos to 0.202 on SafeMVDrive videos, and scene-level collision rate from 0.065 to 3.378, while the realism score from a 2AFC user study (0.560) stays close to the Origin set (0.633). The paper argues that the safety-criticality comes from planned collision-evasion trajectories rather than from visual crashes, because current multi-view video generators degrade when asked to render actual collisions. The evidence also includes the VLM-based selector beating heuristic baselines on F1 (0.710 over 0.654, 0.600, and 0.507) and the two-stage simulation beating collision-only trajectories on both realism and scene-level collision rate.

Load-bearing premise

The pipeline's ground truth is the simulator's verdict on which vehicles can realistically collide with the ego vehicle, and that verdict is used both to train the VLM and to score its accuracy; if the simulator's collision-ability labels do not transfer to real traffic, the whole chain of selection and measured safety-criticality is built on an unvalidated proxy.

Editorial extensions

If this is right

  • SafeMVDrive provides a scalable pipeline for generating safety-critical multi-view video data without collecting dangerous real-world events, which can be used to evaluate and fine-tune vision-based end-to-end planners.
  • The collision-evasion design implies that video models that cannot render collisions can still produce safety-critical data, as long as the trajectory encodes the danger.
  • The GRPO-finetuned VLM selector beats heuristic, random, and SFT baselines on F1, so visual-context selection is a workable replacement for kinematic-rule heuristics in adversarial vehicle choice.
  • The release of the 41-scene multi-view safety-critical video dataset gives the community an initial benchmark for planner robustness in this format.

Reading between the lines

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

  • In our reading, the method's central risk is that the simulator-generated collision annotations are used both as the training signal and as the evaluation reference for the VLM selector, so the reported F1 scores measure agreement with a simulator rather than with real-world accident likelihood.
  • A practical consequence the authors do not explore: the same two-stage evasion pipeline could be applied to other diffusion-based video backbones, and the collision-rate gain would likely vary with the backbone's ability to render near-miss dynamics, meaning the framework is a recipe rather than a fixed dataset.
  • We would expect the collision-evasion trajectories to be most useful for planners that react to visual near-miss cues; planners with strong trajectory prediction may not see the same collision-rate increase, because the evasion stage keeps the ego vehicle in control rather than forcing a crash.
  • An extension we see as natural: using the same selector and two-stage generator on datasets with varied weather and lighting (e.g., Waymo or Argoverse2) would test whether the reported realism and collision-rate effects hold outside NuScenes.
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

5 major / 6 minor

Summary. The paper introduces SafeMVDrive, a framework for generating multi-view, real-world-domain safety-critical driving videos. It combines a GRPO-finetuned VLM-based adversarial vehicle selector, a two-stage collision-evasion trajectory generator built on the CTG diffusion simulator, and the UniMLVG multi-view video generator. The authors claim that this is the first framework to produce high-quality, multi-view safety-critical driving videos from real data, and they evaluate it by measuring UniAD planner collision rates and video realism on NuScenes-derived videos. The main quantitative claim is a '30% increase in collision rate compared to the original NuScenes videos,' which Table 1 reports as sample-level CR rising from 0.004 to 0.202. The paper also reports ablations, a user study, and a VLM-selector comparison against heuristic baselines.

Significance. The research direction is timely and valuable: end-to-end autonomous-driving systems need realistic multi-view safety-critical video data for stress-testing, and a working generative framework would be a useful benchmark and data source. The paper has clear strengths: it provides public code and datasets, evaluates with a downstream planner (UniAD), includes human evaluation and extensive ablations, and its two-stage evasion idea directly addresses the known difficulty of video generators rendering collisions. However, the central empirical claims are not yet reliably established. The headline numerical claim is inconsistent with the reported table, the main method comparison uses unmatched evaluation sets, the VLM selector is trained and evaluated on the same auto-generated labels, and no evidence shows that the planner's collisions are caused by the intended safety-critical events rather than by rendering artifacts. These are fixable concerns, but they are load-bearing for the paper's central message.

major comments (5)
  1. [Section 1 (page 3) and Table 1] The abstract and introduction state that SafeMVDrive produces a '30% increase in collision rate compared to the original NuScenes videos,' but Table 1 reports sample-level CR of 0.004 on Origin versus 0.202 on SafeMVDrive, which is roughly a 50-fold relative increase (about 0.198 absolute), not a 30% increase. This is a substantial numerical inconsistency that directly affects the paper's central claim. In addition, the 'Origin' condition is not the original NuScenes video data; it is NuScenes trajectories re-rendered by UniMLVG. The paper should clarify this distinction and either correct the '30%' statement or explain precisely which comparison it refers to (e.g., a specific time horizon or scene-level metric).
  2. [Appendix E and Table 1] The comparison in Table 1 is not apples-to-apples: SafeMVDrive and Origin collision rates are computed on 41 videos from the same 41 initial scenes, while Naive collision rates are computed on 72 videos generated from a different set of scenes (all 250 base samples). Consequently, the difference between SafeMVDrive and Naive (sample-level CR 0.202 vs 0.267, scene-level CR 3.378 vs 1.815) could be due to scene selection or per-scene difficulty rather than method quality. The authors should evaluate all methods on the same set of initial scenes, or at least report matched-scene results and per-scene statistics, before drawing conclusions about the relative performance of Naive and SafeMVDrive.
  3. [Section 3.2 and Table 2] The VLM selector is trained using GRPO rewards derived from the automated CTG-based annotation (Scoll) and is then evaluated against the same annotation in Table 2. This is circular as a measure of selection quality: the F1-score measures agreement with the simulator's collision-ability judgments, which are never independently validated against real-world traffic or against the final videos' ability to induce planner failures. The paper should add an independent evaluation, for example a human-annotated subset (say 50 scenes) labeled by driving experts, or an evaluation of whether the VLM-selected vehicles are those that actually produce safety-critical events in the generated videos that cause planner collisions.
  4. [Section 4.2 and Section 3.4] The central claim that the elevated UniAD collision rate demonstrates the creation of semantically safety-critical scenarios is not backed by a check that the collisions are caused by the intended near-miss events. Because the two-stage generator produces collision-evasion trajectories without actual contact, a higher CR could result from rendering artifacts such as object warping, flicker, wrong box-to-pixel alignment, or hallucinated objects that degrade UniAD's perception. The Naive row in Table 1 (sample-level CR 0.267 with realism 0.207) is direct evidence that CR and rendering artifacts can be entangled. The authors should provide a spatiotemporal analysis: for each planner collision, verify that the colliding object corresponds to the adversarial vehicle and occurs near the intended evasion time and location, for example by comparing the planner's detected boxes against the rendered trajectory, or by reporting CR on a subset where the adversarial vehicle is confirmed to be visible and to follow its planned path.
  5. [Tables 1 and 4; Appendix A] The headline CR numbers are reported without confidence intervals, significance tests, or per-scene variance. With only 41 SafeMVDrive videos and small absolute collision probabilities, the difference between 0.004 and 0.202 needs error bars or a paired test (e.g., Wilcoxon signed-rank on the same scenes) to rule out sampling noise. Similarly, the user study in Appendix A aggregates 660 answers from 22 participants but reports no inter-rater agreement or per-comparison variance. Please add bootstrap confidence intervals for the CR estimates and realism scores, and report the standard error or another dispersion measure.
minor comments (6)
  1. [Contributions (page 3)] The sentence 'The dataset contains 41 diverse 9-second scenes and can serves as a valuable benchmark' has a subject-verb agreement error: 'can serves' should be 'can serve'.
  2. [Appendix B] There are typos in Appendix B: 'agenets' should be 'agents', and 'utlilize' should be 'utilize'.
  3. [Figure 1] The caption lists 'scene-0802_collision_inference_adv_2_startFrameIndex_71' twice; the duplicate entry should be removed.
  4. [Equation (6)] The definition of N in Eq. (6) as 'N = t/0.5' is ambiguous; please clarify that N is the floor of t/0.5 or the number of planning steps within the horizon, and define I_i consistently as the collision indicator at each planning step.
  5. [Table 3 and Appendix G] In Table 3, the last row has '0.6750.710' without a separating space; fix the formatting. In Appendix G, the list of tables is presented out of numerical order ('Table 6, 8, 7, 9, 10, and 11'); order them numerically.
  6. [Section 3.2] The notation for the adversarial vehicle is inconsistent in places (Vadv vs V_adv); please standardize the math notation throughout the paper.

Circularity Check

1 steps flagged · score 4.0 of 10

VLM selector is trained and scored against the same CTG-generated labels, making its accuracy evaluation circular; the planner-collision benchmark is independent and keeps the central claim partly grounded.

  1. fitted input called prediction [Section 3.2 (Eq. 2) and Section 4.3 (Table 2)]
    "we propose an automated method using a controllable diffusion-based traffic simulator [38] ... This yields annotated data defining the set of effective safety-critical vehicles Scoll for each scene. ... RAcc = 1 if Scoll ̸= ∅ ∧ extract_answer(O) ∈ Scoll ... On 250 validation scenes, we use automated annotation to identify all vehicles that can collide with the ego vehicle. We compare precision, recall, and F1-score of our VLM-based selector."

    The VLM is fine-tuned with GRPO whose accuracy reward (Eq. 2) is scored against Scoll, the CTG-generated collision-feasibility labels. Table 2 then evaluates the VLM's precision/recall against the same automated CTG annotation used to create those training labels. The reported 'accuracy' is therefore agreement with the paper's own label generator, not with an independent real-world notion of safety-criticality; the VLM is effectively fitted to the CTG labeler and then measured against that same labeler on held-out scenes. This makes the selector evaluation self-referential, although it does not by itself force the downstream planner-collision result.

full rationale

The central claim that SafeMVDrive produces safety-critical multi-view videos is supported by an external benchmark: UniAD collision rate on the generated videos (Table 1) and a human 2AFC realism study. Those measurements do not reduce by construction to the method's parameters, so the core derivation is not circular. The one circular element is the VLM selector evaluation: both the GRPO reward and the Table 2 ground truth are produced by the same controllable diffusion traffic simulator (CTG) from [38], making the precision/recall numbers a self-consistency check of the fitted labeler rather than independent validation. The paper is transparent about this automated annotation, but it still overstates the selector's ability to 'accurately identify safety-critical vehicles' in real scenes. The Limitations section's admission that the framework lacks dynamic filtering of outdated guidance signals is a scope limitation, not a circularity. A separate numerical inconsistency (abstract/introduction claim of a '30% increase' versus Table 1's sample-level CR moving from 0.004 to 0.202) is a reporting error, not a circular step. Overall, the independent UniAD and human-study evidence keeps the central safety-criticality claim from reducing to the fitted inputs; the circularity burden is moderate, hence a score of 4.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim depends on several hand-chosen hyperparameters and on the reliability of external pretrained components (CTG, UniMLVG, Qwen2.5-VL, UniAD). The most consequential assumption is that the simulation-based annotation is valid ground truth for both training and evaluating the VLM selector.

free parameters (5)
  • distance threshold D = 25 m
    Hand-chosen in Section 4.1; defines which vehicles are candidates for VLM selection and therefore affects all downstream selection and trajectory generation.
  • collision-stage loss weights alpha, beta, gamma = alpha=1, beta=50, gamma=1
    Chosen in Section 4.1; control the balance between adversarial, no-collision and on-road losses in the collision trajectory simulation (Eq. 4).
  • evasion-stage loss weights beta, gamma = beta=1, gamma=1
    Chosen in Section 4.1; control evasion trajectory realism and drivability (Eq. 5).
  • time-decay factor lambda = 0.9
    Chosen in Section 4.1 and justified by the ablation in Appendix G; weights earlier trajectory predictions across the loss terms.
  • distance penalty thresholds d_penalty and d_ij_penalty = not numerically specified
    Appear in Eq. 3 and Appendix B Eq. 7 as thresholds for collision avoidance; values are not reported, so the loss behavior is not fully reproducible.
assumptions (5)
  • domain assumption The controllable diffusion traffic simulator CTG generates naturalistic trajectories and reliable collision outcomes.
    Used in Section 3.2 to automatically annotate which vehicles can collide with the ego vehicle, and these labels train and evaluate the VLM.
  • domain assumption The multi-view video generator UniMLVG renders control signals into videos that preserve the planned safety-critical dynamics.
    The whole framework (Section 3.4) depends on UniMLVG faithfully realizing the collision-evasion trajectories in visual output; the paper only indirectly checks this via user studies and planner collision rates.
  • domain assumption The UniAD planner's collision rate on generated videos is a meaningful measure of safety-criticality for end-to-end systems.
    Used in Section 4.2 to validate the videos; this assumes planner behavior on synthetic videos transfers to real-world risk.
  • domain assumption NuScenes provides accurate 3D boxes, HD maps, and camera parameters needed for trajectory and video generation.
    All experiments use NuScenes as the real-world domain (Section 4.1).
  • domain assumption GRPO reinforcement learning improves VLM reasoning for spatial multi-view selection as claimed in VLM-R1 style training.
    The method inherits the RL framework from [21; 22] and assumes it transfers to the adversarial vehicle selection task.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SafeMVDrive: Multi-view Safety-Critical Driving Video Synthesis in the Real World Domain." pith.science (2026). https://pith.science/paper/7ZG7UOFL

@misc{pith2026250517727,
  author       = {Pith},
  title        = {Pith review of: SafeMVDrive: Multi-view Safety-Critical Driving Video Synthesis in the Real World Domain},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7ZG7UOFL}},
  note         = {Machine review of arXiv:2505.17727}
}
read the original abstract

Safety-critical scenarios are rare yet pivotal for evaluating and enhancing the robustness of autonomous driving systems. While existing methods generate safety-critical driving trajectories, simulations, or single-view videos, they fall short of meeting the demands of advanced end-to-end autonomous systems (E2E AD), which require real-world, multi-view video data. To bridge this gap, we introduce SafeMVDrive, the first framework designed to generate high-quality, safety-critical, multi-view driving videos grounded in real-world domains. SafeMVDrive strategically integrates a safety-critical trajectory generator with an advanced multi-view video generator. To tackle the challenges inherent in this integration, we first enhance scene understanding ability of the trajectory generator by incorporating visual context -- which is previously unavailable to such generator -- and leveraging a GRPO-finetuned vision-language model to achieve more realistic and context-aware trajectory generation. Second, recognizing that existing multi-view video generators struggle to render realistic collision events, we introduce a two-stage, controllable trajectory generation mechanism that produces collision-evasion trajectories, ensuring both video quality and safety-critical fidelity. Finally, we employ a diffusion-based multi-view video generator to synthesize high-quality safety-critical driving videos from the generated trajectories. Experiments conducted on an E2E AD planner demonstrate a significant increase in collision rate when tested with our generated data, validating the effectiveness of SafeMVDrive in stress-testing planning modules. Our code, examples, and datasets are publicly available at: https://zhoujiawei3.github.io/SafeMVDrive/.

Figures

Figures reproduced from arXiv: 2505.17727 by the authors.

Figure 1
Figure 1. Keyframes from diverse realistic, multi-view, safety-critical videos generated by [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The SafeMVDrive framework for generating realism, multi-view safety-critical videos. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Comparison between the real-world scene (left) and the BEV-rendered non-visual data [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Comparison of videos generated by different methods, only showing front view. Origin is [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Adversarial vehicle selection examples using the GRPO-finetuned VLM. The VLM [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: The questionnaire used to evaluate the realism of videos generated by different baselines in [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: In the prompt, v is substituted with the ego vehicle’s velocity in the given scene. For the non-finetuned VLM, we append “Output final answer (number) in <answer> </answer> tags.” at the end to ensure it outputs the vehicle ID for evaluation. For the GRPO-finetuned VLM…
Figure 7
Figure 7. Figure 7: Original Prompt used in our VLM-based selection. [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

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

    cs.CV 2026-08 conditional novelty 6.0 of 10

    muSync-GS couples weather and road-shape edits in driving videos to a calibrated vehicle-dynamics model, so the synthesized ego motion and telemetry change with the same controls that drive the visual edits.

Reference graph

Works this paper leans on

39 extracted references · 25 canonical work pages · cited by 1 Pith paper

  1. [1]

    Yasasa Abeysirigoonawardena, Florian Shkurti, and Gregory Dudek. 2019. Generating adversar- ial driving scenarios in high-fidelity simulators. In 2019 International Conference on Robotics and Automation (ICRA), pages 8271–8277. IEEE

    work page 2019

  2. [2]

    Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, et al. 2024. Lumiere: A space-time diffusion model for video generation. In SIGGRAPH Asia 2024 Conference Papers, pages 1–11

    work page 2024

  3. [3]

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Do- minik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. 2023. Stable video dif- fusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127

    arXiv 2023

  4. [4]

    Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. 2023. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22563–22575

    work page 2023

  5. [5]

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. 2020. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621–11631

    work page 2020

  6. [6]

    Wei-Jer Chang, Francesco Pittaluga, Masayoshi Tomizuka, Wei Zhan, and Manmohan Chan- draker. 2024. Safe-sim: Safety-critical closed-loop traffic simulation with diffusion-controllable adversaries. In Computer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Proceedings, Part XXI, pages 242–258, Berlin, Heidelberg. ...

    work page 2024

  7. [7]

    Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. 2024. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 14455–14465

    work page 2024

  8. [8]

    Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. 2023. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512

    arXiv 2023

Show all 39 references
  1. [9]

    Rui Chen, Zehuan Wu, Yichen Liu, Yuxin Guo, Jingcheng Ni, Haifeng Xia, and Siyu Xia. 2024. Unimlvg: Unified framework for multi-view long video generation with comprehensive control capabilities for autonomous driving

    2024

  2. [10]

    Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. 2017. CARLA: An open urban driving simulator. In Proceedings of the 1st Annual Conference on Robot Learning, pages 1–16

    2017

  3. [11]

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

    2024

  4. [12]

    Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. 2024. Factorizing text-to-video generation by explicit image conditioning. In European Conference on Computer Vision, pages 205–224....

    2024

  5. [13]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

    2025 arXiv

  6. [14]

    Namgyu Ho, Laura Schmid, and Se-Young Yun. 2022. Large language models are reasoning teachers. arXiv preprint arXiv:2212.10071. 10

    2022 arXiv

  7. [15]

    Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. 2023. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on...

    2023

  8. [16]

    Bo Jiang, Shaoyu Chen, Bencheng Liao, Xingyu Zhang, Wei Yin, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. 2024. Senna: Bridging large vision-language models and end-to-end autonomous driving. arXiv preprint arXiv:2410.22313

    2024 arXiv

  9. [17]

    Cheng Li, Keyuan Zhou, Tong Liu, Yu Wang, Mingqiao Zhuang, Huan-ang Gao, Bu Jin, and Hao Zhao. 2025. Avd2: Accident video diffusion for accident video description. arXiv preprint arXiv:2502.14801

    2025 arXiv

  10. [18]

    Zhenxin Li, Kailin Li, Shihao Wang, Shiyi Lan, Zhiding Yu, Yishen Ji, Zhiqi Li, Ziyue Zhu, Jan Kautz, Zuxuan Wu, et al. 2024. Hydra-mdp: End-to-end multimodal planning with multi-target hydra-distillation. arXiv preprint arXiv:2406.06978

    2024 arXiv

  11. [19]

    Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. 2024. Is ego status all you need for open-loop end-to-end autonomous driving? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14864–14873

    2024

  12. [20]

    Bencheng Liao, Shaoyu Chen, Haoran Yin, Bo Jiang, Cheng Wang, Sixu Yan, Xinbang Zhang, Xiangyu Li, Ying Zhang, Qian Zhang, et al. 2024. Diffusiondrive: Truncated diffusion model for end-to-end autonomous driving. arXiv preprint arXiv:2411.15139

    2024 arXiv

  13. [21]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300

    2024 arXiv

  14. [22]

    Haozhan Shen, Peng Liu, Jingcheng Li, Chunxin Fang, Yibo Ma, Jiajia Liao, Qiaoli Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, Ruochen Xu, and Tiancheng Zhao. 2025. Vlm-r1: A stable and generalizable r1-style large vision-language model. arXiv preprint arXiv:2504.07615

    2025 arXiv

  15. [23]

    Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beißwenger, Ping Luo, Andreas Geiger, and Hongyang Li. 2024. Drivelm: Driving with graph visual question answering. In European Conference on Computer Vision , pages 256–274. Springer

    2024

  16. [24]

    Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. 2020. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on compu...

    2020

  17. [25]

    Qwen Team. 2025. Qwen2.5-vl

    2025

  18. [26]

    Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. 2018. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717

    2018 arXiv

  19. [27]

    Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. 2025. Lavie: High-quality video generation with cascaded latent diffusion models. International Journal of Computer Vision, 133(5):3059–3078

    2025

  20. [28]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837

    2022

  21. [29]

    Yuqing Wen, Yucheng Zhao, Yingfei Liu, Fan Jia, Yanhui Wang, Chong Luo, Chi Zhang, Tiancai Wang, Xiaoyan Sun, and Xiangyu Zhang. 2024. Panacea: Panoramic and controllable video generation for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and ...

    2024

  22. [30]

    Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, et al

  23. [31]

    Jay Zhangjie Wu, Guian Fang, Haoning Wu, Xintao Wang, Yixiao Ge, Xiaodong Cun, David Jun- hao Zhang, Jia-Wei Liu, Yuchao Gu, Rui Zhao, et al. 2024. Towards a better metric for text-to-video generation. arXiv preprint arXiv:2401.07781

    2024 arXiv

  24. [32]

    Chejian Xu, Aleksandr Petiushko, Ding Zhao, and Bo Li. 2025. Diffscene: Diffusion-based safety-critical scenario generation for autonomous vehicles. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 8797–8805

    2025

  25. [33]

    Jiawei Zhang, Chejian Xu, and Bo Li. 2024. Chatscene: Knowledge-enabled safety-critical scenario generation for autonomous vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15459–15469

    2024

  26. [34]

    Wenzhao Zheng, Ruiqi Song, Xianda Guo, Chenming Zhang, and Long Chen. 2024. Genad: Generative end-to-end autonomous driving. In European Conference on Computer Vision, pages 87–104. Springer

    2024

  27. [35]

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. 2024. Llamafactory: Unified efficient fine-tuning of 100+ language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguis- tics (Volume 3:...

    2024

  28. [36]

    Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. 2024. Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404

    2024 arXiv

  29. [37]

    Ziyuan Zhong, Davis Rempe, Yuxiao Chen, Boris Ivanovic, Yulong Cao, Danfei Xu, Marco Pavone, and Baishakhi Ray. 2023. Language-guided traffic simulation via scene-level diffusion. In 7th Annual Conference on Robot Learning

    2023

  30. [38]

    uncertain

    Ziyuan Zhong, Davis Rempe, Danfei Xu, Yuxiao Chen, Sushant Veer, Tong Che, Baishakhi Ray, and Marco Pavone. 2023. Guided conditional diffusion for controllable traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3560–3566. 12 Figu...

    2023

  31. [2023]

    arXiv preprint arXiv:2301.00493

    Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493

    arXiv

Pith tools

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