Pith. sign in

REVIEW 3 major objections 5 minor 32 references

SafeGen: Goal-Conditioned Video Diffusion of Safety-Critical Scenarios for VLM-Based Autonomous Driving

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read SafeGen generates safety-critical driving videos by anchoring video diffusion on a VLM-reasoned catastrophic end state, and shows these videos expose failures in vision-language driving models.

desk verdict SafeGen is a plausible assembly of existing components, but its headline 24.25% improvement rests on a VLM judge metric (JOS) that is not validated, so treat the engineering as solid and the quantitative claim as conditional. read the letter →

arxiv 2607.19701 v1 pith:CUN4EO3X submitted 2026-07-22 cs.CV

Jiangfan Liu , Zexuan Cui , Tianyuan Zhang , Zonglei Jing , Zonghao Ying , Yaoyuan Zhang , Jiakai Wang , Xiaoqi Jiang
show 2 more authors
Aishan Liu Xianglong Liu
This is my paper
classification cs.CV
keywords safety-criticalscenariogenerationgoal-conditionedvideodiffusionvision-languageautonomousdrivingend-stateconditioningvulnerableroadusersVLMjudgemetricboundary-conditionedsafetyevaluation
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 driving scenarios are better generated by reasoning backward from a chosen catastrophic end state than by simulating forward from benign conditions. It builds SafeGen, a goal-conditioned video diffusion pipeline: a vision-language model analyzes an ordinary driving video, picks a frame where a collision is plausible, and writes out a structured end-state specification; a diffusion model then creates a threatening final frame and interpolates backward to the benign first frame. Across three vision-language driving models, videos produced this way raise a VLM-judge score of reasoning misalignment by 24.25% on average over simulator-based baselines, and fine-tuning a driving model on the generated clips improves its accuracy on an unseen accident benchmark by 15.9%. The claim matters because simulator-based scenario generation suffers from a visual and behavioral gap to real driving, while this approach operates directly on real video.

What carries the argument

The load-bearing mechanism is the catastrophic end-state specification s = {Φ_sce, Φ_adv, Φ_pos | T_hat}, together with boundary-conditioned diffusion. The end state is the terminal anchor: it fixes the last frame, the adversarial agent's appearance, and the high-risk insertion region, then a diffusion model interpolates between the benign first frame and that adversarial last frame. The VLM-judge metric JOS = 0.2(10−PA) + 0.3AE + 0.5(10−PC) is the evaluation instrument that turns the generated videos into a number; PA penalizes perception misses, AE rewards anticipation errors, and PC penalizes unsafe planning, with planning weighted most heavily.

What would settle it

Take a sample of SafeGen-generated videos and score them with a jury of experienced human drivers or a closed-loop simulator that reconstructs the scene and lets a planner act; if high-JOS videos are rated non-threatening by humans or are safely avoidable by a planner, the central claim that JOS measures real safety flaws fails. A concrete check: compare per-scene JOS against reported kinematics (SafeGen videos crowd TTC below 0.5s and DRAC into 15–20 m/s²) and see whether the judge rewards scenes where kinematic urgency is genuinely extreme.

Watch

Extended reading notes

Core claim

SafeGen's central discovery is that a predefined catastrophic end-state—formalized as a scenario description, an adversarial pedestrian portrait, and a placement region tied to a chosen critical frame—acts as a strong supervisory signal for video generation. A VLM, grounded by depth and segmentation maps, identifies latent interaction vulnerabilities and specifies that end state. A text-to-image model renders the pedestrian, depth-aware projection places it at a time-to-collision-minimizing point on the ground plane, inpainting fuses it into the last frame, and boundary-conditioned diffusion evolves the intermediate frames from the original benign first frame to the adversarial last frame. E

Load-bearing premise

The load-bearing premise is that the VLM-judge score (JOS) tracks genuine safety-critical reasoning failures; if the judge is fooled by low-level visual artifacts, or if its judgments correlate with the tested models' shared inductive biases, the reported 24.25% improvement would not establish that SafeGen exposes real safety flaws.

Editorial extensions

If this is right

  • VLMAD evaluation can move from simulator-generated hazards to real-video hazards, because SafeGen operates directly on benign driving footage and produces photorealistic, physically grounded VRU insertions.
  • A single pipeline can stress perception, prediction, and planning together: across three models, perception alignment falls 65.38%, anticipation error rises 14.13%, and planning compliance falls 62.59% relative to baseline averages.
  • The catastrophic end-state anchor is the key element: removing it causes a 60.36% average drop in JOS, so any evaluator using SafeGen should preserve the endpoint constraint.
  • Fine-tuning on SafeGen-generated videos transfers to an unseen accident benchmark, improving average accuracy by 15.9% even with a viewpoint mismatch, so the generated data is not only diagnostic but also corrective.
  • Long-tail diversity is achievable in open vocabulary: generated scenes include wheelchair users, Segway users, and construction workers, categories that fixed simulator asset libraries rarely cover.

Reading between the lines

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

  • The paper reports no correlation between JOS and human- or physical-outcome safety; a natural extension would test whether high-JOS videos also elicit emergency braking or near-crash kinematics in closed-loop simulation or real driving.
  • Because the success metric is itself a VLM judge, part of the measured failure could reflect shared inductive biases between the judge and the tested models; comparing judge scores with human driver ratings on the same clips would disentangle genuine safety failures from model-to-model agreement.
  • The fine-tuning result hints at a self-improving loop: use a VLM judge to label generated QA pairs, fine-tune, then re-test with newly generated hazards—but this needs verification that the model learns genuine safety reasoning rather than recognizing generator-specific artifacts.
  • The backward-from-collision formulation could be extended to multi-agent conflicts and non-VRU hazards (vehicles, cyclists, infrastructure), which the paper lists as future work; the same end-state specification would need to encode agent-agent interactions.
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

3 major / 5 minor

Summary. The paper presents SafeGen, a goal-conditioned video diffusion framework that transforms benign driving videos into safety-critical VRU collision scenarios. A VLM first reasons over depth and segmentation priors to infer a catastrophic end state, specified as a scenario description, a VRU portrait, and a placement region; a text-to-image diffusion model then synthesizes the VRU, which is geometrically projected into the scene by minimizing a TTC criterion; finally, a boundary-conditioned video diffusion model interpolates from the benign first frame to the adversarial last frame. Experiments on three VLMADs report a 24.25% average improvement in the proposed Judge Overall Score (JOS) over baselines, a 15.9% downstream fine-tuning gain on the VRU-Accident benchmark, and a 30-driver human study.

Significance. If the JOS metric is valid, SafeGen offers a practical pipeline for exposing cascading perception/anticipation/planning failures in VLMADs and for generating transferable fine-tuning data. The paper's strengths include a clean goal-conditioned formulation, public code release, an external benchmark evaluation, a human-in-the-loop study, and ablations of the main design choices. However, the headline quantitative claims rest on a VLM judge whose scores are not validated against human judgment, kinematic ground truth, or cross-judge stability, and the reported point estimates carry no variance information. These gaps are central to the paper's stated contributions.

major comments (3)
  1. [Sec. 4.1-4.2] The headline 24.25% improvement is measured by JOS, a composite metric whose components PA, AE, and PC are assigned by an unspecified 'high-capacity VLM' judge. The paper provides no evidence that this judge tracks genuine safety-critical reasoning failures: there is no reported correlation with human threat/realism ratings, no agreement with kinematic ground truth (TTC/DRAC), no cross-judge stability check, and no sensitivity analysis for the hand-chosen weights {0.2, 0.3, 0.5}. Because the judge and the tested VLMADs are all web-scale VLMs trained on overlapping data, the judge may share the very perceptual or anticipatory blind spots that SafeGen is designed to expose; a high JOS could then reflect judge artifacts rather than VLMAD safety failures. This is load-bearing: the central claim in the abstract and Sec. 4.2 depends entirely on this unvalidated metric. Please add at least one
  2. [Table 1, Fig. 5] No error bars or repeated-run statistics are reported despite multiple sources of stochasticity: the diffusion generation, the VLM end-state reasoning, and the VLM judge are all sampled. The 24.25% improvement is a point estimate with unknown variance, and the ablations in Fig. 5 similarly report single values. Please report means and standard deviations over at least 3 independent generation/judging runs, or otherwise justify that the differences in Table 1 exceed run-to-run variability.
  3. [Sec. 4.3, Eq. (3.2)] The kinematic urgency metrics TTC and DRAC are claimed to be model-agnostic validations of scenario quality, but they are computed from the same depth-aware geometric projection used to place the VRU (Eq. 3.2 chooses the insertion point by minimizing TTC). Thus the strong TTC distribution shift below 0.5 s is partly guaranteed by construction rather than independently measured. Please provide an independent validation, e.g., human-annotated collision timing, an off-the-shelf tracker/optical-flow estimate, or at least an explicit acknowledgment that the kinematics are derived from the insertion procedure.
minor comments (5)
  1. [Sec. 4.2] The 24.25% figure is computed as the relative improvement over the average of all baseline-model combinations (85.64/12 = 7.14 → 8.87). This should be stated explicitly; a reader may otherwise interpret it as improvement over the best baseline, which gives a much smaller number.
  2. [Sec. 4.1] PA, AE, and PC are defined only by reference to the appendix. The main text needs at least a one-sentence rubric for each so that the JOS definition is self-contained.
  3. [Sec. 4.3, Fig. 4b] FID/FVD values are reported only for SafeGen; there is no comparison against the baselines in Fig. 4b. Adding the corresponding baseline distributions would make the visual-fidelity claim stronger and more interpretable.
  4. [Sec. 4.5] The human study compares SafeGen only with ScenGE, not with the other baselines, and no statistical test or effect size is reported for the Likert ratings. Please add at least a paired significance test or confidence intervals.
  5. [Sec. 4.4] The text says comprehensive hyperparameter ablations are in the supplementary material, but no supplementary material is included in the submission. Either include it or remove the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SafeGen's claims are empirical evaluations, not derivations that reduce to their inputs.

full rationale

SafeGen's central claims are empirical: a 24.25% JOS improvement and a 15.9% downstream fine-tuning improvement. JOS is a defined evaluation metric (Sec. 4.1), not a fitted parameter or a renamed input. The generation pipeline is conditioned on a specified catastrophic end-state, and the evaluation measures whether target VLMADs detect and respond to the synthesized threats; the target models' responses are not produced by the pipeline. The only self-citation is the ScenGE baseline [18], which is authors' prior work, and it is used as a comparison baseline rather than as load-bearing justification for the main claim. The downstream fine-tuning result is evaluated on the external VRU-Accident benchmark, and threat/realism are separately assessed via TTC/DRAC and a human study, providing independent evidence not reducible to the proposed method's own definitions. The possible concern that the VLM judge may share biases with tested VLMADs is a validity/calibration question, not a circularity of the kind where a result is equivalent to its inputs by construction.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim is an empirical evaluation claim, not a derivation. The main ledger entries are the assumptions that JOS (a VLM judge) measures real safety reasoning and that off-the-shelf components faithfully realize the described pipeline. No invented physical entities are introduced. One hand-chosen metric composition (JOS weights) affects every reported number.

free parameters (1)
  • JOS metric weights = 0.2, 0.3, 0.5 for (10-PA), AE, (10-PC)
    The weights in the JOS formula are hand-chosen to emphasize planning compliance; every reported headline improvement depends on this composition. No sensitivity analysis is given in the main text.
assumptions (4)
  • domain assumption A VLM judge's JOS score reflects genuine safety-critical reasoning failures of the tested VLMAD, not low-level artifacts or shared VLM prior biases.
    Central evaluation metric introduced in Sec 4.1; no correlation with human or physical safety outcomes is reported.
  • domain assumption Boundary-conditioned video diffusion (Wan2.1 FLF2V) can generate kinematically plausible intermediate motion between a benign first frame and an injected adversarial last frame.
    Assumed in Eq. (3.3) and Sec 3.3; kinematic plausibility is only indirectly supported by human ratings and TTC statistics, not by physical simulation.
  • domain assumption Baselines (LC, AS, AT, ScenGE) adapted from vehicle-vehicle to human-vehicle conflicts are fair comparators.
    Sec 4.1 states the adaptation without giving the adaptation protocol; the comparative claims depend on this fairness.
  • domain assumption Off-the-shelf components (DepthAnythingV3, MobileSAM, Qwen3-VL-32B, Stable Diffusion 3.5) perform as assumed for the safety-critical generation task.
    The pipeline is built on these pretrained models; failures in depth, segmentation, or reasoning propagate directly into scenario quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SafeGen: Goal-Conditioned Video Diffusion of Safety-Critical Scenarios for VLM-Based Autonomous Driving." pith.science (2026). https://pith.science/paper/CUN4EO3X

@misc{pith2026260719701,
  author       = {Pith},
  title        = {Pith review of: SafeGen: Goal-Conditioned Video Diffusion of Safety-Critical Scenarios for VLM-Based Autonomous Driving},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CUN4EO3X}},
  note         = {Machine review of arXiv:2607.19701}
}
read the original abstract

VLMs are increasingly deployed in AD systems, creating an urgent need for rigorous safety evaluation under rare yet safety-critical scenarios. Among these, interactions with vulnerable road users represent a major source of real-world failures. However, existing safety-critical scenario generation methods predominantly rely on simulator-based pipelines, which suffer from a substantial sim-to-real gap and often fail to capture realistic, diverse, and unforeseen human-vehicle interaction dynamics. We present SafeGen, a goal-conditioned diffusion framework for safety-critical scenario generation in VLMADs. Our key insight is to formulate scenario generation as a goal-conditioned diffusion process, where a predefined catastrophic end-state serves as a strong supervisory signal, guiding the generation of temporally coherent video trajectories that naturally evolve toward safety-critical outcomes. Building on this formulation, we introduce Context Grounded End State Reasoning, which leverages VLMs to analyze benign driving contexts and infer latent vulnerabilities in human-vehicle interactions, producing structured end-state specifications that induce high-risk scenarios. Conditioned on these targets, we further propose End State Conditioned Video Evolution, which grounds semantic threats into physically plausible visual dynamics. Specifically, we instantiate high-risk agents within the scene via depth-aware geometric projection, followed by boundary-conditioned diffusion to generate intermediate frames with consistent motion patterns and temporal coherence. Extensive experiments across 3 VLMADs demonstrate that SafeGen increases the Judge Overall Score, a metric using a VLM judge to evaluate VLMADs' understanding and decision-making, by 24.25% on average compared to SoTA baselines. Furthermore, fine-tuning a VLMAD improves performance in real-world driving scenes by an average of 15.9%.

Figures

Figures reproduced from arXiv: 2607.19701 by the authors.

Figure 1
Figure 1. Given a benign driving video, SafeGen can synthe￾size a physically plausible and temporally coherent safety￾critical scenario video for VLMAD evaluation. improves performance in real-world driving scenes by an average of 15.9%. Our code is available at https://github.com/JoFrc/SafeGen. CCS Concepts • Security and privacy → Social aspects of security and pri￾vacy; • Computing methodologies → Scene understanding. Keyw… view at source ↗
Figure 2
Figure 2. Overview of the proposed SafeGen framework for safety-critical scenario video generation. This spatial optimization process is formulated as: P𝑎𝑑𝑣 = arg min P∈Φpos∩Mground 𝑓𝑑𝑒𝑝 (I𝑇ˆ ) (P) 𝑣𝑒𝑔𝑜 , (3.2) where 𝑓𝑑𝑒𝑝 (I𝑇ˆ ) (P) extracts the localized depth value at the spatial coordinate P, and 𝑣𝑒𝑔𝑜 represents the ego-vehicle velocity. Utiliz￾ing the background-free portrait A𝑎𝑑𝑣 and the spatial geometry derived from the… view at source ↗
Figure 3
Figure 3. Safety-critical scenario videos generated by [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Threat and realism analysis including kinematic [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Impact of generation conditioning mechanisms [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Example frame from VRU-Accident, a dataset con [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Human evaluation, with simulator setup on the left [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 7 linked inside Pith

  1. [1]

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan ...

    arXiv 2026

  2. [2]

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Leo Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. 2024. Video generation models as world simulators.OpenAI Blog1, 8 (2024), 1

    2024

  3. [3]

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. 2020. nuscenes: A multimodal dataset for autonomous driving. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 11621–11631

    2020

  4. [4]

    Yulong Cao, Chaowei Xiao, Anima Anandkumar, Danfei Xu, and Marco Pavone

  5. [5]

    Baiming Chen, Xiang Chen, Qiong Wu, and Liang Li. 2021. Adversarial evaluation of autonomous vehicles in lane-change scenarios.IEEE Transactions on Intelligent Transportation Systems23, 8 (2021), 10333–10342

    2021

  6. [6]

    Wenhao Ding, Baiming Chen, Minjun Xu, and Ding Zhao. 2020. Learning to collide: An adaptive safety-critical scenarios generating method. In2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2243– 2250

    2020

  7. [7]

    Wenhao Ding, Chejian Xu, Mansur Arief, Haohong Lin, Bo Li, and Ding Zhao

  8. [8]

    Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. 2017. CARLA: An Open Urban Driving Simulator. InProceedings of the 1st Annual Conference on Robot Learning. 1–16

    2017

Show all 32 references
  1. [9]

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. 2024. Scaling rectified flow transformers for high-resolution image synthesis. InForty- first International Conference on Ma...

    2024

  2. [10]

    Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. 2024. MagicDrive: Street View Generation with Diverse 3D Geometry Control. InInternational Conference on Learning Representations

    2024

  3. [11]

    Akshay Gopalkrishnan, Ross Greer, and Mohan Trivedi. 2024. Multi-frame, light- weight & efficient vision-language models for question answering in autonomous driving.arXiv preprint arXiv:2403.19838(2024)

    2024 arXiv

  4. [12]

    Wenyi Hong, Wenmeng Yu, Xiaotao Gu, Guo Wang, Guobing Gan, Haomiao Tang, Jiale Cheng, Ji Qi, Junhui Ji, Lihang Pan, et al. 2025. Glm-4.5 v and glm-4.1 v-thinking: Towards versatile multimodal reasoning with scalable reinforcement learning.arXiv preprint arXiv:2507.01006(2025)

    2025 arXiv

  5. [13]

    Ayesha Ishaq, Jean Lahoud, Ketan More, Omkar Thawakar, Ritesh Thawkar, Dinura Dissanayake, Noor Ahsan, Yuhao Li, Fahad Shahbaz Khan, Hisham Cholakkal, et al . 2025. DriveLLM-o1: A step-by-step reasoning dataset and large multimodal model for driving scenario understanding. In2...

    2025

  6. [14]

    Nidhi Kalra and Susan M Paddock. 2016. Driving to safety: How many miles of driving would it take to demonstrate autonomous vehicle reliability?Transporta- tion research part A: policy and practice94 (2016), 182–193

    2016

  7. [15]

    Abdelrahman, and Mohamed Abdel-Aty

    Younggun Kim, Ahmed S. Abdelrahman, and Mohamed Abdel-Aty. 2025. VRU- Accident: A Vision-Language Benchmark for Video Question Answering and Dense Captioning for Accident Scene Understanding. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Work...

    2025

  8. [16]

    Quanyi Li, Zhenghao Peng, Lan Feng, Qihang Zhang, Zhenghai Xue, and Bolei Zhou. 2023. MetaDrive: Composing Diverse Driving Scenarios for Generalizable Reinforcement Learning.IEEE Transactions on Pattern Analysis and Machine Intelligence45, 3 (2023), 3461–3475

    2023

  9. [17]

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

    2025 arXiv

  10. [18]

    Jiangfan Liu, Yongkang Guo, Fangzhi Zhong, Tianyuan Zhang, Zonglei Jing, Siyuan Liang, Jiakai Wang, Mingchuan Zhang, Aishan Liu, and Xianglong Liu

  11. [19]

    Yingzi Ma, Yulong Cao, Jiachen Sun, Marco Pavone, and Chaowei Xiao. 2024. Dolphins: Multimodal language model for driving. InEuropean Conference on Computer Vision. Springer, 403–420

    2024

  12. [20]

    Matthew O’Kelly, Aman Sinha, Hongseok Namkoong, Russ Tedrake, and John C Duchi. 2018. Scalable end-to-end autonomous vehicle testing via rare-event simulation.Advances in neural information processing systems31 (2018)

    2018

  13. [21]

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. 2025. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314(2025)

    2025 arXiv

  14. [22]

    Jingkang Wang, Ava Pun, James Tu, Sivabalan Manivasagam, Abbas Sadat, Sergio Casas, Mengye Ren, and Raquel Urtasun. 2021. Advsim: Generating safety-critical scenarios for self-driving vehicles. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition...

    2021

  15. [23]

    Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, et al. 2025. InternVL3.5: Ad- vancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency. arXiv preprint arXiv:2508.18265(2025)

    2025 arXiv

  16. [24]

    Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu

  17. [25]

    Hao Xiang, Runsheng Xu, Xin Xia, Zhaoliang Zheng, Bolei Zhou, and Jiaqi Ma. 2023. V2xp-asg: Generating adversarial scenes for vehicle-to-everything perception. In2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 3584–3591

    2023

  18. [26]

    Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. 2023. Faster segment anything: Towards lightweight sam for mobile applications.arXiv preprint arXiv:2306.14289(2023)

    2023 arXiv

  19. [27]

    Lvmin Zhang and Maneesh Agrawala. 2025. Packing input frame context in next-frame prediction models for video generation.arXiv preprint: 2504.12626 (2025)

    2025

  20. [28]

    Qingzhao Zhang, Shengtuo Hu, Jiachen Sun, Qi Alfred Chen, and Z Morley Mao. 2022. On adversarial robustness of trajectory prediction for autonomous vehicles. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 15159–15168

    2022

  21. [2022]

    InEuropean Conference on Computer Vision

    Advdo: Realistic adversarial attacks for trajectory prediction. InEuropean Conference on Computer Vision. Springer, 36–52

  22. [2023]

    A survey on safety-critical driving scenario generation—a methodological perspective.IEEE Transactions on Intelligent Transportation Systems24, 7 (2023), 6971–6988

    2023

  23. [2024]

    InEuropean Conference on Computer Vision

    DriveDreamer: Towards Real-World-Drive World Models for Autonomous Driving. InEuropean Conference on Computer Vision. Springer, 55–72

  24. [2026]

    InProceedings of the AAAI Conference on Artificial Intelligence, Vol

    Adversarial generation and collaborative evolution of safety-critical scenar- ios for autonomous vehicles. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 38926–38934

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.