Pith. sign in

REVIEW 4 major objections 5 minor 96 references

Causal-Entity Reflected Egocentric Traffic Accident Video Synthesis

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Gaze-guided diffusion edits the right object into crash videos

desk verdict Useful dataset and a coherent editing pipeline, but the headline causal-sensitivity claim rests on a metric that shares its ground truth with the training signal. read the letter →

arxiv 2506.23263 v1 pith:NHAXVABZ submitted 2025-06-29 cs.CV

Lei-lei Li , Jianwu Fang , Junbin Xiao , Shanmin Pang , Hongkai Yu , Chen Lv , Jianru Xue , Tat-Seng Chua This is my paper · ORCID
classification cs.CV
keywords causalentitygroundingvideodiffusiondrivergazeegocentrictrafficaccidentcounterfactualeditingnormal-to-accidentgenerationtext-to-videoattentiondataset
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

This paper claims that egocentric traffic accident videos can be synthesized with correct causal structure if the generative model knows which objects are the accident participants. It proposes Causal-VidSyn, a video diffusion model that uses accident reason descriptions and driver gaze fixations to select and ground causal-prone visual tokens during training. The authors build Drive-Gaze, a 1.54-million-frame driver gaze dataset over 9,727 accident scenarios, to supply the fixation signal. They report that Causal-VidSyn outperforms prior video diffusion models on frame quality and on a gaze-based causal-sensitivity metric across accident video editing, normal-to-accident generation, and text-to-video tasks. The paper's core argument is that causal grounding, not just visual fidelity, is what separates usable accident synthesis from generic video generation.

What carries the argument

The load-bearing mechanism is the causal-prone token selection and grounding pipeline built on two plug-and-play blocks. CTS receives the noisy vision tokens after each temporal attention layer, fuses them with CLIP-encoded vision tokens and driver gaze tokens through a Gumbel-Softmax gated allocation, scores tokens by importance, and forwards the top quarter as causal-prone tokens while masking a quarter of background tokens to noise. CTG, attached at the end of the backbone, classifies causal versus background tokens using cross-modal attention aligned to the correct accident-reason answer among distractors and enforces a KL divergence between intervened and untouched background tokens. RPFD provides the global contrastive recipe by running two weight-shared diffusion pathways on forward and backward time-ordered frames with semantically reciprocal text prompts, pulling the reconstructed noises apart with a negative cosine similarity loss. These modules are training-only and plug into 3D-Unet or DiT backbones.

What would settle it

Take a fixed set of AEdit prompts and replace the driver gaze maps used in CTS training and Afd evaluation with randomized fixation maps or with object-detection-derived boxes (e.g., from Grounding DINO using the causal entity name). If Causal-VidSyn's Afd advantage over A-OA VD persists under randomized gaze but disappears under object-box anchors, the improvement is not attributable to causal grounding by gaze. Conversely, if a variant trained with ground-truth causal-entity boxes beats the gaze-trained model on Afd, the gaze signal is not the best carrier of causal identity.

Watch

Extended reading notes

Core claim

The central claim is that precise identification of accident participants and their behaviors, aided by accident reason answering and driver gaze-conditioned token selection, is what makes a video diffusion model causally sensitive in egocentric accident scenes. Causal-VidSyn achieves this by contrasting forward and backward time-order diffusion pathways under reciprocal text prompts (RPFD) and by injecting two plug-and-play modules into the diffusion backbone's layers: CTS, which selects causal-prone visual tokens under a gated gaze-vision fusion, and CTG, which grounds those tokens against accident reason-and-question text through causal/background token classification and an intervention loss. The model is trained with ArA and gaze maps only; at inference they are removed, leaving a standard diffusion model that edits or generates videos conditioned on text and frame prompts. The paper reports that this improves causal-entity editing, measured by a new affordance metric Afd that checks whether the edited object appears inside the driver's gazed region.

Load-bearing premise

The ten in-lab observers' gaze fixations on accident videos faithfully mark the objects that truly caused the accident, and the Afd metric's use of those gaze regions as ground truth measures causal-entity editing rather than mere attention saliency.

Editorial extensions

If this is right

  • Accident video editing becomes counterfactually faithful: changing 'pedestrian' to 'motorbike' in the text prompt yields a motorbike at the collision location while background stays intact.
  • Normal-to-accident generation transfers critical-object knowledge to accident-free datasets like BDD-A, producing frames that align with accident text prompts.
  • Text-to-video accident generation improves in semantic alignment, with the model producing collision states that baselines miss.
  • The CTS/CTG modules port to Transformer-based video diffusion models (CogVideoX, Latte), improving their AEdit and T2V scores after fine-tuning.
  • The Drive-Gaze dataset (1.54M frames of fixations over 9,727 accident scenarios, with accident reason texts) supports downstream tasks such as driver attention prediction, accident anticipation, and scanpath prediction.

Reading between the lines

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

  • If the gaze maps are noisy or biased toward non-causal salient regions, the Afd metric's gaze-anchored definition could overstate causal fidelity; a test replacing gaze regions with object-detection boxes would separate 'editing where drivers look' from 'editing the causally involved object'.
  • The reciprocal prompt design suggests a general recipe for injecting causal direction into diffusion models: contrasting forward and backward temporal narratives under semantically inverse text prompts could be applied beyond driving, e.g., to any event where a prevention description exists.
  • The training-only use of ArA and gaze means the method is a fine-tuning recipe for existing backbones; extending CTS/CTG to other tasks like accident anticipation could be tested by using synthetic accident frames to train anticipation models, as the paper's downstream TAA/VD explorations begin to do.
  • The Afd metric, being gaze-based, assumes human attention is the ground truth for causal entities; in multi-agent accidents where the driver's gaze is on one participant but the actual cause involves another, the metric may mis-attribute success.
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

4 major / 5 minor

Summary. The paper proposes Causal-VidSyn, a diffusion-based video generator for egocentric traffic accidents. It introduces a new driver gaze dataset, Drive-Gaze, with 1.54M frames across 9,727 accident videos, and trains a 3D-Unet diffusion model with two causal-grounding mechanisms: a reciprocal prompted frame diffusion (RPFD) that contrasts forward accident descriptions with backward prevention advice, and knowledge-level causal-prone token selection (CTS) with an accident-reason-answering (ArA) guided token grounding (CTG) that uses driver gaze maps to select tokens. The model is evaluated on normal-to-accident diffusion (N2A), accident video editing (AEdit), and text-to-video (T2V) generation, reporting CLIPs, FVD, TempC, and a new Afd metric on AEdit. The main quantitative claim is that Causal-VidSyn surpasses prior video diffusion models in frame quality and causal sensitivity, with the strongest evidence being the Afd margin over Abductive-OA VD in the AEdit task (55.4 vs. 49.4).

Significance. The dataset contribution is potentially significant: Drive-Gaze is claimed to be the largest driver gaze dataset in accident scenarios, and the plug-and-play CTS/CTG modules are portable to DiT-based architectures, with reported gains for CogV-X-T and Latte-T. The paper also performs a useful range of qualitative comparisons and includes a downstream-task exploration. However, the central causal-sensitivity claim is not yet established, because the Afd metric's gaze ground truth is coupled to the gaze-conditioned training objective, and because all quantitative results come from single runs without error bars. The paper's value will depend on independent validation of what Afd actually measures and on more robust statistical evidence for the modest gains in CLIPs and FVD.

major comments (4)
  1. [§5.4, Fig. 5 and §4.2, Eqs. (5)-(6)] The Afd metric is not an independent probe of causal-entity identification. Afd counts an edit as successful when the GroundingDINO box of the target object has IoU>0 with the DADA-2000 driver gaze region. In Stage-2, the CTS block is trained with gaze-conditioned gated fusion and causal-prone token selection, and the training clips come from Drive-Gaze, which is built from the same MM-AU source as DADA-2000. The AEdit test samples 3,000 DADA-2000 clips from the same NC-2-C frame windows. Therefore a high Afd can reflect the model reproducing the gaze-conditioned placement learned during training rather than identifying causal entities. Table 3 makes this concrete: removing gaze from the full model drops Afd from 55.4 to 50.4, close to A-OA VD's 49.4. The paper should provide a held-out evaluation that does not use gaze either in training or in the metric, for example human-annotated causal-participant bounding boxes or a human study of causal-role editing, and should report Afd for a model trained with shuffled or random gaze maps to control for the gaze-conditioning channel.
  2. [Tables 2-4] All quantitative results are single runs with no error bars or significance tests. For example, on N2A the FVD difference between Causal-VidSyn (6192.3) and A-OA VD (6208.2) is about 0.3%, and on AEdit the FVD difference is 5352.9 vs. 5358.2 with TempC slightly worse (0.940 vs. 0.947). The only non-negligible margin appears in Afd. The authors should report mean and standard deviation over at least three seeds or training runs and provide a statistical test (for example a paired bootstrap over clips) before claiming state-of-the-art frame quality and causal sensitivity.
  3. [§5.5, Fig. 8 and §5.4, Fig. 5] Afd conflates object scale with causal success. The paper itself notes in the V2T group that trucks obtain larger Afd values because their large scale easily yields a large IoU with the gaze region, and the IoU>0 threshold privileges large objects. This means Afd is not a clean measure of causal-entity editing quality across object types and can bias model comparisons. The authors should report continuous IoU scores, detection recall of the edited object, and results stratified by object size, so that the metric can be interpreted independently of object scale.
  4. [§3 and §5.4] The assumption that in-lab driver gaze maps from ten observers watching videos faithfully point to causal accident participants is not validated. This assumption is load-bearing because gaze is used both to select causal-prone tokens during training and as the ground-truth region for Afd. The manuscript should provide evidence such as inter-subject agreement of the collected gaze, correlation with the accident-participant annotations available in MM-AU, or a human study showing that gaze regions coincide with the objects responsible for the accident. Without such evidence, the term 'causal-entity reflected' is not supported beyond 'gaze-agreeing editing'.
minor comments (5)
  1. [Abstract] The word 'Egocentricly' should be 'Egocentrically' or 'Ego-centrically'.
  2. [§4.2] The text says 'the reminder ones are treated as the background tokens'; 'reminder' should be 'remainder'.
  3. [§4.1 and §4.2] There are several typos and placeholder artifacts in the source text, including 'simpilicity' for 'simplicity' and the '/care◎-righ◎' markers scattered through Sections 3-5; these should be cleaned before publication.
  4. [§3 and §5.1] The statement 'We take all frames in Drive-Gaze for training use' is inconsistent with the later statement that 6,492 clips are sampled from the 9,727 videos. Please clarify the exact training clip selection procedure and explicitly state that the DADA-2000 clips used for AEdit testing are disjoint from the Drive-Gaze training videos.
  5. [§4.2, Eq. (7)] The notation do(v) in Eq. (7) suggests a causal intervention, but randomly masking a quarter of background tokens to noise is a perturbation rather than a Pearl-style do-operator; the causal vocabulary should be qualified or replaced.

Circularity Check

1 steps flagged · score 6.0 of 10

Afd causal-sensitivity metric uses the same driver-gaze signal that trains the gaze-gated token selector, so the AEdit causal-sensitivity advantage is partly a training-target alignment artifact.

  1. fitted input called prediction [Section 3 (Utilization Ways); Section 4.2 Eq. 5 (CTS gaze gating); Section 5.1 (AEdit); Section 5.4/Fig. 5 (Afd metric)]
    "We take all frames in Drive-Gaze for training use. ... Because of the sudden scene change in egocentric accident videos, we introduce the driver gaze to help the causal-prone token selection in CTS. ... In this task, we adopt the DADA-2000 dataset [16], which also provides the driver fixations for checking whether the noticed objects aligned with text prompts are edited. ... we utilize the gazed regions instead of the object bounding boxes to match a human-preferable causal-entity editing checking ... Afd is the ratio of IOU(, ) > 0 of all checks."

    Stage-❷ trains CTS with gaze-conditioned gating: Eq. 5 forms zgate_v = z_v ⊗ zfu_v, where zfu_v comes from Gumbel-Softmax over concatenated CLIP vision tokens and driver-gaze tokens z_g, and Section 3 states that all Drive-Gaze frames (including DADA-2000's existing gaze data) are used for training. The AEdit evaluation metric Afd then counts, per frame, whether the GroundingDINO box of the edited text-query object has IOU>0 with the DADA-2000 driver gaze region. Thus the same gaze modality is both a training-conditioning signal for selecting causal-prone tokens and the ground-truth region used to define 'causal-entity reflected' editing success.

full rationale

The paper's core diffusion design is not circular in the usual sense: RPFD, CTS, and CTG are trained with external accident-reason text annotations and newly collected driver gaze maps, not with the target metrics FVD, CLIPs, or TempC, and the N2A and T2V comparisons use standard external metrics against several SOTA models. The many self-citations to MM-AU / A-OA VD [17] are heavy, but they are supported by ablations and portability experiments (Tables 2-5) and do not themselves force the central results. The one substantive circularity is the AEdit causal-sensitivity claim: the Afd metric's ground truth is the driver gaze region, and driver gaze is also a training input to the gaze-gated token selection (Eq. 5), with the same DADA-2000 gaze maps used in both training and evaluation. Consequently, the headline causal-sensitivity improvement is partially circular: high Afd can reflect 'the model places objects where the gaze-conditioned training target told it to' rather than an independent measure of causal-entity grounding. Score 6 is appropriate because one central evaluation claim reduces to the training-conditioning signal, while other tasks and metrics retain independent content.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The central claim rests mostly on the quality of the MM-AU text annotations and the in-lab gaze maps, both produced by the authors. The network itself uses standard diffusion and attention machinery; no new physical entities are introduced.

free parameters (5)
  • lambda = 0.2
    Weight of the negative similarity loss LNS in Eq. 4; chosen by hand, not cross-validated.
  • gamma = 0.3
    Weight of the ArA grounding loss L_ArA in Eq. 8; chosen by hand.
  • top-d token proportion = 1/4 of tokens
    Empirical choice for the number of causal-prone tokens selected in CTS; no sensitivity analysis.
  • training steps per stage = 10,000
    Set equally for all three stages; no justification or ablation.
  • gaze map kernel = 50x50 Gaussian
    Used to generate gaze maps from fixations; standard in DADA-2000, but the kernel size affects the gaze maps.
assumptions (4)
  • domain assumption The 3D-Unet backbone from A-OA VD [17] is a suitable base for egocentric accident video diffusion.
    The paper reuses the backbone and MM-AU dataset from the authors' prior work, assuming they transfer to the new tasks.
  • domain assumption MM-AU accident reason and prevention text descriptions are accurate causal labels for the videos.
    Section 3 and 4 rely on these annotations as ground truth for ArA and RPFD; errors would propagate to the training signal.
  • domain assumption Laboratory-collected eye fixations on videos generalize to real driver gaze in accidents.
    Section 3 describes in-lab collection with ten subjects; the paper does not validate against real driving gaze.
  • standard math The Gumbel-Softmax gates in Eq. 5 produce meaningful token selections without introducing harmful bias.
    Gumbel-Softmax is a standard relaxation; no specific issue, but it is a modeling choice.
invented entities (1)
  • causal-prone tokens
    purpose: Token-level units in the diffusion latent space that are selected and grounded as the accidentally relevant content.
    These are internal representations defined by the paper's selection procedure; there is no external benchmark indicating they correspond to true causal objects.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Causal-Entity Reflected Egocentric Traffic Accident Video Synthesis." pith.science (2026). https://pith.science/paper/NHAXVABZ

@misc{pith2026250623263,
  author       = {Pith},
  title        = {Pith review of: Causal-Entity Reflected Egocentric Traffic Accident Video Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NHAXVABZ}},
  note         = {Machine review of arXiv:2506.23263}
}
read the original abstract

Egocentricly comprehending the causes and effects of car accidents is crucial for the safety of self-driving cars, and synthesizing causal-entity reflected accident videos can facilitate the capability test to respond to unaffordable accidents in reality. However, incorporating causal relations as seen in real-world videos into synthetic videos remains challenging. This work argues that precisely identifying the accident participants and capturing their related behaviors are of critical importance. In this regard, we propose a novel diffusion model, Causal-VidSyn, for synthesizing egocentric traffic accident videos. To enable causal entity grounding in video diffusion, Causal-VidSyn leverages the cause descriptions and driver fixations to identify the accident participants and behaviors, facilitated by accident reason answering and gaze-conditioned selection modules. To support Causal-VidSyn, we further construct Drive-Gaze, the largest driver gaze dataset (with 1.54M frames of fixations) in driving accident scenarios. Extensive experiments show that Causal-VidSyn surpasses state-of-the-art video diffusion models in terms of frame quality and causal sensitivity in various tasks, including accident video editing, normal-to-accident video diffusion, and text-to-video generation.

Figures

Figures reproduced from arXiv: 2506.23263 by the authors.

Figure 1
Figure 1. We show the inability of two state-of-the-art diffusion [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The training schema of Causal-VidSyn. It mainly includes three stages: The direct optimization of e f with the forward time order diffusion (Stage-0, omitted in this figure), where the pre-trained Stable Diffusion [58] is used for model initialization. The reciprocal prompted frame diffusion (RPFD) in Stage-❶ contrasts two diffusion pathways conditioned by reciprocal time order vision frames (Vf&Vr) and semantically… view at source ↗
Figure 3
Figure 3. The details of Cross-Layer Causal-Prone Token Tran [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Sample visualizations of N2A task by Latte* [ [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: We visualize AEdit results of one crossing situation by [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Visualizations of one N2A example and an AEdit sample with the ablation [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: The CLIPs and Afd values of Causal-VidSyn, w.r.t., fine-grained causal-entity editing from “pedestrian” to “vehicle” (P2V), “vehicle” to “pedestrian” (V2P), “vehicle” to “motorcy￾cle/motorbike” (V2M), “vehicle" to “cyclist” (V2C), “vehicle to “truck” (V2T), and “motorc…
Figure 1
Figure 1. Figure 1: The detailed injection workflow of CTS and CTG in Causal-VidSyn, interleaved with multiple SA, CA, and TA blocks in 3D-Unet or DiT-based video diffusion models. The dimension indexes are defined as: B denotes the batch size, F = 16 is the video frame length, the maximu…
Figure 2
Figure 2. Figure 2: We visualize two N2A examples and four AEdit samples with the ablation response checking of our Causal-VidSyn , CogV-X [PITH_FULL_IMAGE:figures/full_fig_p013_2.png]
Figure 3
Figure 3. Figure 3: Sample visualizations of N2A task by Latte* [ [PITH_FULL_IMAGE:figures/full_fig_p014_3.png]
Figure 4
Figure 4. Figure 4: Sample visualizations of T2V task by Free-bloom [ [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: We visualize more AEdit results of four different situations (rainy, dark, strong light, and shadow) by LAMP [ [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Comparisons between Causal-VidSyn and four commercial model: Pika-1.5, Vidu-1.5 (I2V), Kling-AI, and HunyuanVideo-I2V. [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: The TAA and decision Tasks. R is the real data, and S is the synthesis data driven by C (our Causal-VidSyn model). 4. Comparison with Commercial Models To verify the SOTA performance of our Causal-VidSyn, we also take four popular and famous commercial models, in￾cludi…
Figure 9
Figure 9. Figure 9: Some failure cases in N2A and AEdit tasks. [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

96 extracted references · 73 canonical work pages

  1. [1]

    Sonia Baee, Erfan Pakdamanian, Inki Kim, Lu Feng, Vicente Ordonez, and Laura E. Barnes. MEDIRL: predicting the visual attention of drivers via maximum entropy deep inverse reinforcement learning. In ICCV, pages 13158–13168, 2021. 2

    2021

  2. [2]

    DRIVE: deep reinforced accident anticipation with visual explanation

    Wentao Bao, Qi Yu, and Yu Kong. DRIVE: deep reinforced accident anticipation with visual explanation. In ICCV, pages 7599–7608, 2021. 3

    2021

  3. [3]

    Align your latents: High-resolution video synthesis with la- tent diffusion models

    Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dock- horn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with la- tent diffusion models. In CVPR, pages 22563–22575, 2023. 2

    2023

  4. [4]

    Causal question answer- ing with reinforcement learning

    Lukas Blübaum and Stefan Heindorf. Causal question answer- ing with reinforcement learning. In WWW, pages 2204–2215,

  5. [5]

    Sihite, and Laurent Itti

    Ali Borji, Dicky N. Sihite, and Laurent Itti. Computational modeling of top-down visual attention in interactive environ- ments. In BMVC, pages 1–12, 2011. 2

    2011

  6. [6]

    Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom

    Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11618– 11628, 2020. 6

    2020

  7. [7]

    Counterfactual samples synthesizing and training for robust visual question answering

    Long Chen, Yuhang Zheng, Yulei Niu, Hanwang Zhang, and Jun Xiao. Counterfactual samples synthesizing and training for robust visual question answering. IEEE Trans. Pattern Anal. Mach. Intell., 45(11):13218–13234, 2023. 5

    2023

  8. [8]

    Predicting human scanpaths in visual question answering

    Xianyu Chen, Ming Jiang, and Qi Zhao. Predicting human scanpaths in visual question answering. In CVPR, pages 10876–10885, 2021. 3

    2021

Show all 96 references
  1. [9]

    Fblnet: Feed- back loop network for driver attention prediction

    Yilong Chen, Zhixiong Nan, and Tao Xiang. Fblnet: Feed- back loop network for driver attention prediction. In ICCV, pages 13325–13334, 2023. 3

    2023

  2. [10]

    Learning spatial adaptation and temporal coherence in diffusion models for video super- resolution

    Zhikai Chen, Fuchen Long, Zhaofan Qiu, Ting Yao, Wengang Zhou, Jiebo Luo, and Tao Mei. Learning spatial adaptation and temporal coherence in diffusion models for video super- resolution. In CVPR, pages 9232–9241, 2024. 2

    2024

  3. [11]

    Medm: Mediating image diffusion models for video- to-video translation with temporal correspondence guidance

    Ernie Chu, Tzuhsuan Huang, Shuo-Yen Lin, and Jun-Cheng Chen. Medm: Mediating image diffusion models for video- to-video translation with temporal correspondence guidance. In AAAI, pages 1353–1361, 2024. 1

    2024

  4. [12]

    Tao Deng, Hongmei Yan, Long Qin, Thuyen Ngo, and B. S. Manjunath. How do drivers allocate their potential attention? driving fixation prediction via convolutional neural networks. IEEE Trans. Intell. Transp. Syst., 21(5):2146–2154, 2020. 2

    2020

  5. [13]

    Isha Dua, Thrupthi Ann John, Riya Gupta, and C. V . Jawahar. DGAZE: driver gaze mapping on road. In IROS, pages 5946– 5953, 2020. 2

    2020

  6. [14]

    Structure and content-guided video synthesis with diffusion models

    Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In ICCV, pages 7312–7322, 2023. 1

    2023

  7. [15]

    DADA-2000: can driving accident be predicted by driver attention? analyzed by A benchmark

    Jianwu Fang, Dingxin Yan, Jiahuan Qiao, Jianru Xue, He Wang, and Sen Li. DADA-2000: can driving accident be predicted by driver attention? analyzed by A benchmark. In ITSC, pages 4303–4309, 2019. 2, 3, 4, 7

    2000

  8. [16]

    DADA: driver attention prediction in driving accident scenarios

    Jianwu Fang, Dingxin Yan, Jiahuan Qiao, Jianru Xue, and Hongkai Yu. DADA: driver attention prediction in driving accident scenarios. IEEE Trans. Intell. Transp. Syst., 23(6): 4959–4971, 2022. 2, 6, 7

    2022

  9. [17]

    Abductive ego- view accident video understanding for safe driving perception

    Jianwu Fang, Lei-Lei Li, Junfei Zhou, Junbin Xiao, Hongkai Yu, Chen Lv, Jianru Xue, and Tat-Seng Chua. Abductive ego- view accident video understanding for safe driving perception. In CVPR, pages 22030–22040, 2024. 1, 2, 3, 4, 5, 6, 7, 8

    2024

  10. [18]

    Constructing personalized situation awareness dataset for hazard perception, comprehension, pro- jection, and action of drivers

    Shun Gan, Quan Li, Qingfan Wang, WenTao Chen, Detong Qin, and Bingbing Nie. Constructing personalized situation awareness dataset for hazard perception, comprehension, pro- jection, and action of drivers. In ITSC, pages 1697–1704,

  11. [19]

    Multisource adaption for driver attention prediction in arbitrary driving scenes

    Shun Gan, Xizhe Pei, Yulong Ge, Qingfan Wang, Shi Shang, Shengbo Eben Li, and Bingbing Nie. Multisource adaption for driver attention prediction in arbitrary driving scenes. IEEE Trans. Intell. Transp. Syst., 23(11):20912 – 20925, 2022. 3

    2022

  12. [20]

    Magicdrive: Street view generation with diverse 3D geometry control

    Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. Magicdrive: Street view generation with diverse 3D geometry control. In ICLR, 2024. 2

    2024

  13. [21]

    attended awareness

    Deepak Gopinath, Guy Rosman, Simon Stent, Katsuya Tera- hata, Luke Fletcher, Brenna Argall, and John Leonard. Maad: A model and dataset for" attended awareness" in driving. In ICCV, pages 3426–3436, 2021. 2

    2021

  14. [22]

    Drivinggen: Ef- ficient safety-critical driving video generation with latent diffusion models

    Zipeng Guo, Yuchen Zhou, and Chao Gou. Drivinggen: Ef- ficient safety-critical driving video generation with latent diffusion models. In ICME, pages 1–6, 2024. 2

    2024

  15. [23]

    Genvideo: One-shot target-image and shape aware video editing using t2i diffusion models

    Sai Sree Harsha, Ambareesh Revanur, Dhwanit Agarwal, and Shradha Agrawal. Genvideo: One-shot target-image and shape aware video editing using t2i diffusion models. pages 7559–7568, 2024. 2

    2024

  16. [24]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. NeurIPS, 33:6840–6851, 2020. 1

    2020

  17. [25]

    Gritsenko, William Chan, Mohammad Norouzi, and David J

    Jonathan Ho, Tim Salimans, Alexey A. Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffu- sion models. In NeurIPS, 2022. 1

    2022

  18. [26]

    Gaia-1: A generative world model for autonomous driving

    Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Cor- rado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 2

    2023 arXiv

  19. [27]

    Videoshield: Regulating diffusion- based video generation models via watermarking

    Runyi Hu, Jie Zhang, Yiming Li, Jiwei Li, Qing Guo, Han Qiu, and Tianwei Zhang. Videoshield: Regulating diffusion- based video generation models via watermarking. In ICLR,

  20. [28]

    Free-bloom: Zero-shot text-to-video generator with llm director and ldm animator

    Hanzhuo Huang, Yufan Feng, Cheng Shi, Lan Xu, Jingyi Yu, and Sibei Yang. Free-bloom: Zero-shot text-to-video generator with llm director and ldm animator. NeurIPS, 36,

  21. [29]

    Categorical reparam- eterization with gumbel-softmax

    Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparam- eterization with gumbel-softmax. In ICLR, 2017. 5

    2017

  22. [30]

    Diffusion-based causal representation learning

    Amir Mohammad Karimi-Mamaghan, Andrea Dittadi, Ste- fan Bauer, Karl Henrik Johansson, and Francesco Quinzan. Diffusion-based causal representation learning. Entropy, 26 (7):556, 2024. 2

    2024

  23. [31]

    Look both ways: Self-supervising driver gaze estimation and road scene saliency

    Isaac Kasahara, Simon Stent, and Hyun Soo Park. Look both ways: Self-supervising driver gaze estimation and road scene saliency. In ECCV, pages 126–142, 2022. 2

    2022

  24. [32]

    Text2video-zero: Text-to-image diffusion models are zero-shot video generators

    Levon Khachatryan et al. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In ICCV, pages 15908–15918, 2023. 1, 6, 4

    2023

  25. [33]

    Causal diffusion autoencoders: Toward counterfactual gener- ation via diffusion probabilistic models

    Aneesh Komanduri, Chen Zhao, Feng Chen, and Xintao Wu. Causal diffusion autoencoders: Toward counterfactual gener- ation via diffusion probabilistic models. In CVPR Workshop,

  26. [34]

    Avd2: Acci- dent video diffusion for accident video description

    Cheng Li, Keyuan Zhou, Tong Liu, Yu Wang, Mingqiao Zhuang, Huan-ang Gao, Bu Jin, and Hao Zhao. Avd2: Acci- dent video diffusion for accident video description. In ICRA,

  27. [35]

    T2v- turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design

    Jiachen Li, Qian Long, Jian Zheng, Xiaofeng Gao, Robinson Piramuthu, Wenhu Chen, and William Yang Wang. T2v- turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design. In ICLR, 2025. 2

    2025

  28. [36]

    Cognitive traffic accident anticipation

    Lei-Lei Li, Jianwu Fang, and Jianru Xue. Cognitive traffic accident anticipation. IEEE Intell. Transp. Syst. Mag., 16(5): 17–32, 2024. 3, 6

    2024

  29. [37]

    Image content generation with causal reasoning

    Xiaochuan Li, Baoyu Fan, Runze Zhang, Liang Jin, Di Wang, Zhenhua Guo, Yaqian Zhao, and Rengang Li. Image content generation with causal reasoning. In AAAI, pages 13646– 13654, 2024. 2

    2024

  30. [38]

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

    Xiaofan Li, Yifu Zhang, and Xiaoqing Ye. Drivingdiffusion: Layout-guided multi-view driving scenarios video generation with latent diffusion model. In ECCV, pages 469–485, 2024. 2

    2024

  31. [39]

    Equivariant and invariant grounding for video question an- swering

    Yicong Li, Xiang Wang, Junbin Xiao, and Tat-Seng Chua. Equivariant and invariant grounding for video question an- swering. In ACM MM, pages 4714–4722, 2022. 5

    2022

  32. [40]

    Invariant grounding for video question answering

    Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Invariant grounding for video question answering. In CVPR, pages 2928–2937, 2022. 5

    2022

  33. [41]

    Diffusiondrive: Truncated diffusion model for end-to-end autonomous driving

    Bencheng Liao, Shaoyu Chen, Haoran Yin, Bo Jiang, Cheng Wang, Sixu Yan, Xinbang Zhang, Xiangyu Li, Ying Zhang, Qian Zhang, and Xinggang Wang. Diffusiondrive: Truncated diffusion model for end-to-end autonomous driving. InCVPR,

  34. [42]

    Mo- tionclone: Training-free motion cloning for controllable video generation

    Pengyang Ling, Jiazi Bu, Pan Zhang, Xiaoyi Dong, Yuhang Zang, Tong Wu, Huaian Chen, Jiaqi Wang, and Yi Jin. Mo- tionclone: Training-free motion cloning for controllable video generation. In ICLR, 2025. 6, 7, 3

    2025

  35. [43]

    Ce-vdg: Counterfactual entropy-based bias reduction for video-grounded dialogue generation

    Hongcheng Liu, Pingjie Wang, Zhiyuan Zhu, Yanfeng Wang, and Yu Wang. Ce-vdg: Counterfactual entropy-based bias reduction for video-grounded dialogue generation. In LREC- COLING, pages 2958–2968, 2024. 2

    2024

  36. [44]

    Li, and Ge Li

    Ruyang Liu, Jingjia Huang, Thomas H. Li, and Ge Li. Causal- ity compensated attention for contextual biased visual recog- nition. In ICLR, 2023. 7

    2023

  37. [45]

    Grounding DINO: mar- rying DINO with grounded pre-training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: mar- rying DINO with grounded pre-training for open-set object detection. In ECCV, pages 38–55, 2024. 7

    2024

  38. [46]

    Cross-modal causal relational reasoning for event-level visual question answering

    Yang Liu, Guanbin Li, and Liang Lin. Cross-modal causal relational reasoning for event-level visual question answering. IEEE Trans. Pattern Anal. Mach. Intell., 45(10):11624–11641,

  39. [47]

    Videofusion: Decomposed diffusion models for high-quality video generation

    Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tie- niu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In CVPR, pages 10209–10218,

  40. [48]

    Latte: Latent diffusion transformer for video generation

    Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zi- wei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. CoRR, abs/2401.03048, 2024. 1, 2, 6, 7, 8, 3, 4

    2024 arXiv

  41. [49]

    Mastering context-to-label representation transformation for event causality identification with diffusion models

    Hieu Man, Franck Dernoncourt, and Thien Huu Nguyen. Mastering context-to-label representation transformation for event causality identification with diffusion models. In AAAI, pages 18760–18768, 2024. 2

    2024

  42. [50]

    Stealing stable diffusion prior for robust monocular depth estimation

    Yifan Mao, Jian Liu, and Xianming Liu. Stealing stable diffusion prior for robust monocular depth estimation. arXiv preprint arXiv:2403.05056, 2024. 2

    2024 arXiv

  43. [51]

    Safe driving cars

    Nature. Safe driving cars. Nat. Mach. Intell., 4(2):95–96,

  44. [52]

    Driver attention tracking and analysis

    Dat Viet Thanh Nguyen, Anh Tran, Nam Vu, Cuong Pham, and Minh Hoai. Driver attention tracking and analysis. arXiv preprint arXiv:2404.07122, 2024. 2

    2024 arXiv

  45. [53]

    Predicting the driver’s focus of attention: The dr(eye)ve project

    Andrea Palazzi, Davide Abati, Simone Calderara, Francesco Solera, and Rita Cucchiara. Predicting the driver’s focus of attention: The dr(eye)ve project. IEEE Trans. Pattern Anal. Mach. Intell., 41(7):1720–1733, 2019. 2

    2019

  46. [54]

    Generating driv- ing scenes with diffusion

    Ethan Pronovost, Kai Wang, and Nick Roy. Generating driv- ing scenes with diffusion. arXiv preprint arXiv:2305.18452,

    arXiv

  47. [55]

    Fatezero: Fusing attentions for zero-shot text-based video editing

    Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In ICCV, pages 15886–15896, 2023. 6

    2023

  48. [56]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748–8763, 2021. 4, 5, 1

    2021

  49. [57]

    Review on synergizing the metaverse and ai-driven synthetic data: enhancing virtual realms and activity recognition in computer vision.Vis

    Megani Rajendran, Chek Tien Tan, Indriyati Atmosukarto, Aik Beng Ng, and Simon See. Review on synergizing the metaverse and ai-driven synthetic data: enhancing virtual realms and activity recognition in computer vision.Vis. Intell., 2(1):27, 2024. 2

    2024

  50. [58]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. InCVPR, pages 10674– 10685, 2022. 3

    2022

  51. [59]

    Tsaftaris

    Pedro Sanchez and Sotirios A. Tsaftaris. Diffusion causal models for counterfactual estimation. In CLeaR, pages 647– 668, 2022. 2

    2022

  52. [60]

    O’Neil, and Sotirios A

    Pedro Sanchez, Xiao Liu, Alison Q. O’Neil, and Sotirios A. Tsaftaris. Diffusion models for causal discovery via topologi- cal ordering. In ICLR, 2023. 2

    2023

  53. [61]

    Cocatt: A cognitive-conditioned driver attention dataset

    Yuan Shen, Niviru Wijayaratne, Pranav Sriram, Aamir Hasan, Peter Du, and Katherine Driggs Campbell. Cocatt: A cognitive-conditioned driver attention dataset. In ITSC, pages 32–39, 2022. 2

    2022

  54. [62]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 5

    2021

  55. [63]

    Text2street: Controllable text-to-image genera- tion for street views

    Jinming Su, Songen Gu, Yiting Duan, Xingyue Chen, and Junfeng Luo. Text2street: Controllable text-to-image genera- tion for street views. arXiv preprint arXiv:2402.04504, 2024. 2

    2024 arXiv

  56. [64]

    Add-it: Training-free object insertion in images with pretrained diffusion models

    Yoad Tewel, Rinon Gal, Dvir Samuel, Yuval Atzmon, Lior Wolf, and Gal Chechik. Add-it: Training-free object insertion in images with pretrained diffusion models. In ICLR, 2025. 7

    2025

  57. [65]

    Cherise Threewitt. U. s. news, 10 vehicles that are almost self- driving in 2024. https://cars.usnews.com/cars- trucks/advice/cars-that-are-almost-self- driving, 2024. 1

    2024

  58. [66]

    Driving as well as on a sunny day? predicting driver’s fixation in rainy weather conditions via a dual-branch visual model

    Han Tian, Tao Deng, and Hongmei Yan. Driving as well as on a sunny day? predicting driver’s fixation in rainy weather conditions via a dual-branch visual model. IEEE CAA J. Autom. Sinica, 9(7):1335–1338, 2022. 2

    2022

  59. [67]

    To- wards accurate generative models of video: A new metric & challenges

    Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. To- wards accurate generative models of video: A new metric & challenges. CoRR, abs/1812.01717, 2018. 6

    2018 arXiv

  60. [68]

    Seedvr: Seed- ing infinity in diffusion transformer towards generic video restoration

    Jianyi Wang, Zhijie Lin, Meng Wei, Yang Zhao, Ceyuan Yang, Fei Xiao, Chen Change Loy, and Lu Jiang. Seedvr: Seed- ing infinity in diffusion transformer towards generic video restoration. In CVPR, 2025. 2, 6

    2025

  61. [69]

    Reinforced causal explainer for graph neural networks

    Xiang Wang, Yingxin Wu, An Zhang, Fuli Feng, Xiangnan He, and Tat-Seng Chua. Reinforced causal explainer for graph neural networks. IEEE Trans. Pattern Anal. Mach. Intell., 45 (2):2297–2309, 2023. 2

    2023

  62. [70]

    Drivedreamer: Towards real-world-driven world models for autonomous driving

    Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. In ECCV, 2024. 2

    2024

  63. [71]

    Lavie: High-quality video generation with cascaded latent diffusion models

    Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. Int. J. Comput. Vis.,

  64. [72]

    Detail-enhancing framework for reference-based image super-resolution

    Zihan Wang, Ziliang Xiong, Hongying Tang, and Xiaobing Yuan. Detail-enhancing framework for reference-based image super-resolution. arXiv preprint arXiv:2405.00431, 2024. 2

    2024 arXiv

  65. [73]

    Panacea: Panoramic and controllable video generation for autonomous driving

    Yuqing Wen, Yucheng Zhao, Yingfei Liu, Fan Jia, Yanhui Wang, Chong Luo, Chi Zhang, Tiancai Wang, Xiaoyan Sun, and Xiangyu Zhang. Panacea: Panoramic and controllable video generation for autonomous driving. In CVPR, pages 6902–6912, 2024. 2

    2024

  66. [74]

    Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation

    Jay Zhangjie Wu et al. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, pages 7589–7599, 2023. 6, 7, 8, 1

    2023

  67. [75]

    Lamp: Learn a motion pattern for few-shot video generation

    Ruiqi Wu, Liangyu Chen, Tong Yang, Chunle Guo, Chongyi Li, and Xiangyu Zhang. Lamp: Learn a motion pattern for few-shot video generation. In CVPR, pages 7089–7098, 2024. 6, 7, 3, 4, 5

    2024

  68. [76]

    Predicting driver attention in critical situations

    Ye Xia, Danqing Zhang, Jinkyu Kim, Ken Nakayama, Karl Zipser, and David Whitney. Predicting driver attention in critical situations. In ACCV, pages 658–674, 2018. 2, 6

    2018

  69. [77]

    Versvideo: Leveraging enhanced tempo- ral diffusion models for versatile video generation

    Jinxi Xiang, Ricong Huang, Jun Zhang, Guanbin Li, Xiao Han, and Yang Wei. Versvideo: Leveraging enhanced tempo- ral diffusion models for versatile video generation. In ICLR,

  70. [78]

    Video diffusion models are training-free motion interpreter and con- troller

    Zeqi Xiao, Yifan Zhou, Shuai Yang, and Xingang Pan. Video diffusion models are training-free motion interpreter and con- troller. CoRR, abs/2405.14864, 2024. 2, 6

    2024 arXiv

  71. [79]

    Streetcrafter: Street view synthesis with controllable video diffusion models

    Yunzhi Yan, Zhen Xu, Haotong Lin, Haian Jin, Haoyu Guo, Yida Wang, Kun Zhan, Xianpeng Lang, Hujun Bao, Xiaowei Zhou, and Sida Peng. Streetcrafter: Street view synthesis with controllable video diffusion models. In CVPR, 2025. 2

    2025

  72. [80]

    Deconfounded image captioning: A causal retrospect

    Xu Yang, Hanwang Zhang, and Jianfei Cai. Deconfounded image captioning: A causal retrospect. IEEE Trans. Pattern Anal. Mach. Intell., 45(11):12996–13010, 2023. 2

    2023

  73. [81]

    Videograin: Modulating space-time attention for multi- grained video editing

    Xiangpeng Yang, Linchao Zhu, Hehe Fan, and Yi Yang. Videograin: Modulating space-time attention for multi- grained video editing. 2025. 1

    2025

  74. [82]

    Cogvideox: Text-to-video diffu- sion models with an expert transformer

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffu- sion models with an expert transformer. In ICLR, 2025. 1, 2, 6, 7, 8, 3, 4

    2025

  75. [83]

    Stdiff: Spatio- temporal diffusion for continuous stochastic video prediction

    Xi Ye and Guillaume-Alexandre Bilodeau. Stdiff: Spatio- temporal diffusion for continuous stochastic video prediction. In AAAI, pages 6666–6674, 2024. 1

    2024

  76. [84]

    Inflation with diffusion: Efficient temporal adaptation for text-to-video super-resolution

    Xin Yuan, Jinoo Baek, Keyang Xu, Omer Tov, and Hongliang Fei. Inflation with diffusion: Efficient temporal adaptation for text-to-video super-resolution. In WACV, pages 489–496,

  77. [85]

    Boosting causal discovery via adaptive sample reweighting

    An Zhang, Fangfu Liu, Wenchang Ma, Zhibo Cai, Xiang Wang, and Tat-Seng Chua. Boosting causal discovery via adaptive sample reweighting. In ICLR, 2023. 2

    2023

  78. [86]

    How to unify grounding and causation

    Wenjun Zhang. How to unify grounding and causation. Syn- these, 202(24):1–18, 2023. 2

    2023

  79. [87]

    Controlvideo: Training- free controllable text-to-video generation

    Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training- free controllable text-to-video generation. In ICLR, 2024. 6

    2024

  80. [88]

    Towards consistent video editing with text-to-image diffusion models

    Zicheng Zhang, Bonan Li, Xuecheng Nie, Congying Han, Tiande Guo, and Luoqi Liu. Towards consistent video editing with text-to-image diffusion models. In NeurIPS, 2023. 2

    2023

  81. [89]

    Storydiffusion: Consistent self-attention for long-range image and video generation

    Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation. In ICLR, 2024. 2 Supplementary Material of Causal-Entity Reflected Egocentric Traffic Accident Video Synthesis PE SA CA T...

    2024

  82. [90]

    1, where different at- tention modules, i.e., SA, CA, and TA, are fed into low- rank adaptation (LoRA) trainer 2 for fast fine-tuning

    More Details of Implementation To be clear for re-reproduction, we provide the injection workflow of CTS and CTG in Fig. 1, where different at- tention modules, i.e., SA, CA, and TA, are fed into low- rank adaptation (LoRA) trainer 2 for fast fine-tuning. In Stage-❷, we mainly...

  83. [91]

    +RPFD (!

    More Details of CTS and CTG 2.1. The Architecture of Sampling Adapter As denoted in Fig. 1, ze(l) v ∈ R(Bhlwl)×F ×Cl. To match the dimension of zgate v in Eq. 4 stated in the main paper, a bilinear interpolation (BintP) is applied to adjust the token dimension as R(Bhpwp)×F ×C...

  84. [92]

    large object issue

    More Evaluations of Causal-VidSyn For a solid evaluation, we offer more evaluations mainly from the visualizations of N2A, T2V , and AEdit tasks. 3.1. More Visualizations on N2A and T2V Tasks N2A Evaluation: We present more ego-car involved visu- alizations of the N2A task in ...

  85. [93]

    pedestrian

    Comparison with Commercial Models To verify the SOTA performance of our Causal-VidSyn, we also take four popular and famous commercial models, in- cluding Pika-1.55, Vidu-1.56, Kling AI7, and newly released HunyuanVideo-I2V8, for egocentric traffic accident video generation. F...

  86. [94]

    Downstream Task Explorations We explore two downstream tasks using our synthetic data: traffic accident anticipation (TAA) and visual decision (VD). Fig. 7 shows the pipelines of them. In the TAA task, we take Cog-TAA [36] as the baseline, and take the accident- free BDD-A [76...

    2000

  87. [95]

    overtak- ing

    Failure Case Analysis In addition, we also show the limitations of our Causal- VidSyn by analyzing some failure cases, as shown in Fig. 9. In this analysis, we take several samples in N2A and AEdit tasks because of the demand for causal-sensitivity checking. For the failure ca...

  88. [96]

    In addition, we claim that all authors have solid contributions to this work

    Ethics Statement The misuse including the creation of deceptive accident con- tent for evidence collection may have negative societal im- pacts, and we advocate positive use for deep accident un- derstanding, such as accident anticipation. In addition, we claim that all author...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.