Pith. sign in

REVIEW 4 major objections 4 minor 46 references

Direct Preference Optimization-Enhanced Multi-Guided Diffusion Model for Traffic Scenario Generation

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

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

Pith's one-line read MuDi-Pro shows DPO fine-tuning improves rule-following in guided traffic scenario generation.

desk verdict The DPO mechanism as written does not match what is implemented, so the central controllability claim is not supported, but the non-DPO backbone and ablations are solid enough to merit a careful revision. read the letter →

arxiv 2502.12178 v1 pith:CZIS2VFK submitted 2025-02-14 cs.LG cs.MA

Seungjun Yu , Kisung Kim , Daejung Kim , Haewook Han , Jinhan Lee This is my paper
classification cs.LGcs.MA
keywords trafficscenariogenerationdiffusionmodeldirectpreferenceoptimizationguidedsamplingmulti-tasklearningtrajectorynuScenescontrollability
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

Traffic scenario generation for autonomous-vehicle testing must balance realism, diversity, and controllability, but guided diffusion sampling that pushes trajectories toward traffic rules tends to pull them away from real driving behavior. This paper argues that the tension can be resolved by a two-stage training strategy: a scene-level diffusion transformer trained on real driving data, then fine-tuned with Direct Preference Optimization (DPO) over preferences defined by the guidance scores themselves. The result, MuDi-Pro, is a single model that handles multiple combinations of guidance inputs while improving rule adherence. On the nuScenes dataset the fine-tuned model lowers rule-violation and collision metrics compared with its non-fine-tuned backbone, which is what would matter if the approach is right: safer, more controllable synthetic scenarios without losing realism or diversity.

What carries the argument

The argument rests on three components. First, a scene-level diffusion transformer learns a traffic prior from real driving data, using classifier-free sampling and clean-trajectory guided sampling so that arbitrary guidance functions can steer the denoising process. Second, a guidance conditional layer, analogous to task-conditioning in multi-task learning, embeds the active guidance set into a latent vector that modulates the transformer blocks, letting one model handle many guidance combinations. Third, DPO fine-tuning optimizes preferences: for each prompt, the model samples a winning and a losing trajectory, scores them by the guidance loss, and updates through $L_{\text{DPO}} = -\log\sigma(-\beta(d_w - d_l))$, where $d_i$ is the difference in $L_2$ error between the current model and a frozen reference model against the stored trajectory. That surrogate replaces the log-probability ratio of the DPO objective with measurable trajectory errors, avoiding non-differentiable guidance gradients.

What would settle it

Compute the exact log-probabilities of the stored winning and losing trajectories under both the target and reference models, and compare the gradient of the true DPO loss to the gradient of the Algorithm 1 surrogate on the same sampled pairs; if the two disagree in sign or magnitude, the fine-tuning is not performing direct preference optimization. Alternatively, relabel the preference pairs with human-rated realism instead of guidance scores and check whether the rule-following improvements persist.

Watch

Extended reading notes

Core claim

The paper's central claim is that fine-tuning a multi-guided diffusion model with DPO, using winning and losing trajectory pairs scored by the same guidance functions that drive sampling, makes guided sampling more effective. Concretely, MuDi-Pro is shown to reduce the rule metric from 52.4 to 43.5 and vehicle collisions from 0.20% to 0.03% for a combined no-collision and no-off-road rule, while keeping realism scores essentially unchanged, and also to improve over its backbone on a three-rule combination. The authors also claim that the guidance conditional layer, a multi-task-style addition that encodes which guidance is active, allows one model to serve many guidance combinations rather than requiring separate fine-tuning per rule.

Load-bearing premise

Algorithm 1 assumes the $L_2$ error between a sampled trajectory and a stored winning or losing sample can stand in for the model's log-probability of that trajectory, so that minimizing the difference of these errors is equivalent to optimizing the DPO objective; the paper does not derive or test this equivalence.

Editorial extensions

If this is right

  • If MuDi-Pro is right, a single diffusion model can be fine-tuned once and then steer by many rule combinations, removing the need to train separate models per guidance.
  • The DPO treatment of guidance preferences offers a generic recipe for fine-tuning guided generative models whenever guidance gradients are expensive or non-differentiable.
  • The reported reductions in off-road and vehicle collision rates, with comparable realism, would make synthetic scenario generation more usable for safety testing of autonomous vehicles.
  • The sample mode, which does not see future information, maintains diversity while improving controllability, so the gains are not limited to reconstruction settings.

Reading between the lines

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

  • A limitation the paper does not dwell on: because the winning and losing labels come from the same guidance scores used as evaluation metrics, the reported rule-following improvements may partly reflect alignment with the metric rather than with genuinely realistic driving; a human-preference or independent-realness evaluation would separate these.
  • The $L_2$ surrogate for log-probabilities in Algorithm 1 is the load-bearing approximation; if it breaks, the fine-tuning is no longer DPO, so testing the surrogate against exact likelihoods would clarify the mechanism.
  • The authors note classifier-free sampling is limited to trajectory space; extending the same DPO fine-tuning to latent-space sampling would test whether the controllability gains carry over.
  • If the surrogate is validated, the same fine-tuning scheme could apply to other domains where guidance is defined by rule-based losses, such as pedestrian or robot motion generation.
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 / 4 minor

Summary. The paper proposes MuDi-Pro, a traffic-scenario generation model built on a diffusion transformer, classifier-free sampling with reconstruction guidance, a guidance-conditional layer for multi-guided control, and a DPO-style fine-tuning stage intended to reduce the prior shift caused by guided sampling. Training is split into a backbone stage, a guidance-conditional-layer stage (MuDi), and a final preference fine-tuning stage (MuDi-Pro). Experiments on nuScenes compare the backbone, controllability, fine-tuned model, and ablations across metrics for stability, realism, controllability, and data-driven fidelity.

Significance. If the claimed mechanism worked as described, the paper would be a practically useful contribution to controllable traffic generation: the idea of conditioning one diffusion model on multiple guidance functions and then fine-tuning it to reduce guided-sampling distortion is attractive, and the paper includes useful ablations (unicycle dynamics, transformer backbone, data filtering, context design). The nuScenes data-filtering observation about parked agents with pre-existing collisions is also practically relevant. However, the central DPO mechanism is not actually derived or verified, the preference labels are the same quantities used as the controllability evaluation metric, and one of the headline claims in Section IV-B is contradicted by the paper's own Table I. The paper provides no code or other reproducibility artifact, so the empirical claims rest entirely on the textual description. The strengths do not outweigh these load-bearing concerns at the current stage.

major comments (4)
  1. [III-C, Eq. (5), Algorithm 1] Algorithm 1 does not implement the DPO objective in Eq. (5). Eq. (5) is the standard DPO loss over log-likelihood ratios of fixed winning and losing samples, while Algorithm 1 (lines 8-12) replaces the log-probability ratio with d_i = E_{i,theta} - E_{i,ref}, where E_{i,j} = L2(T_i, T_{i,j}) and T_{i,theta}, T_{i,ref} are freshly sampled trajectories from the target and reference models. The paper gives no derivation showing that an L2 error to a newly sampled trajectory estimates or proxies log p_theta(T_i|c) - log p_ref(T_i|c); for diffusion models, the log-likelihood is not an L2 distance to an arbitrary sample. Standard Diffusion-DPO, as in reference [34], optimizes denoising losses on fixed preference pairs, not random samples drawn from the current policy. Without a derivation or an ablation comparing this surrogate with the stated DPO objective, the controllability improvements in Table III cannot be attributed to direct preference optimization, which is the paper's central contribution.
  2. [III-C, Tables II and III] The preference dataset is constructed from guidance scores, and the 'rule' controllability metric in Tables II and III is the final guidance loss. Because the fine-tuning objective directly optimizes the same guidance scores that are later reported as the controllability metric, the reduction from 52.4 to 43.5 in Table III is partly a fitted result rather than independent evidence of improved rule-following. The paper needs either an external evaluation of controllability that was not used as a training signal, or an explicit statement that controllability is defined as this guidance loss and that the comparison is therefore a direct measure of training objective, not a downstream validation of rule compliance.
  3. [IV-B, Table I] The text in Section IV-B states that 'ours (sample) also outperforms baselines with all metrics except for the veh metric value,' but Table I directly contradicts this. Under the paper's own statement that lower pos and ang indicate better data-driven ability, Ours(sample) has pos=2.738 and ang=2.682, which are worse than every baseline, and map=4.84 is worse than CTG (2.31), CTG++ (3.92), and STRIVE(recon) (2.67). Only the real metric (2.0) is best among all compared entries. This false summary undermines the data-driven-ability claim for the sample mode and should be corrected.
  4. [IV-C, Table II] Section IV-C says the authors 'apply four single-rule cases and also conduct experiments with three multiple-rule cases,' but Table II reports only two multiple-rule combinations (rule 1+2 and rule 1+2+4). A third combination such as rule 1+2+5 appears in a figure caption but is not tabulated. Since multi-guidance handling is a core contribution, the experimental record is incomplete as presented and should be clarified or completed.
minor comments (4)
  1. [III-C, Eq. (5)] The notation 'pref' in Eq. (5) should be 'p_ref' for consistency with the surrounding text, and the subscript placement in the preference ratio should be made explicit to avoid confusion with multiplication.
  2. [Algorithm 1] Line 1 contains a typo ('lossing' should be 'losing'), and the notation 'i ∼ [w,l]' is not standard; the set should be written explicitly as i in {w,l}.
  3. [IV-E, Table IV] The ablation rows 'MuDi-Pro(decoder)' and 'MuDi-Pro(blocks)' are not defined in the text; it should be stated which parameters are fine-tuned in each variant and whether the guidance conditional layer is included.
  4. [IV, metrics] No standard deviations, confidence intervals, or multiple-seed comparisons are reported for any metric; given the small differences in some comparisons (e.g., real in Table III), this limits the statistical support for the empirical conclusions.

Circularity Check

1 steps flagged · score 6.0 of 10

MuDi-Pro's rule-controllability improvement is a re-measurement of its own DPO preference signal, making the headline controllability gain partly a fitted result.

  1. fitted input called prediction [Section III-C 'Fine-tuning via DPO' and Section IV-A 'Metrics'; Table III]
    "We devise a preference dataset predicated on the guidance score, which is directly derived from the guidance loss during the diffusion model’s sampling phase, to provide heuristic-based preferences for intuitiveness... Controllability is assessed by how well the trajectories follow guidance, measured by the final guided loss (rule)."

    The preference labels that define winning and losing samples are the guidance scores, and the controllability metric 'rule' is the same final guided loss. The DPO objective is trained to prefer the lower-guidance-loss sample, so the reported rule decreases in Table III (e.g., 52.4 to 43.5 for rule 1+2) are a re-measurement of the training signal itself. The evaluation metric is not independent of the optimized objective; the rule improvement is largely a fitted outcome of the preference construction rather than an independent prediction about controllability.

full rationale

The backbone comparison against CTG/STRIVE and the realism/diversity metrics are externally anchored on nuScenes and do not reduce to the training objective. However, the central controllability claim for MuDi-Pro is partially circular: the DPO preference dataset is constructed from guidance scores, and the 'rule' evaluation metric is the final guided loss from the same guidance functions. Consequently, the rule improvements in Table III substantially reflect the objective used for fine-tuning. The paper does not rely on self-citations. A separate non-circular concern is that Algorithm 1 replaces the log-probability ratio in Eq. (5) with L2 distances to sampled trajectories without derivation; this is a correctness gap rather than circularity, so it is not scored here. Overall, the controllability finding is partially forced by construction, giving a score of 6.

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

The central claim relies on a set of hand-chosen hyperparameters (alpha, w, beta) and, most importantly, on the assumption that guidance scores define both the preference signal and the evaluation metric. No new physical entities are introduced.

free parameters (5)
  • guidance strength alpha = not reported
    Scales the guidance gradient in Eq. (4); chosen by hand for each rule, no value or sweep is given.
  • CFS weight w = 1.0, 0.5, 0.0
    Controls how much future-conditioned sampling is used in Eq. (3); evaluated at three fixed values.
  • DPO scale beta = not reported
    Regularization strength in Eq. (5); no value or selection procedure is provided.
  • guidance rule weights = not reported
    The composite guidance losses for collision, off-road, goal point, and target speed are combined without listing weights or normalization.
  • data filtering threshold = agents with past collisions (13.9% of training set)
    Agents whose past trajectories contain collisions are excluded from loss and evaluation; this ad hoc criterion changes the training and test distribution.
assumptions (5)
  • standard math Diffusion forward and reverse process with a fixed variance schedule
    Used in Eq. (1)-(2), retained from DDPM and Improved DDPM.
  • domain assumption Unicycle dynamics model
    Assumed to capture vehicle motion; the ablation shows removing it degrades results.
  • domain assumption Future trajectories are available at test time in reconstruction mode
    The model is trained with future-conditioning dropout; w=1.0 uses future information, which is not available in real deployment.
  • ad hoc to paper Guidance-score preference is a valid proxy for controllability and realism
    The DPO preference labels are computed from guidance losses, and the same losses are used as evaluation metrics in Tables II-IV.
  • domain assumption nuScenes dataset represents real-world traffic after filtering
    The paper uses nuScenes and removes agents with past collisions, assuming the remainder is clean enough for training and evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Direct Preference Optimization-Enhanced Multi-Guided Diffusion Model for Traffic Scenario Generation." pith.science (2026). https://pith.science/paper/CZIS2VFK

@misc{pith2026250212178,
  author       = {Pith},
  title        = {Pith review of: Direct Preference Optimization-Enhanced Multi-Guided Diffusion Model for Traffic Scenario Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CZIS2VFK}},
  note         = {Machine review of arXiv:2502.12178}
}
read the original abstract

Diffusion-based models are recognized for their effectiveness in using real-world driving data to generate realistic and diverse traffic scenarios. These models employ guided sampling to incorporate specific traffic preferences and enhance scenario realism. However, guiding the sampling process to conform to traffic rules and preferences can result in deviations from real-world traffic priors and potentially leading to unrealistic behaviors. To address this challenge, we introduce a multi-guided diffusion model that utilizes a novel training strategy to closely adhere to traffic priors, even when employing various combinations of guides. This model adopts a multi-task learning framework, enabling a single diffusion model to process various guide inputs. For increased guided sampling precision, our model is fine-tuned using the Direct Preference Optimization (DPO) algorithm. This algorithm optimizes preferences based on guide scores, effectively navigating the complexities and challenges associated with the expensive and often non-differentiable gradient calculations during the guided sampling fine-tuning process. Evaluated using the nuScenes dataset our model provides a strong baseline for balancing realism, diversity and controllability in the traffic scenario generation.

Figures

Figures reproduced from arXiv: 2502.12178 by the authors.

Figure 1
Figure 1. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Model architecture of training and inference. (a) The model is trained to predict clean trajectories from noisy ones. The input trajectory and traffic information are first encoded, then processed through Transformer blocks. (b) In the inference phase, classifier-free sampling controls the amount of future feature information provided to the sample, while guided sampling directs the sample generation towards the des… view at source ↗
Figure 4
Figure 4. Examples of sample data exhibiting collisions from [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Qualitative results of MuDi and MuDi-Pro. (a) Qualitative results demonstrate that our model can produce diverse samples with varying w. (b) Qualitative results of comparing MuDi-Pro with MuDi. MuDi-Pro produces plausible and realistic trajectories in scenes where MuDi…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 24 canonical work pages

  1. [34]

    Diffusion model alignment using direct preference optimization

    Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. arXiv preprint arXiv:2311.12908 , 2023

    arXiv 2023

  2. [1]

    Data-driven Traffic Simulation: A Comprehensive Review

    Di Chen, Meixin Zhu, Hao Yang, Xuesong Wang, and Yinhai Wang. Data-driven traffic simulation: A comprehensive review.arXiv preprint arXiv:2310.15975, 2023

    work page Pith review arXiv 2023

  3. [2]

    Driving to safety: How many miles of driving would it take to demonstrate autonomous vehicle reliability? Transportation Research Part A: Policy and Practice , 94:182–193, 2016

    Nidhi Kalra and Susan M Paddock. Driving to safety: How many miles of driving would it take to demonstrate autonomous vehicle reliability? Transportation Research Part A: Policy and Practice , 94:182–193, 2016

    work page 2016

  4. [3]

    Language- guided traffic simulation via scene-level diffusion

    Ziyuan Zhong, Davis Rempe, Yuxiao Chen, Boris Ivanovic, Yulong Cao, Danfei Xu, Marco Pavone, and Baishakhi Ray. Language- guided traffic simulation via scene-level diffusion. arXiv preprint arXiv:2306.06344, 2023

    arXiv 2023

  5. [4]

    Reinforcement Learning with Human Feedback for Realistic Traffic Simulation

    Yulong Cao, Boris Ivanovic, Chaowei Xiao, and Marco Pavone. Reinforcement learning with human feedback for realistic traffic simulation. arXiv preprint arXiv:2309.00709 , 2023

    work page Pith review arXiv 2023

  6. [5]

    Realgen: Retrieval augmented generation for controllable traffic scenarios

    Wenhao Ding, Yulong Cao, Ding Zhao, Chaowei Xiao, and Marco Pavone. Realgen: Retrieval augmented generation for controllable traffic scenarios. arXiv preprint arXiv:2312.13303 , 2023

    arXiv 2023

  7. [6]

    Scenecontrol: Diffusion for controllable traffic scene generation

    Jack Lu, Kelvin Wong, Chris Zhang, Simon Suo, and Raquel Urtasun. Scenecontrol: Diffusion for controllable traffic scene generation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16908–16914. IEEE, 2024

    work page 2024

  8. [7]

    Direct preference optimiza- tion: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Man- ning, Stefano Ermon, and Chelsea Finn. Direct preference optimiza- tion: Your language model is secretly a reward model. Advances in Neural Information Processing Systems , 36, 2024

    work page 2024

Show all 46 references
  1. [8]

    Task switching network for multi-task learning

    Guolei Sun, Thomas Probst, Danda Pani Paudel, Nikola Popovi ´c, Menelaos Kanakis, Jagruti Patel, Dengxin Dai, and Luc Van Gool. Task switching network for multi-task learning. In Proceedings of the IEEE/CVF international conference on computer vision , pages 8291– 8300, 2021

    2021

  2. [9]

    Parameter-efficient multi-task fine-tuning for trans- formers via shared hypernetworks

    Rabeeh Karimi Mahabadi, Sebastian Ruder, Mostafa Dehghani, and James Henderson. Parameter-efficient multi-task fine-tuning for trans- formers via shared hypernetworks. arXiv preprint arXiv:2106.04489 , 2021

    2021 arXiv

  3. [10]

    nuscenes: A multimodal dataset for autonomous driving

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

    2020

  4. [11]

    Toward benchmarking of microscopic traffic flow models

    Elmar Brockfeld, Reinhart D K ¨uhne, Alexander Skabardonis, and Peter Wagner. Toward benchmarking of microscopic traffic flow models. Transportation research record, 1852(1):124–129, 2003

    2003

  5. [12]

    Microscopic traffic simulation using sumo

    Pablo Alvarez Lopez, Michael Behrisch, Laura Bieker-Walz, Jakob Erdmann, Yun-Pang Fl ¨otter¨od, Robert Hilbrich, Leonhard L ¨ucken, Johannes Rummel, Peter Wagner, and Evamarie Wießner. Microscopic traffic simulation using sumo. In 2018 21st international conference on intellig...

    2018

  6. [13]

    Multipath: Multiple probabilistic anchor trajectory hypothe- ses for behavior prediction

    Yuning Chai, Benjamin Sapp, Mayank Bansal, and Dragomir Anguelov. Multipath: Multiple probabilistic anchor trajectory hypothe- ses for behavior prediction. arXiv preprint arXiv:1910.05449 , 2019

    1910 arXiv

  7. [14]

    Scept: Scene- consistent, policy-based trajectory predictions for planning

    Yuxiao Chen, Boris Ivanovic, and Marco Pavone. Scept: Scene- consistent, policy-based trajectory predictions for planning. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17103–17112, 2022

    2022

  8. [15]

    Trajectron++: Dynamically-feasible trajectory forecasting with heterogeneous data

    Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectron++: Dynamically-feasible trajectory forecasting with heterogeneous data. In Computer Vision–ECCV 2020: 16th Eu- ropean Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVIII 16, pages...

    2020

  9. [16]

    Trace and pace: Con- trollable pedestrian animation via guided trajectory diffusion

    Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Con- trollable pedestrian animation via guided trajectory diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, ...

    2023

  10. [17]

    Auto-encoding variational bayes

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 , 2013

    2013 arXiv

  11. [18]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  12. [19]

    Motiondiffuser: Controllable multi- agent motion prediction using diffusion

    Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi- agent motion prediction using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9644–9653, 2023

    2023

  13. [20]

    Scenedm: Scene-level multi-agent trajectory generation with consistent diffusion models

    Zhiming Guo, Xing Gao, Jianlan Zhou, Xinyu Cai, and Botian Shi. Scenedm: Scene-level multi-agent trajectory generation with consistent diffusion models. arXiv preprint arXiv:2311.15736 , 2023

    2023 arXiv

  14. [21]

    Generating driving scenes with diffusion

    Ethan Pronovost, Kai Wang, and Nick Roy. Generating driving scenes with diffusion. arXiv preprint arXiv:2305.18452 , 2023

    2023 arXiv

  15. [22]

    Scenario diffusion: Controllable driving scenario generation with diffusion

    Ethan Pronovost, Meghana Reddy Ganesina, Noureldin Hendy, Zeyu Wang, Andres Morales, Kai Wang, and Nick Roy. Scenario diffusion: Controllable driving scenario generation with diffusion. Advances in Neural Information Processing Systems , 36, 2024

    2024

  16. [23]

    Planning with diffusion for flexible behavior synthesis

    Michael Janner, Yilun Du, Joshua B Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. arXiv preprint arXiv:2205.09991, 2022

    2022 arXiv

  17. [24]

    Diffscene: Diffusion-based safety-critical scenario generation for au- tonomous vehicles

    Chejian Xu, Ding Zhao, Alberto Sangiovanni-Vincentelli, and Bo Li. Diffscene: Diffusion-based safety-critical scenario generation for au- tonomous vehicles. In The Second Workshop on New Frontiers in Adversarial Machine Learning , 2023

    2023

  18. [25]

    King: Generating safety-critical driving scenarios for robust imitation via kinematics gradients

    Niklas Hanselmann, Katrin Renz, Kashyap Chitta, Apratim Bhat- tacharyya, and Andreas Geiger. King: Generating safety-critical driving scenarios for robust imitation via kinematics gradients. In European Conference on Computer Vision , pages 335–352. Springer, 2022

    2022

  19. [26]

    A diffusion-model of joint interactive navigation

    Matthew Niedoba, Jonathan Lavington, Yunpeng Liu, Vasileios Li- outas, Justice Sefas, Xiaoxuan Liang, Dylan Green, Setareh Dabiri, Berend Zwartsenberg, Adam Scibior, et al. A diffusion-model of joint interactive navigation. Advances in Neural Information Processing Systems, 36, 2024

    2024

  20. [27]

    Behav- iorgpt: Smart agent simulation for autonomous driving with next-patch prediction

    Zikang Zhou, Haibo Hu, Xinhong Chen, Jianping Wang, Nan Guan, Kui Wu, Yung-Hui Li, Yu-Kai Huang, and Chun Jason Xue. Behav- iorgpt: Smart agent simulation for autonomous driving with next-patch prediction. arXiv preprint arXiv:2405.17372 , 2024

    2024 arXiv

  21. [28]

    Language-driven interactive traffic trajectory generation

    Junkai Xia, Chenxin Xu, Qingyao Xu, Chen Xie, Yanfeng Wang, and Siheng Chen. Language-driven interactive traffic trajectory generation. arXiv preprint arXiv:2405.15388 , 2024

    2024 arXiv

  22. [29]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 2250...

    2023

  23. [30]

    An image is worth one word: Personalizing text-to-image generation using textual inversion

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618 , 2022

    2022 arXiv

  24. [31]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836– 3847, 2023

    2023

  25. [32]

    Low-rank adaptation for fast text-to-image diffusion fine- tuning, 2023

    Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine- tuning, 2023

    2023

  26. [33]

    Svdiff: Compact parameter space for diffusion fine-tuning

    Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. arXiv preprint arXiv:2303.11305 , 2023

    2023 arXiv

  27. [35]

    Aligndiff: Aligning diverse human preferences via behavior-customisable diffu- sion model

    Zibin Dong, Yifu Yuan, Jianye Hao, Fei Ni, Yao Mu, Yan Zheng, Yujing Hu, Tangjie Lv, Changjie Fan, and Zhipeng Hu. Aligndiff: Aligning diverse human preferences via behavior-customisable diffu- sion model. arXiv preprint arXiv:2310.02054 , 2023

    arXiv

  28. [36]

    Feedback efficient online fine- tuning of diffusion models

    Masatoshi Uehara, Yulai Zhao, Kevin Black, Ehsan Hajiramezanali, Gabriele Scalia, Nathaniel Lee Diamant, Alex M Tseng, Sergey Levine, and Tommaso Biancalani. Feedback efficient online fine- tuning of diffusion models. arXiv preprint arXiv:2402.16359 , 2024

    2024 arXiv

  29. [37]

    Guided conditional dif- fusion for controllable traffic simulation

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

    2023

  30. [38]

    A. D. Luca, G. Oriolo, and C. Samson. Feedback Control of a Nonholonomic Car-Like Robot , pages 171–249. Springer, Berlin, Germany, 1988

    1988

  31. [39]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Ma- chine Learning, pages 8162–8171. PMLR, 2021

    2021

  32. [40]

    Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mo- hammad Norouzi, and David J. Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022

    2022 arXiv

  33. [41]

    Accurate, large minibatch sgd: Training imagenet in 1 hour

    Priya Goyal, Piotr Doll ´ar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaim- ing He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677 , 2017

    2017 arXiv

  34. [42]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Confer- ence on Computer Vision , pages 4195–4205, 2023

    2023

  35. [43]

    Using human feedback to fine-tune diffusion models without any reward model

    Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Qimai Li, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. arXiv preprint arXiv:2311.13231, 2023

    2023 arXiv

  36. [44]

    Exploration-driven policy optimization in rlhf: Theoretical insights on efficient data utilization

    Yihan Du, Anna Winnicki, Gal Dalal, Shie Mannor, and R Srikant. Exploration-driven policy optimization in rlhf: Theoretical insights on efficient data utilization. arXiv preprint arXiv:2402.10342 , 2024

    2024 arXiv

  37. [45]

    Bits: Bi-level imitation for traffic simulation

    Danfei Xu, Yuxiao Chen, Boris Ivanovic, and Marco Pavone. Bits: Bi-level imitation for traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA) , pages 2929–2936. IEEE, 2023

    2023

  38. [46]

    Generating useful accident-prone driving scenarios via a learned traffic prior

    Davis Rempe, Jonah Philion, Leonidas J Guibas, Sanja Fidler, and Or Litany. Generating useful accident-prone driving scenarios via a learned traffic prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17305–17315, 2022

    2022

Pith tools

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