- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
The paper studies the physical plausibility in AR-diffusion video models with few-step generation. It creates a new benchmark , PLV-Bench, which has real reference videos, and use a VLM to conduct evaluation. It also proposes PhysDMD, a method to augment the DMD with a reward training objective using world model (V-JEPA) signal. They show that their method can perform better than vanilla-DMD on Physics-IQ-Verified benchmark, and also their own benchmark.
Critical strengths
- The scope of the problem (physics plausibility in video generative models) is relevant, and the efficiency for video models is also important.
- They provides a new benchmakr and a new method, for evaluation and improving physics plausibility in AR video generative models.
Critical weaknesses
- The novelty of the paper seems limited. Using V-JEPA (world model) as a signal for physics plausibility has been studied in WMReward [1], where they use it during the inference time for search and guidance. Using RL to tilt the distribution during DMD has been studied in RTDMD [2], where they show that this combination can improve the text-to-image models' performance.
- The contribution of the introduction of the new benchmark seems incremental. Apart from Physics-IQ and Physics-IQ-Verified, there have been other I2V physics benchmakrs that contain real references, e.g., Physion-Eval [3] and PAI-Bench [4].
- The presentation of the PLV-bench is not sufficiently clear. The most concerning part is the validation of the evaluation protocol of the proposed benchmakr. There is no validation of the VLM judge performance, and it is thus hard to assess the reliability of the proposed new benchmark.
- The usefulness of the proposed method on open-source AR video generative models is missing. The paper applies their method on their self-trained ckpt, and shows improvment. It would be more convincing to apply their method on open-source models.
[1] https://arxiv.org/abs/2601.10553
[2] https://arxiv.org/abs/2605.26108
Questions for authors
- please explain the difference and contribution of the proposed method compared to existing literature.
- please explain the difference and contribution of the proposed benchmarks compared to existing literature.
- Do you have human evaluations on the reliability of the VLM judge?
- Is it possible to apply your proposed method to open sourced models? If so, how would be the results?
- what is the plus minus in the tables? Are they standard deviation or confidence internal or others?
- It seems that PhysDMD can recover the teacher model's physics performance, but cannot surpass it, as shown in ablation. Is it expected?
- Please double check the citation of Physics-IQ-Verified. It seems that the authors cited the original version but not the verified version.
5 comments
Response to Reviewer gGjJ [Part 2]
Q2: Differences Between PLV-Bench and Existing Physics Benchmarks
A2: We thank the reviewer for highlighting Physion-Eval [3] and PAI-Bench [4]. While these benchmarks utilize real-world references, PLV-Bench fundamentally differs in its evaluation objective and how these references are used to assess sustained physical evolution.
Differences from Existing Benchmarks: Physion-Eval identifies temporally localized violations via expert annotation, and PAI-Bench broadly assesses physical perception. In contrast, PLV-Bench explicitly targets sustained physical evolution in autoregressive (AR) video generation, tackling the compounding errors and omitted intermediate stages inherent to extended long-horizon rollouts.
Methodological Design of PLV-Bench:
- Long-Horizon Physical Processes: We systematically curate 125 real-world videos spanning at least 10 seconds across 16 physical phenomena. We employ rigorous automated filtering and human verification to anchor multi-stage physical events.
- Reference-Guided Phase-Level Evaluation: Rather than directly comparing generated videos with reference frames, we construct reference-specific rubrics describing core events, ordered physical phases, and causal constraints. These rubrics are frozen before evaluating generated candidates, allowing plausible variations in timing and trajectories.
- Decoupled Assessment Dimensions: We independently assess Event Completion, Physical Plausibility, Temporal Evolution, and Appearance Consistency, strictly distinguishing true structural physics failures from superficial appearance degradation or incomplete progressions.
Qualitative Examples: Figure 3 visualizes the phase-level evaluation of ordered physical processes. Additionally, Appendices E.1 and F detail further reference-guided scoring examples (Figs. 9–11) and common generation failures (Figs. 16–18), respectively.
Revision: We clarified these distinctions (App. A.1), expanded the protocol (Sec. 2.3), and added qualitative examples (Apps. E.1, F).
W3 & Q3: Reliability and Human Validation of VLM-Based Physical Plausibility Evaluation
A3: We thank the reviewer for their feedback on the reliability of our VLM-based evaluation. To address this, we detail our reference-guided protocol and provide new evidence through evaluator repeatability and human validation.
Evaluation Protocol: PLV-Bench employs a two-stage process. First, the VLM constructs a frozen, reference-specific rubric (capturing core events, ordered phases, and causal constraints) strictly from the real video. Anonymized candidates are then evaluated independently via phase-level alignment rather than rigid frame-wise matching, accommodating plausible variations in timing and trajectories.
Evaluator Repeatability: Repeated GPT-5.6-Sol evaluations on 50 randomly sampled PLV-Bench cases yielded exact score agreements of 82% to 88% across all four dimensions, demonstrating high consistency.
Human–VLM Agreement: We further conducted a human study involving 11 participants and 60 generated videos spanning all 16 physical categories. Before annotation, all participants were familiarized with the evaluation guidelines, dimension definitions, and five-point scoring criteria. Each participant evaluated 40 videos, yielding 1,320 dimension-level judgments.
| Scoring Dimension | Exact Match ↑ | 1-Point Difference |
|---|---|---|
| Event Completion | 80.9% | 13.0% |
| Physical Plausibility | 77.7% | 11.6% |
| Temporal Evolution | 83.4% | 10.0% |
| Overall | 80.7% | 11.5% |
Results: The VLM achieves 80.7% exact agreement with human assessments.
Revision: We expanded the evaluation protocol (Sec. 2.3, App. B.5), added human validation results (Sec. 4.6, Table 5), and provided qualitative examples (App. E.1, Figs. 9–11).
Reference:
[3] Zhang, Qin, et al. "Physion-eval: Evaluating physical realism in generated video via human reasoning." arXiv preprint arXiv:2603.19607 (2026).
[4] Zhou, Fengzhe, et al. "Pai-bench: A comprehensive benchmark for physical ai." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.
Response to Reviewer gGjJ [Part 3]
W4 & Q4: Generalization to Open-Source Autoregressive Video Models
A4: We thank the reviewer for suggesting validation on existing open-source AR video models. Our primary PhysDMD model is built upon the open-source Wan2.1-14B backbone, which we causally adapt and distill into a four-step AR generator before physics-reinforced post-training.
Generalization to Causal Forcing++: To further validate the transferability of our framework, we applied PhysDMD directly to the pretrained Causal Forcing++, an independently developed open-source 1.3B AR video model with two-step inference. We preserved its original architecture and sampling configuration, applying only additional physics-reinforced post-training. All variants were evaluated on the Physics-IQ-Verified benchmark using four independent random seeds, with results reported as mean ± standard deviation.
| Method | NFE | ST-IoU ↑ | Sp-IoU ↑ | wSp-IoU ↑ | MSE-based ↑ | Overall ↑ |
|---|---|---|---|---|---|---|
| Causal Forcing++ | 2 | 23.19 ± 1.61 | 32.81 ± 0.68 | 18.48 ± 0.30 | 16.58 ± 0.31 | 22.76 ± 0.35 |
| + PhysDMD (50 steps) | 2 | 23.60 ± 1.52 | 32.75 ± 0.71 | 18.29 ± 0.51 | 16.88 ± 0.28 | 22.88 ± 0.30 |
| + PhysDMD (100 steps) | 2 | 23.50 ± 1.49 | 33.02 ± 0.95 | 18.41 ± 0.58 | 17.13 ± 0.39 | 23.02 ± 0.62 |
| + PhysDMD (150 steps) | 2 | 23.53 ± 1.08 | 33.22 ± 0.81 | 18.54 ± 0.39 | 17.63 ± 0.45 | 23.23 ± 0.31 |
Empirical Results: PhysDMD progressively improves the Overall score from 22.76 to 23.23 within 150 additional training steps in Physics-IQ-Verified Benchmark, yielding improvements across all four metrics at the final checkpoint. Importantly, these gains are achieved while retaining the original two-step inference budget, introducing zero inference-time world-model overhead.
Revision: We added these results (Sec. 4.5, Table 4) and detailed the training configurations and intermediate checkpoints (App. D.7, Table 11).
Q5: Clarification of Statistical Reporting (Mean ± Standard Deviation)
A5: Thank you for pointing this out. Following the Physics-IQ evaluation protocol, we perform inference using four different random seeds and report the results as mean ± standard deviation. To avoid ambiguity, we have clarified this notation in the revised table caption.
Q6: Is It Expected That PhysDMD Approaches but Does Not Surpass the Teacher?
A6: We thank the reviewer for this insightful question. As stated in our Problem Formulation (Line 133), our primary objective is to recover the physical plausibility degraded during few-step AR video distillation while preserving inference efficiency, rather than strictly surpassing the teacher's generation quality.
- Theoretical Justification: Standard DMD minimizes a reverse KL divergence [5]. This mode-seeking behavior inherently bounds student capacity [6], and distribution matching alone fails to explicitly enforce physical plausibility.
- Empirical Restoration: As shown in Table 2, Vanilla-DMD drops significantly behind the 100-NFE AR teacher (25.31 vs. 28.22 Overall), a performance gap that continued standard DMD training fails to resolve (25.99). In contrast, PhysDMD effectively restores these physical dynamics, achieving 28.10 Overall in just 4 NFE to closely approach the teacher.
- Fine-Grained Improvements: Notably, PhysDMD actually outperforms the teacher on specific fine-grained metrics, including mean Sp-IoU (31.13 vs. 30.09) and MSE-based scores (32.24 vs. 31.54).
Our conclusion in Sec. 6 is therefore that PhysDMD approaches the multi-step baseline while retaining four-step sampling and requiring no inference-time world model.
7. Correction of the Physics-IQ-Verified Citation.
Thank you for pointing this out. We have revised the paper accordingly and have updated the reference.
We would like to sincerely thank you once again for your time and effort in reviewing our paper.
Reference:
[5] Yin, Tianwei, et al. "One-step diffusion with distribution matching distillation." CVPR 2024.
[6] Wu, Tianhe, et al. "Diversity-preserved distribution matching distillation for fast visual synthesis." ICML 2026.
Response to Reviewer gGjJ [Part 1]
We sincerely thank the reviewer for their constructive feedback and for recognizing the importance of physical plausibility in efficient video generation. In response, we have clarified our methodological and benchmark contributions, strengthened PLV-Bench validation through human evaluation and additional analyses, and demonstrated PhysDMD's generalizability on an open-source AR video model. We have also improved our experimental reporting and citations. Detailed responses are provided below.
A1: We thank the reviewer for highlighting WMReward [1] and RTDMD [2]. PhysDMD fundamentally differs from these works in its motivation, supervision formulation, and target generation setting:
Results and Core Contribution: PhysDMD improves the Physics-IQ-Verified Overall score from 25.31 to 28.10, approaching the multi-step teacher (28.22) while retaining four-step inference. On PLV-Bench, Physical Plausibility improves from 3.64 to 3.77 over Vanilla-DMD. Our core contribution is the design and integration of reference-calibrated dynamics supervision tailored to few-step autoregressive video distillation, which integrates world-model conditioning with direct dynamics alignment and reward-guided optimization across different denoising stages.
Our empirical analyses further motivate two key design choices:
1. Reference-Calibrated Reward: Directly minimizing prediction energy can favor static, easily predictable motion. To investigate this issue, we evaluated 1,000 real-video clips, comparing continuations slowed to 75% of their original speed against frozen counterparts.
Our symmetric relative-Huber formulation penalizes both excessively high and abnormally low prediction energies relative to the paired real transition. Paired calibration improves the preference rate by 27.9% over shuffled references, demonstrating the necessity of transition-specific supervision.
2. History-Decoupled Supervision for AR Generation: AR generation depends on previously generated frames, making world-model energy sensitive to historical errors. In a separate 1,000-clip diagnostic, replacing clean real history with generated history reversed the generated-versus-real energy ordering in 39.7% of cases.
To address this, PhysDMD decouples generation conditioning from reward evaluation: the generator is trained with augmented histories (i.e., real-video history features with augmentations) to improve robustness to imperfect rollouts, while the world model always evaluates generated and real continuations against the identical clean history. This calibrated signal supports two complementary objectives: Direct Dynamics Alignment supervises the final denoising prediction, while Reward Optimization guides intermediate stochastic transitions through GRPO. Both operate without requiring full-history self-rollouts.
Revision: We clarified our differences in Section 5.3, and expanded the supervision details in Section 3.2 and Appendices C.1–C.2.
References:
[1] Yuan, Jianhao, et al. "Inference-time physics alignment of video generative models with latent world models." arXiv (2026).
[2] Huang, Yushi, et al. "Reinforcing Few-step Generators via Reward-Tilted Distribution Matching." arXiv (2026).