acceptodds
Withdrawn from ICLR 2027

PhysGuide: Physics-Principles-Guided Learning for Physically Plausible Video Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Text-to-video (T2V) models have achieved remarkable visual fidelity, yet their generated videos still frequently violate physical laws. Preference-based post-training provides a promising way to improve physical plausibility, but existing approaches often represent physics supervision using a scalar signal, which collapses the principle-wise structure of physical behavior and provides limited information about how preferred and rejected videos differ physically. We introduce PhysGuide, a physics principles-guided preference optimization algorithm for physically plausible video generation. We construct a set of physics principle vectors from real-world demonstrations and project preferred and rejected videos into this representation space. PhysGuide uses their principle-wise differences to independently modulate the preferred and rejected policy rewards, producing physics-guided rewards along individual physical dimensions. We further introduce a physics-violation divergence penalty that captures the overall magnitude of physical disagreement between the pair, so that more severe violations produce a stronger corrective signal during optimization. We evaluate PhysGuide on PhyGenBench and VideoPhy and observe consistent improvements over the baselines, while preserving general video quality on VBench. Ablations further show that the principle-specific physics-guided rewards and physics-violation divergence penalty contribute complementary improvements to physical plausibility.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video Generation, Physics Plausibility

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_9CyQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PhysGuide improves the physical plausibility of generated videos through physics-guided preference optimization. It constructs physics prototypes from real videos and textual descriptions, then uses prototype associations to modulate preference rewards and introduce an adaptive margin. Experiments report improvements on PhyGenBench and VideoPhy, alongside improved VBench scores.

Critical strengths

1 The reported improvements are substantial. Compared with standard DPO, PhysGuide improves PhyGenBench from 53.1 to 63.1 and VideoPhy from 40.7 to 45.3. The VBench results also suggest that these gains do not sacrifice general video quality.

2 The ablations provide useful evidence for both reward modulation and the divergence-based margin. Comparisons against random grouping and predefined physics categories further support the proposed prototype construction.

Critical weaknesses

1 The interpretation of prototype distance as physical-violation severity is insufficiently validated. The projection measures association with physical phenomena, which does not necessarily indicate whether those phenomena are depicted correctly. Appendix A.3 demonstrates robustness to color changes, but its comparisons across different videos also change scene content and semantics. Consequently, it does not establish that the distance isolates physical violations or tracks their severity.

2 Important implementation details are missing, including the video encoder and its relationship to the generator, the EMA update procedure, and how the policy log-ratio is computed for the diffusion model. It is also unclear whether gradients flow through the prototype projections and adaptive margin. These details matter because an evolving representation could reduce the margin by changing the representation itself, rather than improving generated physics.

Questions for authors

1 Can you validate the proposed distance using matched scenes with controlled physical violations of increasing severity, while separately varying appearance?

2 Which video encoder and diffusion log-ratio estimator are used? Are gradients stopped through the projections and margin, and how are the encoder and prototypes updated?

3 Were the preference-training baselines evaluated using the same training pairs and comparable optimization budgets?

4 comments

We thank the reviewer for their thoughtful comments and constructive feedback on the formulation of our physics-principle-guided objective, clarifying the implementation details, and the experimental validation of our approach.


  1. The interpretation of prototype distance as physical-violation severity is insufficiently validated.

We clarify that our physics-violation divergence term measures differences in association across multiple physics principles, rather than directly measuring the severity of a single physical violation.

Each video is represented by a set of scalar association scores corresponding to different physical principles, such as collision dynamics, momentum conservation, and object interactions. Physically implausible videos may exhibit weaker associations with the relevant principles compared to physically plausible videos. Consequently, videos with more severe physical violations may have smaller association scores than those with milder violations. When comparing physically plausible and implausible videos, larger differences in their principle-association distributions can lead to a higher physics-violation divergence penalty, g(xw,xl)g(x^w,x^l).

Empirical validation. To investigate this relationship, we conduct an analysis on the PhysicsIQ test set. We use the PhysicsIQ evaluation error as a proxy for physical violation severity, where higher error values indicate greater deviations from expected physical behavior, while lower values indicate more physically plausible dynamics. For each evaluated video pair, we compute the corresponding physics-violation divergence penalty, g(xw,xl)g(x^w,x^l), and examine its relationship with the measured physics error. The results are summarized below:

Error valuesAvg. divergence penalty
Higher violation: [50, 100]0.84
Lower violation: [0, 50)0.35

Our analysis shows that higher physics evaluation errors are associated with larger divergence penalties (0.84 vs. 0.35). This suggests that the proposed divergence term captures principle-level discrepancies that are correlated with the severity of physical violations.

AuthorsOpenReview ↗

Rebuttal - 2

  1. Important implementation details are missing

Gradient Flow. We clarify that gradients flow exclusively through the differentiable policy rewards rθwr_\theta^w and rθlr_\theta^l, updating only the trainable LoRA parameters of the HunyuanVideo model, while the pretrained backbone remains frozen. The physics-principle representations qwq^w and qlq^l, and the divergence term g(xw,xl)g(x^w,x^l), do not receive gradients and are detached from the computational graph.

Consequently, for the scenario πθ=πref\pi_\theta = \pi_{\mathrm{ref}}, the model cannot directly minimize the preference loss by shrinking the distance gg between the preferred and rejected representations, which addresses the reviewer's concern about representation collapse. Since the LoRA-adapted video features evolve during training, we update the physics-principle prototypes using an exponential moving average (EMA) of the corresponding cluster embeddings.

Diffusion Likelihood Estimation. We clarify that we do not explicitly compute the exact video likelihoods. Instead, following diffusion/flow-matching DPO [1], we approximate the policy log-likelihood ratio using the difference in velocity-prediction errors between the trainable HunyuanVideo model and a frozen reference model.

For each preferred and rejected video, we encode the video into latent space, sample a random timestep and Gaussian noise, and construct noisy latents using the flow-matching schedule. Both the trainable policy and the reference model predict the velocity from the same noisy latents.

We estimate the implicit policy reward as:

r^θ(x,c)=\hat{r}_{\theta}(x,c) =

=βeff[∥vref(zt,t,c)−ut∥22−∥vθ(zt,t,c)−ut∥22]=\beta_{\mathrm{eff}} \left[ \lVert v_{\mathrm{ref}}(z_t,t,c) - u_t \rVert_2^2 - \lVert v_{\theta}(z_t,t,c) - u_t \rVert_2^2 \right]

where ut=ϵ−z0u_t=\epsilon-z_0 is the target flow velocity and βeff\beta_{\mathrm{eff}} includes the loss scaling.

This provides a differentiable surrogate for the policy reward, with gradients propagated through the trainable model while the reference model remains frozen. The estimated rewards for the preferred and rejected videos are then used in our principle-guided preference objective (Eq. 8).

EMA Updates. All codes live in the embedding space of an encoder fϕf_\phi that maps a video xx to a feature vector in Rd\mathbb{R}^d. Instead of keeping this encoder fixed, we update its parameters with an exponential moving average (EMA) of the corresponding encoder parameters in the generator:

ϕt=αϕt−1+(1−α)θtenc,\phi_t = \alpha \phi_{t-1} + (1-\alpha)\theta_t^{\mathrm{enc}},

where θtenc\theta_t^{\mathrm{enc}} denotes the visual encoder parameters of the generator at DPO step tt, and α∈(0,1)\alpha \in (0,1) is the EMA decay coefficient.

Video and Text Encoder. We clarify that the video encoder fϕf_\phi and text encoder are part of the pretrained HunyuanVideo model used for preference optimization. The pretrained backbone weights are frozen, while the LoRA parameters are trainable. Video features are extracted from the HunyuanVideo diffusion transformer (DiT) hidden states.

[1]. Wallace, Bram, et al. "Diffusion model alignment using direct preference optimization." 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024.

Thank you for clarifying the detached gradient paths and reward estimator, and for providing the matched PhyGDPO comparison. These address several of my concerns. However, the two-bin PhysicsIQ analysis does not yet resolve my main concern: differences in scene content and physical phenomena may still confound the relationship between divergence and violation severity. Could you provide a controlled within-scene comparison, or narrow the interpretation of this term accordingly? Please also clarify whether EMA is applied to encoder parameters, prototype embeddings, or both. I am currently maintaining my rating pending these clarifications.

We thank the reviewer for their response and feedback. We discuss the concerns as follows:

Relationship between divergence and violation severity (controlled physical violations of increasing severity) To evaluate the relationship between divergence and violation severity in a controlled within-scene comparison, we select 10 scenes from Physics-IQ (which will be winner videos) and generate 10 loser videos using different HunyuanVideo seeds while keeping the conditioning input fixed. This setting keeps the scene context fixed and only the physics phenomenon changes. For each scene, the generated videos are ranked according to the severity of their physical violations based on PhysicsIQ scores, from 1 (least severe violation, high PhysicsIQ score) to 10 (most severe violation, low PhysicsIQ score). The PhysicsIQ scores are in a range of [0-100] where 100 is better. We then compute the physics-violation divergence (g(xw,xl)) for every winner-loser pair and rank the divergence scores as well. Next, we compute the Spearman score between the ranking and report the results below.

Scene IDPhysics-IQ Score (Mean ± SD) (High is better)Divergence g (Mean ± SD) (Low is better)Spearman rho
S171.2 ± 10.10.24 ± 0.12-0.867
S267.5 ± 12.20.33 ± 0.14-0.745
S360.2 ± 15.80.35 ± 0.17-0.879
S456.8 ± 17.60.37 ± 0.16-0.733
S552.1 ± 15.60.39 ± 0.15-0.939
S643.8 ± 15.90.46 ± 0.18-0.842
S738.2 ± 18.10.51 ± 0.21-0.988
S829.4 ± 11.30.59 ± 0.16-0.830
S925.7 ± 9.50.64 ± 0.14-0.782
S1022.6 ± 8.20.69 ± 0.12-0.939
Mean scene-wise rho——-0.854

From the above table, we can see that the negative within-scene Spearman correlations support that videos with lower physical plausibility scores exhibit greater divergence from their corresponding physically plausible winners.

EMA application. We apply the exponential-moving-average (EMA) to the encoder parameters only and use it to update the physics principle vectors throughout training.

Reviewer_FjyoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PhysGuide, a physics-principle-guided preference optimization framework for improving the physical plausibility of text-to-video generation. The method constructs physics principle vectors from real-world videos and incorporates principle-specific reward modulation and a physics-violation divergence penalty into preference optimization. Experiments on PhyGenBench and VideoPhy demonstrate improved physical plausibility over existing baselines, while evaluations on VBench suggest that general video quality is also preserved.

Critical strengths
  1. The paper introduces a physics-principle representation that preserves multi-dimensional physical information, addressing the limitations of scalar physics rewards. The principle-specific reward modulation provides a more structured learning signal for distinguishing physically plausible and implausible videos.

  2. The method demonstrates consistent improvements on PhyGenBench and VideoPhy while maintaining general video quality on VBench. The ablation studies further investigate the complementary contributions of physics-guided rewards and the physics-violation divergence penalty, supporting the effectiveness of the proposed design.

Critical weaknesses

1、Figure 2: The background color of the “Overall Objective Loss Function” box should be adjusted to improve visual contrast and consistency with the rest of the figure. In addition, a brief explanation of the objective function should be provided to clarify how the physics-guided rewards and the physics-violation divergence penalty jointly contribute to the optimization.

2、The paper uses a VLM (GPT-5.4) to identify physical principles in videos and construct physics principle vectors, but it remains unclear whether the VLM is fine-tuned for physical reasoning or used as an off-the-shelf model. More importantly, the paper does not sufficiently validate the reliability of the generated physical annotations. Since these annotations directly determine the clustering and initialization of the physics principle vectors, evaluating their accuracy would help establish the reliability of the proposed physics representation.

3、Although PhysGuide incorporates physics-principle representations into the preference optimization objective and demonstrates improved qualitative results, it remains unclear whether the policy learns a more generalizable understanding of physical phenomena. The current visualizations mainly compare generated videos and do not directly validate whether the learned representations capture the intended physical principles. Additional analyses of principle-wise representations and generalization to unseen physical interactions would strengthen the claim that the proposed method improves physical understanding rather than merely enhancing the generated outcomes on selected examples.

4、Figure 2: Many mathematical symbols in the figure appear to be non-selectable. The authors should ensure that mathematical notations are rendered as selectable text when preparing the figure to improve accessibility and usability.

Questions for authors

see Weaknesses

1 comment

We thank the reviewer for their thoughtful comments and constructive feedback on the figures, the accuracy of VLM for constructing physics principle vectors, and generalization of our method.


1. Figure 2: background color; a brief explanation of the objective function

We thank the reviewer for the suggestion. We will update Figure 2 to improve the background color, visual contrast, and consistency of the “Overall Objective Loss Function” box. We also clarify that the two components of our objective play complementary roles. The physics-guided rewards use principle-wise differences between the preferred and rejected videos to separately modulate their policy rewards, while the physics-violation divergence term is calculated as the overall difference between their physics-principle representations. Together, these components encourage the model to favor physically plausible videos. We will add this explanation to the figure caption for clarity.


2. Whether the VLM is fine-tuned for physical reasoning or used as an off-the-shelf model. More importantly, the paper does not sufficiently validate the reliability of the generated physical annotations.

We clarify that GPT-5.4 is used as an off-the-shelf model without additional fine-tuning for physical reasoning. To evaluate the reliability of its annotations, we conducted a human study using 20 videos covering all eight physics-principle clusters. Two participants independently identified the physical principles present in each video using a shared taxonomy. We compared the GPT-5.4 annotations with the human annotations and obtained 96% agreement, demonstrating strong consistency between the VLM-generated annotations and human judgments. This supports the reliability of the physics annotations used to construct our principle vectors.


3. It remains unclear whether the policy learns a more generalizable understanding of physical phenomena.

Generalization to Unseen Physical Interactions. We train PhysGuide on the PhysicsIQ dataset and evaluate it on two independent physics benchmarks, VideoPhy and PhyGenBench. These benchmarks contain physical interactions and scenarios that are not present in the PhysicsIQ training dataset. To evaluate generalization, we identify a subset of unseen physical scenarios from PhyGenBench that are absent from PhysicsIQ but involve physical principles represented during training. We evaluate the base HunyuanVideo model, standard DPO, and PhysGuide on this subset using identical prompts, random seeds, and evaluation settings.

MethodMechanicsOpticsThermalMaterialOverall
HunyuanVideo35.255.833.743.143.3
HunyuanVideo + DPO44.358.740.257.651.3
Ours (PhysGuide)54.771.760.560.262.5

Analysis. PhysGuide consistently outperforms both the base HunyuanVideo model and standard DPO across all four physical categories on unseen interactions, achieving an overall score of 62.5 compared to 51.3 for DPO and 43.3 for HunyuanVideo. These results demonstrate that the improvements extend beyond the specific scenarios encountered during training. This suggests that PhysGuide learns transferable physics-related representations that improve generation on novel interactions involving familiar physical principles, rather than simply optimizing for previously seen scenarios.


4. Figure 2: Many mathematical symbols in the figure appear to be non-selectable.

We thank the reviewer for pointing this out. We will update Figure 2 to ensure that all mathematical symbols and notations are rendered as selectable text rather than embedded images, improving the accessibility, readability, and usability of the figure.

Reviewer_irmgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PhysGuide post-trains a text-to-video model using physics-principle representations. It clusters VLM descriptions of real videos, averages video features into prototypes, and uses winner–loser assignment differences to modulate DPO-style rewards, with a distance-based margin. Using HunyuanVideo, 200 real winner videos, and 600 generated losers, it reports improvements on PhyGenBench, VideoPhy, and VBench.

Critical strengths

S1. Relevant post-training objective. Using multiple physical cues to guide preference learning addresses an important limitation of coarse feedback. The method preserves the underlying generation architecture and standard text-to-video inference interface.

S2. Meaningful reported benchmark improvements. PhyGenBench improves from 45.6 to 63.1 and VideoPhy from 33.4 to 45.3 over HunyuanVideo. The corresponding advantages over PhysHPO are 8.7 and 3.7 points, respectively.

S3. Useful component and grouping controls. Table 2 separates standard preference training, reward modulation, and the additional margin. Table 3 includes random grouping and a comparison with predefined Physics-IQ groups at the same K, supporting the relevance of prototype construction.

S4. Complementary quality and human evaluation. VBench Total improves from 81.4 to 84.3. The reported human preference rates of 96% over the base model and 78% over standard preference training provide additional evidence of practical benefit.

Critical weaknesses

W1. The benefit of principle-wise optimization is not isolated from adaptive reward weighting. Eqs. (6)–(8) transform the same two full-video scalar rewards. With detached representations, the policy gradient remains a weighted combination of their gradients. A matched scalar-statistic reweighting/adaptive-margin control is needed to establish the specific benefit of retaining principle dimensions; adaptive DPO margins already have precedent, e.g., AlphaDPO.

W2. Principle-assignment distance is not validated as violation severity. Eq. (3) measures relative association with phenomenon clusters, rather than correctness of each physical principle. Consequently, Eq. (7) need not increase monotonically with error severity. Appendix A.3's tint controls are useful, but comparisons between different videos change both physical behavior and visual content, leaving this interpretation unresolved.

W3. Crucial implementation and gradient paths are unspecified. The paper does not identify the video/text encoders, explain the diffusion likelihood estimator in Eq. (4), or fully specify EMA updates and gradient flow through q and g. When πθ = πref, Eq. (8) becomes softplus(γg); if g receives gradients, this introduces a direct incentive to shrink representation distance. Detaching versus learning these signals therefore materially changes the algorithm.

W4. Advanced-baseline comparisons are not demonstrably matched. Table 2 provides a useful standard-DPO control, but the paper does not state whether PhysHPO/PhyGDPO were rerun on the same 200-winner/600-loser data with matched optimization budgets. Baseline checkpoints and evaluation configurations also need documentation before the head-to-head gains can be attributed to algorithmic differences.

W5. The ablations do not fully establish the claimed complementarity. Table 2 adds components progressively but omits a divergence-only variant, leaving its standalone contribution unresolved. No EMA ablation is reported despite being announced in Section 4. Defaults and sensitivity for the modulation strength, margin strength, and projection temperature are also missing.

W6. Evaluation reliability is insufficiently documented. The precise benchmark metrics, evaluator checkpoints, and evaluation splits are not specified. The human study reports large preference rates without participant counts, comparison counts, or uncertainty estimates. These omissions limit independent assessment of the quantitative and human-evaluation gains.

Questions for authors
  1. Please specify the implemented diffusion reward estimator and whether q/g are detached, including how the encoder and EMA prototypes are updated.

  2. Were advanced baselines rerun on the same 200-winner/600-loser data with matched budgets? Please clarify the evaluation metrics and human-study sample sizes supporting the reported gains.

3 comments

We thank the reviewer for their thoughtful comments and constructive feedback on the formulation of our physics-principle-guided objective, clarifying the implementation details, and the experimental validation of our approach including the baseline and ablations.


W1. The benefit of principle-wise optimization is not isolated from adaptive reward weighting.

To isolate this contribution, we introduce a scalar-statistic reweighting DPO baseline that aggregates the principle-wise discrepancies into a single reward-scaling factor while retaining the same physics-violation divergence margin. We define the principle-wise discrepancy and physics-violation divergence as:

dk=qkw−qkl,g=∥qw−ql∥22.d_k = q_k^w - q_k^l, \qquad g = \|q^w - q^l\|_2^2.

Derivation of scalar weighting factor. We average the principle-specific scaling factors for the preferred and dispreferred videos:

skw=1+λ[dk]+s_k^w = 1 + \lambda[d_k]_+

skl=1+λ[−dk]+s_k^l = 1 + \lambda[-d_k]_+

sw=1+λK∑k=1K[dk]+s^w = 1 + \frac{\lambda}{K}\sum_{k=1}^{K}[d_k]_+

sl=1+λK∑k=1K[−dk]+s^l = 1 + \frac{\lambda}{K}\sum_{k=1}^{K}[-d_k]_+

We define a single scalar as the average of these two factors:

s=sw+sl2s = \frac{s^w+s^l}{2}

After substituting and since [d_k]+ + [-d_k]+ = |d_k|, we obtain:

s=1+λ2K∣qw−ql∣1\boxed{s = 1 + \frac{\lambda}{2K}\\|q^w-q^l\\|_1}

Lscalar=−log⁡σ(s(rw−rl)−γg)\boxed{\mathcal{L}_{\mathrm{scalar}} = -\log \sigma\left(s(r^w-r^l)-\gamma g\right)}

We conduct the experiment using the same training dataset and optimization parameters as PhysGuide. The results are reported below:

MethodMechanicsOpticsThermalMaterialOverall
DPO46.161.042.158.753.1
Scalar ss47.362.543.459.654.3
Ours (PhysGuide)55.572.160.661.563.1

Analysis. The scalar-statistic reweighting baseline improves over vanilla DPO (54.3 vs. 53.1 overall) but remains substantially below PhysGuide (63.1 overall). This comparison demonstrates that aggregating principle-wise discrepancies into a single scalar is insufficient to achieve the gains of PhysGuide. The results support the benefit of retaining individual principle-wise discrepancies within the preference optimization objective, rather than collapsing them into a single scalar weighting factor.


W2. Principle-assignment distance is not validated as violation severity.

We clarify that our physics-violation divergence term measures differences in association across multiple physics principles, rather than directly measuring the severity of a single physical violation.

Each video is represented by a set of scalar association scores corresponding to different physical principles, such as collision dynamics, momentum conservation, and object interactions. Physically implausible videos may exhibit weaker associations with the relevant principles compared to physically plausible videos. Consequently, videos with more severe physical violations may have smaller association scores than those with milder violations. When comparing physically plausible and implausible videos, larger differences in their principle-association distributions can lead to a higher physics-violation divergence penalty, g(xw,xl)g(x^w,x^l).

Empirical validation. To investigate this relationship, we conduct an analysis on the PhysicsIQ test set. We use the PhysicsIQ evaluation error as a proxy for physical violation severity, where higher error values indicate greater deviations from expected physical behavior, while lower values indicate more physically plausible dynamics. For each evaluated video pair, we compute the corresponding physics-violation divergence penalty, g(xw,xl)g(x^w,x^l), and examine its relationship with the measured physics error. The results are summarized below:

Error valuesAvg. divergence penalty
Higher violation: [50, 100]0.84
Lower violation: [0, 50)0.35

Our analysis shows that higher physics evaluation errors are associated with larger divergence penalties (0.84 vs. 0.35). This suggests that the proposed divergence term captures principle-level discrepancies that are correlated with the severity of physical violations.

AuthorsOpenReview ↗

Rebuttal - 2

W3. Crucial implementation and gradient paths are unspecified.

Gradient Flow. We clarify that gradients flow exclusively through the differentiable policy rewards rθwr_\theta^w and rθlr_\theta^l, updating only the trainable LoRA parameters of the HunyuanVideo model, while the pretrained backbone remains frozen. The physics-principle representations qwq^w and qlq^l, and the divergence term g(xw,xl)g(x^w,x^l), do not receive gradients and are detached from the computational graph.

Consequently, for the scenario πθ=πref\pi_\theta = \pi_{\mathrm{ref}}, the model cannot directly minimize the preference loss by shrinking the distance gg between the preferred and rejected representations, which addresses the reviewer's concern about representation collapse. Since the LoRA-adapted video features evolve during training, we update the physics-principle prototypes using an exponential moving average (EMA) of the corresponding cluster embeddings.

Diffusion Likelihood Estimation. We clarify that we do not explicitly compute the exact video likelihoods. Instead, following diffusion/flow-matching DPO [1], we approximate the policy log-likelihood ratio using the difference in velocity-prediction errors between the trainable HunyuanVideo model and a frozen reference model.

For each preferred and rejected video, we encode the video into latent space, sample a random timestep and Gaussian noise, and construct noisy latents using the flow-matching schedule. Both the trainable policy and the reference model predict the velocity from the same noisy latents.

We estimate the implicit policy reward as:

r^θ(x,c)=\hat{r}_{\theta}(x,c) =

=βeff[∥vref(zt,t,c)−ut∥22−∥vθ(zt,t,c)−ut∥22]=\beta_{\mathrm{eff}} \left[ \lVert v_{\mathrm{ref}}(z_t,t,c) - u_t \rVert_2^2 - \lVert v_{\theta}(z_t,t,c) - u_t \rVert_2^2 \right]

where ut=ϵ−z0u_t=\epsilon-z_0 is the target flow velocity and βeff\beta_{\mathrm{eff}} includes the loss scaling.

This provides a differentiable surrogate for the policy reward, with gradients propagated through the trainable model while the reference model remains frozen. The estimated rewards for the preferred and rejected videos are then used in our principle-guided preference objective (Eq. 8).

EMA Updates. All codes live in the embedding space of an encoder fϕf_\phi that maps a video xx to a feature vector in Rd\mathbb{R}^d. Instead of keeping this encoder fixed, we update its parameters with an exponential moving average (EMA) of the corresponding encoder parameters in the generator:

ϕt=αϕt−1+(1−α)θtenc,\phi_t = \alpha \phi_{t-1} + (1-\alpha)\theta_t^{\mathrm{enc}},

where θtenc\theta_t^{\mathrm{enc}} denotes the visual encoder parameters of the generator at DPO step tt, and α∈(0,1)\alpha \in (0,1) is the EMA decay coefficient.

Video and Text Encoder. We clarify that the video encoder fϕf_\phi and text encoder are part of the pretrained HunyuanVideo model used for preference optimization. The pretrained backbone weights are frozen, while the LoRA parameters are trainable. Video features are extracted from the HunyuanVideo diffusion transformer (DiT) hidden states.

[1]. Wallace, Bram, et al. "Diffusion model alignment using direct preference optimization." 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024.


W4. Advanced-baseline comparisons are not demonstrably matched.

We clarify that our standard DPO baseline in Table 2 is trained using the same 200 winner and 600 loser videos, base HunyuanVideo model, and optimization settings as PhysGuide, ensuring that the improvements can be attributed to the proposed physics-guided objective.

For a fair comparison with advanced baselines, we fine-tuned PhyGDPO using the same 200 winner and 600 loser videos and matched optimization settings as PhysGuide. We could not reproduce PhysHPO under the same settings because its implementation code is not publicly available.

MethodMechanicsOpticsThermalMaterialOverall
PhyGDPO55.060.058.047.055.0
PhyGDPO w/ PhysicsIQ dataset53.157.356.145.853.1
Ours (PhysGuide)55.572.160.661.563.1

Analysis. PhysGuide continues to outperform PhyGDPO even when both methods are trained on the same preference dataset with matched optimization settings, achieving an overall score of 63.1 compared to 53.1. This controlled comparison suggests that the improvements are not simply due to differences in training data or optimization configurations, but support the effectiveness of our proposed physics-principle-guided objective.

AuthorsOpenReview ↗

Rebuttal - 3

W5. The ablations do not fully establish the claimed complementarity.

We conduct two additional ablation experiments: (1) a divergence-only variant without principle-wise scaling, and (2) a No EMA variant without exponential moving average updates. The results are reported below:

MethodMechanicsOpticsThermalMaterialOverall
Divergence-only51.266.353.359.258.3
No EMA55.171.860.361.262.8
Ours (PhysGuide)55.572.160.661.563.1

Analysis. The additional ablations show that both variants perform worse than the full PhysGuide model. Removing principle-wise scaling reduces the overall score from 63.1 to 58.3, highlighting the importance of incorporating individual physics-principle discrepancies into preference optimization. Removing EMA updates results in a smaller performance drop from 63.1 to 62.8, suggesting that EMA updates provide a modest improvement in performance.

W6. Evaluation reliability is insufficiently documented.

PhyGenBench evaluates generated videos using two metrics: Semantic Alignment (SA) and Physical Commonsense Alignment (PCA). SA measures whether the generated video contains the objects and actions described in the text prompt. PCA evaluates physical correctness across three aspects: (1) whether the expected physical phenomena occur, (2) whether physical events happen in the correct order, and (3) whether the overall motion follows realistic physical behavior. These aspects are evaluated using vision-language models, including VQAScore, GPT-4o, LLaVA-Interleave, and InternVideo2. The three PCA scores are converted to a 0–3 scale, averaged, and rounded down to obtain the final physical commonsense score. Higher scores indicate better physical correctness.

VideoPhy evaluates generated videos using two metrics: Semantic Adherence (SA) and Physical Commonsense (PC). A trained video-language model assesses whether the video follows the text prompt and exhibits physically plausible behavior. Scores are computed from normalized probabilities of “Yes” and “No” predictions, with higher scores indicating better adherence and plausibility.

For the evaluator checkpoints, we report the results directly from the PhyGDPO and PhysHPO papers, using the scores obtained from their respective model checkpoints and evaluation settings. Our training dataset consists of 200 samples from PhysicsIQ, while evaluation is conducted on 160 prompts from PhyGenBench and 688 prompts from VideoPhy.

We thank the reviewer for highlighting the need for clearer evaluation details and uncertainty estimates. We have added details about the benchmark metrics, evaluator checkpoints, and evaluation splits. Our human study includes 10 participants, each evaluating 20 videos, resulting in 300 evaluations. We report that PhysGuide was preferred in 288/300 (96.0%) comparisons, with a Wilson 95% confidence interval of [93.14%, 97.70%].

Comments on the paper

1 comment
AuthorsOpenReview ↗

Thanks and request for discussion

Dear reviewers,

We thank you once again for your feedback on our work. With the rebuttal/discussion period nearing a close, we would like to hear from reviewers on whether our rebuttal addressed your concerns. We are eager to clarify any remaining questions, help resolve any remaining concerns, and more importantly, look for ways to improve the paper overall.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.