acceptodds
Under review as a conference paper at ICLR 2027

DuoMatching: Joint–Marginal Distribution Matching for Few-Step Video Generation

45 views

Abstract

Streaming video generation has benefited from distribution matching distillation (DMD), which matches the joint distribution of video frames to a video teacher’s approximation of the real video distribution. Although this joint matching mitigates drift during autoregressive rollouts, limitations remain in visual quality and semantic alignment. To address these limitations, we propose DuoMatching, a distribution matching framework that approximates the real video distribution through a unified joint-marginal formulation. On top of existing joint matching formulations, the additional marginal matching objective provides dedicated frame-level supervision from an image generator, transferring complementary visual and semantic priors from it. To apply this frame-level supervision in video generation, we introduce LatentBridge to resolve the latent representation mismatch between the video student and the image teacher. Latent Variation Sampling further distributes such frame-level supervision across distinct temporal segments, reducing redundancy. Experiments demonstrate that DuoMatching improves visual quality, composition, and semantic alignment while largely preserving motion dynamics. Human evaluations show overall preference rates above 80% against all evaluated baselines.

Primary area:
generative models
Keywords:
Video Generation, Distribution Matching

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_NVGuOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper adds a frame-level marginal DMD term from an image teacher (Qwen-Image) on top of joint DMD for few-step video distillation. LatentBridge maps video latents to image latents, and LVS chooses which frames to supervise. It improves visual quality and semantic scores on both causal and bidirectional students.

Critical strengths

The idea is simple and gives consistent gains across all NFE settings at no inference cost. The LatentBridge ablation (Tab. 4) clearly shows that applying marginal DMD directly to video latents collapses motion.

Critical weaknesses

With Wan-14B as the image teacher the gain is small (80.67→81.43), so most of the improvement comes from the stronger Qwen-Image prior, and the closest prior image-teacher distillation methods (MCM, Zhu et al. 2024) are not compared. The theory assumes the key conclusion (that the image teacher's marginal is closer to real frames), and temporal quality is checked only with coarse VBench metrics.

Questions for authors

How does the method compare with Zhu et al. 2024 and with a frame-level reward baseline using the same image prior? How is "Total" computed, and what are the official VBench temporal-consistency scores?

1 comment

We really appreciate the reviewer for the detailed and constructive feedback. We are pleased that you recognize the simplicity and usefulness of our approach and its consistent gains without additional inference cost. Below, we address your concerns point by point.


Question 1: How does the method compare with Zhu et al. 2024?

Response: Thank you for this suggestion. We adapt MCM and AVDM² to Wan for comparison. For AVDM², we additionally provide our LatentBridge to enable frame-level score matching with the same Qwen-Image teacher.

Table 1. VBench Total comparison on Wan under matched generation modes and 4-step budgets.

MethodVBench Total ↑
MCM (Wan adaptation)76.89
AVDM² (Wan adaptation with LatentBridge)78.44
DuoMatching83.51

Question 2: How is "Total" computed, and what are the official VBench temporal-consistency scores?

Response: We follow the official VBench aggregation:

Total = 100 × (0.8 × Quality + 0.2 × Semantic)

Quality = (Subject Consistency + Background Consistency + Temporal Flickering + Motion Smoothness + Aesthetic Quality + Imaging Quality + 0.5 × Dynamic Degree) / 6.5

We additionally report the official VBench temporal-consistency metrics below.

Table 2. VBench temporal-consistency scores.

MethodNFESubject Consistency ↑Background Consistency ↑Temporal Flickering ↑
Causal Forcing++294.2993.2194.86
DuoMatching295.0695.5896.11
Causal Forcing++498.1296.3698.27
DuoMatching498.6697.2297.11

Concern 3: With Wan-14B as the image teacher the gain is small (80.67→81.43), so most of the improvement comes from the stronger Qwen-Image prior. How does the method compare with a frame-level reward baseline using the same image prior?

Response: Thank you for raising this point. We would like to clarify that a 0.76-point improvement on VBench is non-trivial. For context, Causal Forcing reports a 0.30-point improvement over Self Forcing in VBench Total (83.74 → 84.04).

We agree that Qwen-Image provides a stronger prior, while effectively transferring this prior is a key contribution of our method. To examine the role of the transfer mechanism, we also evaluate a DMD2-style alternative that combines joint video DMD with a frame-level adversarial loss, using 10,000 Qwen-Image-generated images to train the discriminator. Its lower performance compared with DuoMatching supports the effectiveness of our approach in transferring the image prior.

Table 3. Comparison of transfer mechanisms using the same Qwen-Image prior source under matched 2-step settings.

MethodVBench Total ↑
Causal Forcing++80.67
Joint DMD + frame-level GAN (DMD2-style)80.48
DuoMatching83.53

Concern 4: The theory assumes the key conclusion (that the image teacher's marginal is closer to real frames).

Response: Thank you for raising this concern. To empirically examine its marginal-quality assumption, we use the same video prompts to generate 1,000 Qwen-Image images and Wan2.1-14B videos, then compare the images and sampled video frames against 1,000 randomly sampled OpenVidHD reference frames. Qwen-Image achieves a lower FID (61.79 vs. 66.85), providing empirical evidence consistent with this assumption.

Reviewer_EwaMOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses the limited frame-level quality of few-step video generators distilled with distribution matching distillation (DMD), especially autoregressive/streaming generators. . Existing video DMD methods primarily match the joint distribution of video frames using a pretrained video teacher. The authors argue that this can preserve temporal consistency but leave deficiencies in individual-frame visual quality and semantic alignment. To address this the authors introduce an image teacher to improve individual-frame appearance, fine-grained details, and prompt alignment. This setup is made practical by 2 components - First, LatentBridge - a small model that maps from the existing latent space to the latent space of the image generation models and secondly Latent Variation Sampling (LVS), a way to select frames from different temporal segments to send to the judge model. The method is evaluated on causal and bidirectional video generation. The reported results show improvements in visual quality and semantic alignment, with generally comparable temporal quality.

Critical strengths
  1. The thought of improving prompt adherance and visual quality while maintaining the existing temporal consistency in Joint DMD based Models is convincing and an important challenge. Use of LatentBridge is interesting and mapping temporally compressed video latents into image-compatible latents avoids repeatedly decoding and re-encoding video frames. The reported memory overhead is small compared with the direct baseline. Latent Variation Sampling is a simple, cheap, and effective way to allocate frame-level supervision while providing modest gains from uniform sampling.
  2. The method achieves consistent gains across settings and improvements appear in 1-, 2-, and 4-step causal generation and in bidirectional generation. The paper investigates the image teacher, LatentBridge, the number of sampled frames, and the marginal loss weight. These experiments help understand the contributions of individual components. The ablation in Table 3 is esspecially helpful since it shows a monotone trend between image-teacher strength (HPSv3) and student quality.
Critical weaknesses
  1. The conceptual novelty of the joint-marginal formulation is not sufficiently established relative to prior image-teacher-based distribution matching approaches, particularly AVDM2. Moreover there are gap between the theory and the actual algorithm. Appendix C assumes "matched marginal capacity," i.e., that the image teacher PiP_i is the optimal fit within the same feasible family F={MP:P∈P}\mathcal{F}=\{MP : P\in\mathcal{P}\} as the video model's frame marginals. In practice PiP_i is an independently trained still-image model (Qwen-Image) with different data and architecture. The key premise DKL(m∗∥Pi)≤DKL(m∗∥mv)D_{\mathrm{KL}}(m^{\ast} \Vert P_i) \le D_{\mathrm{KL}}(m^{\ast} \Vert m_v), on which Appendix D also depends, is therefore assumed rather than shown. Moreover Appendix F assumes a fixed scalar frame-level loss ℓ\ell, whereas the marginal DMD update depends on a fake score estimator that evolves during training.
  2. Empirically, there is no distribution-level metric against real videos, such as FVD or frame-level FID/KID against real video frames. The reported metrics (VBench Aesthetic/Imaging Quality, human preference for visual quality) plausibly reward still-image aesthetics. Still images and video frames differ systematically, in terms of motion blur and framing, lighting. The qualitative results in the paper as well as in the supplementary, also show noticeable stylistic shifts relative to the baseline. So the gains are equally consistent with the student moving toward the image teacher's aesthetic rather than toward P∗P^*.
Questions for authors
  1. Can you report FVD, or frame-level FID/KID against real video frames, for DuoMatching versus Causal Forcing++ and CausVid? The paper's central claim is a closer approximation to P∗P^{\ast}. Evidence that DuoMatching moves closer to real video statistics, and not only toward the image teacher's aesthetic, would substantially strengthen my assessment of Weakness 2.
  2. How do the results in Appendices C, D, and F apply to the implemented method, given three differences? (a) Qwen-Image is not the optimal fit within the video model's marginal family, so the premise DKL(m∗∥Pi)≤DKL(m∗∥mv)D_{\mathrm{KL}}(m^{\ast} \Vert P_i) \le D_{\mathrm{KL}}(m^{\ast} \Vert m_v) is an assumption that is not explored. (b) LVS makes frame selection depend on the generated sample, whereas Appendix D assumes a fixed linear operator MM. (c) The marginal DMD update depends on an evolving fake score estimator, whereas Appendix F assumes a fixed loss ℓ\ell.
  3. In the 4-step setting, DuoMatching improves the overall score but reduces Dynamic Degree from 80.56 to 76.67 compared with Causal Forcing++. Human evaluation also shows approximately equal preference for temporal/motion quality (49.57%). Could the authors explain this degradation and provide statistical significance or confidence intervals for the temporal evaluation? This would help determine whether the method improves visual quality at the expense of motion characteristics.
  4. Please report wall-clock training time and total compute for both DuoMatching and the baseline under identical settings, together with the size and parameterization of the image fake score estimator. Given the size of Qwen-Image, how do the teacher and fake estimator fit within the $\sim$3 GB peak-memory increase reported in Table 4 (sharding, offloading, LoRA, separate devices) (Table 4 reports peak memory of 56.49 GB without and 59.41 GB with marginal DMD.)? Please also give the values of RR and TwarmT_{\mathrm{warm}} in Algorithm 1.
  5. It would be better if the authors provide results over multiple seeds, especially where margins are small.
  6. How is the marginal DMD objective (Eq. 6) differs from AVDM2's image-teacher score distribution matching, aside from the latent mapping. It is important to understand which of the DuoMatching's contributions do you consider conceptually new rather than enabling engineering?
  7. Did you try Decode–Encode restricted to only the KK sampled frames, with gradient checkpointing? If that fits in memory, how does it compare with LatentBridge in quality and cost?
5 comments

We really appreciate the reviewer for the detailed and constructive feedback. We are pleased that you find our motivation convincing and our experimental analysis informative.

We believe our response has addressed most of your concerns and we respectfully ask if you're comfortable to re-consider the rating in light of our efforts and dedication. Below, we provide detailed responses to all the comments:


Question 1: Can you report FVD, or frame-level FID/KID against real video frames, for DuoMatching versus Causal Forcing++ and CausVid? The paper's central claim is a closer approximation to P∗P^*. Evidence that DuoMatching moves closer to real video statistics, and not only toward the image teacher's aesthetic, would substantially strengthen my assessment of Weakness 2.

Response: We sample 512 high quality videos from OpenVidHD with high motion, and calculate the FVD with our generated 512 videos with the same prompt.

Table 1. FVD comparison across generation modes and NFE settings.

MethodFVD ↓
Bidirectional · 4 NFE
CausVid1155.51
DuoMatching864.97
Causal frame-wise · 1 NFE
Causal Forcing++1193.98
One-Forcing1301.14
DuoMatching934.56
Causal frame-wise · 2 NFE
Causal Forcing++977.63
DuoMatching720.44
Causal chunk-wise · 4 NFE
CausVid1194.67
Self Forcing1078.06
Reward Forcing1280.68
Causal Forcing++1065.55
DuoMatching767.03

The lower FVD scores support improved alignment with real videos.


Question 2: How do the results in Appendices C, D, and F apply to the implemented method, given three differences?

(a) Qwen-Image is not the optimal fit within the video model's marginal family, so the premise DKL(m∗∥Pi)≤DKL(m∗∥mv)D_{\mathrm{KL}}(m^* \| P_i) \leq D_{\mathrm{KL}}(m^* \| m_v) is an assumption that is not explored.

(b) LVS makes frame selection depend on the generated sample, whereas Appendix D assumes a fixed linear operator MM.

(c) The marginal DMD update depends on an evolving fake score estimator, whereas Appendix F assumes a fixed loss ℓ\ell.

Response: (a) Thank you for highlighting this distinction. Appendix C provides an idealized comparison to motivate marginal supervision, while Appendix D does not require the image teacher to be optimal. To empirically examine its marginal-quality assumption, we use the same video prompts to generate 1,000 Qwen-Image images and Wan2.1-14B videos, then compare the images and sampled video frames against 1,000 randomly sampled OpenVidHD reference frames. Qwen-Image achieves a lower FID (61.79 vs. 66.85), providing empirical evidence consistent with this assumption.

(b) We would like to clarify that Appendix D already represents MM as a sampling kernel, (MQ)(y)=∫M(y∣x)Q(x) dx(MQ)(y)=\int M(y\mid x)Q(x)\,dx. “Fixed” refers to the conditional sampling rule, not identical frame indices across videos. LVS may depend on each sample while still inducing a linear mapping on distributions. Therefore, this sample dependence does not invalidate the proof, provided that the same LVS rule defines m∗m^* and mvm_v, and the stated assumptions hold for these induced marginals. We will make this interpretation explicit in Appendix D.

(c) We would like to clarify that the fixed-loss assumption concerns a before–after comparison of a generator update. For DMD, Appendix F can be applied under the stated conditions to a local surrogate with the score-derived supervision signal held fixed during this comparison. Subsequent fake-estimator updates define a new surrogate. We will clarify this interpretation in Appendix F.


Question 3: In the 4-step setting, DuoMatching improves the overall score but reduces Dynamic Degree from 80.56 to 76.67 compared with Causal Forcing++. Human evaluation also shows approximately equal preference for temporal/motion quality (49.57%). Could the authors explain this degradation and provide statistical significance or confidence intervals for the temporal evaluation? This would help determine whether the method improves visual quality at the expense of motion characteristics.

Response: We acknowledge the decrease in Dynamic Degree in the 4-step setting. This automatic metric measures motion degree rather than overall perceived motion quality. In the corresponding human evaluation, DuoMatching receives a temporal-quality preference rate of 49.57% (95% CI: [46.87%, 53.55%]), indicating no statistically clear preference for either method.

Question 4: Please report wall-clock training time and total compute for both DuoMatching and the baseline under identical settings, together with the size and parameterization of the image fake score estimator. Given the size of Qwen-Image, how do the teacher and fake estimator fit within the ~3 GB peak-memory increase reported in Table 4 (sharding, offloading, LoRA, separate devices) (Table 4 reports peak memory of 56.49 GB without and 59.41 GB with marginal DMD.)? Please also give the values of RR and TwarmT_{\mathrm{warm}} in Algorithm 1.

Response: Thank you for requesting these implementation details. On the same eight-GPU setup, reaching 1,000 training iterations takes 50 h 33 min for Causal Forcing++ and 54 h 22 min for DuoMatching, corresponding to 404.4 and 434.9 GPU-hours.

All models are sharded with FSDP across the same eight GPUs, without CPU offloading or additional devices. For the Qwen-Image configuration, we use Wan2.1-1.3B as the image fake score estimator, leveraging their shared latent space, rather than maintaining a second Qwen-Image model. This engineering optimization provides comparable performance while reducing memory consumption. Together with FSDP, it helps explain the modest peak-memory increase reported in Table 4. For other image teachers, LoRA-based adaptation is another possible way to reduce the fake estimator’s training memory.

We set R=5R=5 and Twarm=50T_{\mathrm{warm}}=50 in Algorithm 1. We will explicitly include these implementation details in the revised manuscript.


Question 5: It would be better if the authors provide results over multiple seeds, especially where margins are small.

Response: Thank you for this suggestion. Our original VBench evaluation already uses five inference seeds per prompt. We now explicitly report the variability across these seeds and additionally evaluate DuoMatching across three independent training seeds.

Table 2. VBench Total scores across the five inference seeds used in the original evaluation (mean ± std).

Method2-step VBench Total ↑4-step VBench Total ↑
Causal Forcing++80.67 ± 0.0682.79 ± 0.10
DuoMatching83.53 ± 0.0783.51 ± 0.08

Table 3. VBench Total scores across three independent training seeds for DuoMatching (mean ± std), with each checkpoint’s score first aggregated over the same five inference seeds. Causal Forcing++ results are from a single reference checkpoint.

Method2-step VBench Total ↑4-step VBench Total ↑
Causal Forcing++80.6782.79
DuoMatching83.52 ± 0.1183.47 ± 0.05

DuoMatching achieves higher mean scores in both evaluations, supporting the robustness of the reported gains.

Question 6: How is the marginal DMD objective (Eq. 6) differs from AVDM²'s image-teacher score distribution matching, aside from the latent mapping. It is important to understand which of the DuoMatching's contributions do you consider conceptually new rather than enabling engineering?

Response: Thank you for encouraging us to clarify this distinction.

While both methods use image-teacher and fake-estimator scores for frame-level supervision, our conceptual contribution lies in formulating and analyzing the complementarity between joint and marginal DMD.

AVDM² compares marginal and joint score matching as alternative auxiliary objectives within a video-GAN framework. Its adversarial supervision is distinct from joint video DMD, and its reported advantage of marginal over joint score matching does not hold in modern video models.

We instead investigate why explicit marginal supervision remains beneficial when joint video DMD already implicitly constrains the marginals. Joint-marginal distribution matching is our core training objective. Our theoretical analysis examines this complementarity under approximate teacher distributions, with a conceptual illustration provided in Figure 2. Our controlled ablation further supports it: under the same video adversarial supervision, joint–marginal DMD achieves 83.57, outperforming joint-only DMD (82.80) and marginal-only DMD (78.44).

Table 4. Ablation of joint and marginal DMD on Wan under the same video-GAN training setup. The Frame-level DMD is deployed through LatentBridge.

MethodJoint DMDMarginal DMDVideo GANVBench Total ↑
DMD2 applied to video✓—✓82.80
AVDM² (Frame-level DMD + GAN)—✓✓78.44
DuoMatching + GAN✓✓✓83.57

Question 7: Did you try Decode–Encode restricted to only the KK sampled frames, with gradient checkpointing? If that fits in memory, how does it compare with LatentBridge in quality and cost?

Response: Thank you for this suggestion. After moving to 96 GB GPUs, we fit Decode–Encode with 4 sampled frames and gradient checkpointing in memory without OOM. At 120 training iterations, LatentBridge and Decode–Encode achieve similar VBench Total scores (79.56 vs. 79.57). Under identical training settings, LatentBridge is approximately 9.36× faster per iteration and reduces peak allocated GPU memory by 32.4%.

Table 5. Quality and cost comparison under identical training settings. VBench Total is evaluated after 120 training iterations. Runtime (mean ± std) and peak allocated GPU memory per GPU are measured over ten consecutive iterations after 20 warm-up iterations.

MethodVBench Total at 120 iterations ↑Training time per iteration (s) ↓Peak GPU memory (GB/GPU) ↓
LatentBridge79.56195.72 ± 3.2459.48
Decode–Encode (4 frames, gradient checkpointing)79.571832.70 ± 10.7487.97
Reviewer_w9yZOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

Distribution Matching Distillation addresses the computational bottleneck of video generation models by distilling multistep diffusion into few-step generators. The authors identify a gap in these methods: preserving frame-level visual quality and prompt adherence.

The authors propose DuoMatching, which adds a marginal distribution constraint in training. A frozen pretrained image generator provides frame-level DMD supervision on sampled frames, while the video teacher keeps temporal coherence. The two terms are combined as L = L_joint + ω*L_marginal.

Contributions:

  1. Joint-marginal formulation
  2. LatentBridge: A lightweight network maps temporally compressed video latents to per-frame latents in the image teacher's latent space.
  3. Latent Variation Sampling: The latent sequence is split at the largest changes between adjacent slices, and one slice is sampled per segment (K = 4). This spreads frame-level supervision across the video.

The authors test both causal students at 1, 2, and 4 steps and a bidirectional student, all built on Wan2.1-1.3B. On VBench, DuoMatching beats CausVid, Self Forcing, Reward Forcing, One-Forcing, and Causal Forcing++ on visual quality and prompt alignment. In a side-by-side study with 23 raters, people preferred DuoMatching over 80% of the time against every baseline. Ablations show that stronger image teachers give better students, that LatentBridge keeps motion intact where direct latent matching does not, and that LVS frame selection and a marginal weight of ω = 0.4 give the best balance between quality and motion.

Critical strengths

The authors identify a real gap: few-step DMD video generators lose per-frame fidelity and prompt adherence, as Figure 1 clearly shows.

LatentBridge makes image-teacher supervision practical for video generation models; Table 4 shows it preserves motion and improves quality.

The image prior visibly improves frame quality: The marginal term improves causal students at 1, 2, and 4 steps as well as a bidirectional student, with no added inference cost. Student quality rises steadily as the image teacher improves, which suggests image priors actually transfer to the video student.

Critical weaknesses

Evaluation: The VBench improvements over the strongest baseline are under one point, and no variance across seeds is reported. The evaluation also leans heavily on VBench and frame-level quality scores, with no video-level metrics such as FVD. The human study appears to use only about 10 prompts per baseline, and if these are the same as the supplementary prompts, most of them describe slow, calm scenes. It would help to know how these prompts were chosen. As it stands, the claim that motion is preserved isn't strongly supported; testing on a larger set of prompts with more varied motion would make it much more convincing.

Semantic-alignment gains also appear limited to within-frame attributes; scene-level and multi-step prompt errors persist in several supplementary examples.

Table 3 shows that image priors transfer, but because teacher quality increases together with model size and no same-teacher control is included, it does not show that the joint-marginal formulation itself, rather than added teacher knowledge, is responsible for the gains.

Questions for authors

The concerns are mentioned in the weakness section.

3 comments

We really appreciate the reviewer for the detailed and constructive feedback. We are pleased to see that you recognize the importance of the problem and the consistent improvements achieved by our method.

We believe our response has addressed most of your concerns and we respectfully ask if you're comfortable to re-consider the rating in light of our efforts and dedication. Below, we provide detailed responses to all the comments:


Concern 1: The VBench improvements over the strongest baseline are under one point, and no variance across seeds is reported.

Response:

Our original VBench evaluation already uses five inference seeds per prompt, following VBench official protocols. We now explicitly report the variability across these seeds and additionally evaluate DuoMatching across three independent training seeds.

Table 1. VBench Total scores across the five inference seeds used in the original evaluation (mean ± std).

Method2-step VBench Total ↑4-step VBench Total ↑
Causal Forcing++80.67 ± 0.0682.79 ± 0.10
DuoMatching83.53 ± 0.0783.51 ± 0.08

Table 2. VBench Total scores across three independent training seeds for DuoMatching (mean ± std), with each checkpoint’s score first aggregated over the same five inference seeds. Causal Forcing++ results are from a single reference checkpoint.

Method2-step VBench Total ↑4-step VBench Total ↑
Causal Forcing++80.6782.79
DuoMatching83.52 ± 0.1183.47 ± 0.05

The improvements remain consistent across both training and inference seeds, supporting the effectiveness of our method.


Concern 2: The evaluation leans heavily on VBench and frame-level quality scores, with no video-level metrics such as FVD.

Response: We sample 512 high quality videos from OpenVidHD with high motion, and calculate the FVD with our generated 512 videos with the same prompt.

Table 3. FVD comparison across generation modes and NFE settings.

MethodFVD ↓
Bidirectional · 4 NFE
CausVid1155.51
DuoMatching864.97
Causal frame-wise · 1 NFE
Causal Forcing++1193.98
One-Forcing1301.14
DuoMatching934.56
Causal frame-wise · 2 NFE
Causal Forcing++977.63
DuoMatching720.44
Causal chunk-wise · 4 NFE
CausVid1194.67
Self Forcing1078.06
Reward Forcing1280.68
Causal Forcing++1065.55
DuoMatching767.03

The lower FVD scores indicate that videos generated by DuoMatching more closely match the real-video distribution.


Concern 3: The human study appears to use only about 10 prompts per baseline, and if these are the same as the supplementary prompts, most of them describe slow, calm scenes. It would help to know how these prompts were chosen. As it stands, the claim that motion is preserved isn't strongly supported; testing on a larger set of prompts with more varied motion would make it much more convincing.

Response: To ensure a fair comparison, we randomly sample 10 demo prompts from Causal Forcing for all methods in the original user study. These differ from the prompts shown in our supplementary material.

To further evaluate motion quality, we curate 50 additional prompts describing more dynamic scenes with various motions. Under identical generation settings, Causal Forcing++(2-step)videos achieve a higher VBench Dynamic Degree on this set than on the original prompt set (90 → 94.8). We recruit 20 raters to evaluate paired videos from the two methods on these 50 prompts. We additionally evaluate motion magnitude to assess whether the improvement in visual quality comes at the expense of reduced motion.

Table 4. Human preference for 2-step DuoMatching over 2-step Causal Forcing++ on 50 additional dynamic prompts. Results are reported as preference rates with 95% confidence intervals.

ComparisonVisual qualitySemantic alignmentTemporal qualityMotion magnitudeOverall
DuoMatching vs. Causal Forcing++ (2-step)80.8% [76.0%, 85.4%]78.9% [73.4%, 84.0%]50.1% [47.4%, 56.2%]49.8% [47.2%, 51.9%]80.1% [75.0%, 84.9%]

The additional human evaluation shows no clear preference between the methods in temporal quality or motion magnitude, while favoring DuoMatching in visual quality and semantic alignment.

Concern 4: Semantic-alignment gains also appear limited to within-frame attributes; scene-level and multi-step prompt errors persist in several supplementary examples.

Response: Thank you for pointing out this limitation. Our marginal matching objective substantially improves within-frame semantic alignment, while temporal semantics, such as multi-step event ordering, remain a limitation; we will clarify this scope in the revised manuscript.


Concern 5: Table 3 shows that image priors transfer, but because teacher quality increases together with model size and no same-teacher control is included, it does not show that the joint-marginal formulation itself, rather than added teacher knowledge, is responsible for the gains.

Response: We would like to clarify that Table 3 includes a same-teacher control using Wan2.1-14B as the image teacher. Adding marginal matching improves the Total score from 80.67 to 81.43 without introducing additional teacher knowledge, supporting its effectiveness.

Dear Reviewer w9yZ,

Thank you again for your thoughtful review. We’ve updated the PDF with the additional experiments and clarifications discussed in our response. We’d welcome your feedback and, if you feel the revisions address your concerns, we would be grateful if you would consider revisiting your rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.