acceptodds
Withdrawn from ICLR 2027

Stabilizing Long-Horizon Streaming DMD with Drift-Controlled Adaptive TTUR

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Distribution matching distillation (DMD) relies on an online fake-score model that tracks an evolving generator distribution; tracking errors can perturb generator updates. Fixed two-timescale update rules (TTUR) use constant critic-to-generator ratios that can be poorly suited to the varying tracking burden in long-horizon streaming DMD. We introduce Drift-Gated TTUR, which adapts only optimizer timing. A periodic fixed-paired probe holds prompts, latents, diffusion timesteps, and forward noise fixed while measuring temporal changes in the on-policy fake- score field as a one-sided indicator of elevated change in the coupled generator– critic dynamics. After a short calibration, exceeding a frozen threshold triggers a skipped generator update and additional critic recovery; otherwise, the original schedule is retained. DMD objectives and inference remain unchanged, with no gradient through the probe or controller. On the official 946-prompt VBench evaluation in LongLive, our method achieves Total and Quality scores of 81.05 and 83.59, compared with 80.16 and 82.93 for the uniform-schedule control. The Total gain is +0.89 with a paired 95% confidence interval of [+0.60,+1.17] across seven matched inference seeds. A random intervention schedule scores 79.58, while removing critic recovery from the uniform schedule yields 80.03. These controls indicate that neither additional critic optimization nor schedule irregularity alone explains the gain, supporting drift-conditioned optimizer timing as an effective control variable for long-horizon streaming DMD.

Primary area:
generative models
Keywords:
Distribution Matching Distillation, Streaming Video Generation, Adaptive TTUR, Training Stability, Score Distillation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_vxQ3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Drift-Gated TTUR, an adaptive training update scheduler for long-horizon streaming video generation with DMD. The proposed method measures changes in the prediction of the fake model using a small fixed probe set. When the measured drift exceed a threshold, the method skips the generator update and execute additional fake model update. The experiments are conducted using LongLive on VBench.

Critical strengths
  1. This paper proposes a simple yet effective method to improve the training stablilty of DMD, requiring no changes to the original DMD objective and inference pipeline.

  2. The authors compare the proposed method with uniform, random, and skip-only schedules.

Critical weaknesses
  1. Adaptive generator-discriminator update scheduling has already been explored in previous work [1]. The proposed Drift-Gated TTUR follows a similar idea, with the main difference being the application on DMD.

  2. It is unclear whether the proposed drift measurement is a reliable indicator of critic tracking difficulty. The measured drift reflects changes in both the generator and the fake-score model, rather than directly measuring the critic's estimation error. The paper also uses only two fixed probe samples (m=2), without examining how the number or selection of probe samples affects the results.

  3. The results on VBench is promising, however, Figure 4 shows that the drift remains highly variable throughout training, with noticeably larger fluctuations after the initial stage. However, the paper only shows the drift curve for the proposed method. Without comparing the drift dynamics against Fixed TTUR or other scheduling strategies, it is difficult to determine whether the proposed method actually reduces long-term drift or mainly changes the timing of optimizer updates.

  4. Besides, all controlled experiments are conducted using LongLive. Although the authors include comparisons with several official checkpoints, they do not evaluate previous works, like Causal Forcing [2], Causal Forcing++ [3] and etc, or demonstrate that the proposed controller can improve other DMD-based video generation frameworks.

[1] Accelerated WGAN update strategy with loss change rate balancing

[2] Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation

[3] Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation

Questions for authors
  1. Adaptive generator-discriminator update strategies have already been studied in previous work, such as [1]. What is the main difference between Drift-Gated TTUR and these methods beyond the choice of scheduling signal?

  2. The proposed drift measurement uses only two fixed probe samples (m=2). How sensitive is the method to the number and selection of probe samples? Would using more samples or different probe sets lead to more reliable scheduling decisions?

  3. Figure 4 shows that drift remains relatively high and fluctuates throughout training. Does Drift-Gated TTUR actually reduce the overall drift compared with Fixed TTUR? Could the authors provide drift curves for both methods under the same training settings, along with statistics such as the average drift and frequency of high-drift events?

  4. Could the authors provide results from multiple independent training runs? The current confidence intervals only account for inference-time variation.

  5. The proposed method is only evaluated within the LongLive training framework. Could the authors explain whether the same strategy can be applied to other autoregressive video distillation methods, such as Causal Forcing?

  6. In the first example of Figure 5, the results from Drift-Gated TTUR appear somewhat oversaturated compared with other methods. Could the authors provide more qualitative examples, preferably randomly selected, to show whether this behavior is consistent across different prompts?

[1] Accelerated WGAN update strategy with loss change rate balancing

4 comments

W1/Q1: Distinction from prior adaptive GAN scheduling

We thank the reviewer for pointing out Ouyang et al. (2021). We agree that adapting generator–discriminator update frequencies is not new in itself, and have added an explicit citation and discussion in Related Work.

Our contribution differs in the optimization setting, monitoring signal, and intervention mechanism. Ouyang et al. adapt WGAN training using relative changes in generator and discriminator losses. In contrast, DMD relies on an online fake-score model that must track the evolving generator distribution, and its estimation error directly perturbs the practical generator update direction.

Motivated by this structure, we introduce a fixed-paired probe that measures temporal changes in the on-policy fake-score field while holding the probe variables fixed. The controller uses elevated drift to selectively skip generator updates and allocate critic recovery steps, without modifying the DMD objectives.

We therefore position our contribution not as the first adaptive two-timescale schedule, but as a DMD-specific drift-conditioned controller, supported by a local tracking analysis and controlled scheduling experiments.

W2/Q2: Interpretation and sensitivity of the drift probe

We agree that the measured drift is not a direct estimator of critic tracking error. As clarified in Section 3.2 and Appendix B.3 (Eq. (36)), the statistic captures temporal changes arising from both generator-induced input shifts and critic parameter updates. We therefore use it only as a one-sided intervention signal: elevated drift triggers recovery, whereas low drift is not interpreted as evidence of accurate critic tracking. To examine probe sensitivity, we conducted an additional ablation (Exp. 084) varying probe-bank composition and size, with κ = 3 and M = 40 and separate threshold calibration for each configuration.

Probe configurationVBench Total ↑Threshold τ
Default Bank A (m = 2)80.930.0927
Independent Bank B (m = 2)80.580.0693
Extended Bank A (m = 4)79.340.1105

Replacing the probe bank leads to a relatively modest reduction in Total (−0.35), while increasing the bank size to four produces a larger decrease (−1.59) and approximately 23% additional training overhead. These results indicate that scheduling is not dependent exclusively on the original probe samples, but is sensitive to probe-bank configuration. Increasing the number of probe samples does not necessarily improve scheduling performance under the tested settings. The ablation is reported in Appendix E.2, Table E.1.

W3/Q3: Drift dynamics relative to Fixed TTUR

We thank the reviewer for this suggestion. We added a comparison of Drift-Gated TTUR and Fixed 5:1 using the same fixed-paired probe, evaluated every five outer iterations. Fixed TTUR records drift passively, with the gate disabled. We analyze three independent training seeds (42, 43, and 45), with 359 post-calibration probe evaluations per arm and seed.

Training seedMean drift (ours / Fixed)High-drift rate (ours / Fixed)
420.1025 / 0.368945.1% / 99.7%
430.1079 / 0.299043.2% / 98.1%
450.1042 / 0.316743.7% / 99.4%

Across seeds, the ratio of mean drift under Fixed TTUR to that under Drift-Gated TTUR is approximately 3.1×. Mean high-drift event frequency decreases from approximately 99% to 44%. The temporal curves show increasingly elevated drift under Fixed TTUR and substantially lower drift under adaptive scheduling at later stages. We have expanded Figure 4 to include both schedules and added the remaining curves and three-seed statistics in Appendix F.

W4/Q4: Robustness across independent training runs

We added two independent training replications (three runs total). Uniform matches the adaptive run's generator/critic update counts separately for each run. Evaluation uses 946 VBench prompts and seven inference seeds (Appendix G).

Training runDrift-Gated TTURFixed 5:1UniformΔ vs. FixedΔ vs. Uniform
Run 1 (original)81.0580.1480.16+0.91+0.89
Run 2 (added)80.4579.9480.28+0.51+0.17
Run 3 (added)81.3580.4679.96+0.89+1.39
Mean ± SD across training runs80.95 ± 0.4680.18 ± 0.2680.13 ± 0.16+0.77 ± 0.23+0.82 ± 0.61

Our method outperforms both controls in all three runs, averaging +0.82 ± 0.61 over Uniform and +0.77 ± 0.23 over Fixed (mean ± SD across training runs).

W4/Q5. Applicability beyond LongLive

We thank the reviewer for raising this question. The motivation behind our controller is not specific to LongLive, but arises from the generator–fake-score coupling inherent in DMD. As the generator distribution evolves, the online fake-score model must continually adapt, making the required critic update budget potentially time-varying.

Our method operates at the optimizer-scheduling level and is largely orthogonal to the student architecture, rollout strategy, and preceding initialization stages. It only requires monitoring the online fake-score model through fixed-paired probes and adjusting generator/critic updates, without modifying the DMD objective or inference pipeline.

In particular, Causal Forcing differs from Self Forcing primarily in its initialization pipeline: it introduces an additional teacher-forcing stage to train an autoregressive teacher, followed by causal ODE initialization. The subsequent DMD training procedure follows Self Forcing, making our controller conceptually compatible with Causal Forcing and potentially complementary to its improved initialization, without requiring changes to that pipeline.

To further examine generality, we conducted an additional short-horizon T2V experiment (81 frames, 16 fps, without long-horizon rollout). Drift-Gated TTUR improves VBench Total from 79.64 to 80.90 over Fixed 5:1 and increases Dynamic Degree from 0.153 to 0.597 (Appendix C).

These results show that the scheduling benefit extends beyond long-horizon rollout within the LongLive-based training pipeline. Direct evaluation in Causal Forcing or Causal Forcing++ remains future work.

To further examine whether our improvements depend on long-horizon autoregressive rollout, we conducted an additional short-horizon T2V experiment (81 frames, 16 fps, without long-horizon rollout). Drift-Gated TTUR improves VBench Total from 79.64 to 80.90 over Fixed 5:1 and increases Dynamic Degree from 0.153 to 0.597 (Appendix C).

These results demonstrate that the scheduling benefit extends beyond long-horizon rollout within the LongLive-based DMD pipeline, supporting the broader applicability of the underlying mechanism. Direct evaluation on Causal Forcing and other DMD frameworks remains future work. We have clarified this distinction in Section 3.2.

Q6. Color saturation and qualitative coverage

We added six prompt groups (24 videos) in Appendix H (Figures 9–10), spanning objects, landmarks, and outdoor scenes. All four methods use matched temporal positions and original aspect ratios, without color correction. The examples include both neutral and vivid scenes, broadening direct inspection of color behavior across prompts. Section 5.6 now references these comparisons.

We hope these additions address the concerns and encourage reconsideration of the rating.

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts to address my concerns.

However, I still have some concerns about the reliability and novelty of the proposed scheduling strategy.

First, the new probe ablation shows noticeable sensitivity to the probe configuration. In particular, increasing the probe size from m=2 to m=4 reduces VBench Total from 80.93 to 79.34, despite recalibrating the threshold. This is a relatively large drop compared with the reported average improvement over Uniform Scheduling. Could the authors clarify how the probe configuration affects the frequency of recovery steps and the resulting generator/critic update counts?

Second, although the authors have clarified the difference from previous adaptive GAN scheduling methods, it is still unclear whether fake-score drift provides a meaningful advantage over simpler adaptive signals, such as critic loss changes or gradient statistics.

Overall, I appreciate the improvements in the rebuttal, particularly the additional training runs and drift comparisons. However, the sensitivity to probe configuration and the limited evidence for the distinct advantage of the proposed scheduling signal remain important concerns. I therefore plan to maintain my current rating.

1. Probe-bank sensitivity. We thank the reviewer for encouraging us to look more closely at the controller behavior. Motivated by this comment, we conducted an additional controlled experiment to better understand the interaction between probe-bank size and threshold calibration.

In the original m=4 run, calibration yielded a higher threshold than m=2 (0.1105 vs. 0.0927), accompanied by fewer recovery events (146 vs. 162), more generator updates (254 vs. 238), and fewer recovery critic updates (730 vs. 810). This helped us identify that applying the same calibration rule across probe configurations does not necessarily lead to comparable effective update schedules.

We therefore repeated the m=4 experiment with a threshold approximately matching the default m=2 setting, while keeping all other settings unchanged. VBench Total improved from 79.34 to 80.72, recovering approximately 87% of the original performance drop and narrowing the gap to m=2 (80.93) to just 0.21 points.

These findings suggest that drift-conditioned scheduling is not limited to the original m=2 probe configuration, provided that the threshold is appropriately calibrated. The substantial recovery in performance also suggests that the original degradation was largely associated with changes in the effective update schedule, rather than probe-bank size alone. We appreciate the reviewer bringing attention to this interaction, as it highlights an important consideration when applying the controller across different probe configurations.

We have included this additional control in Section 5.4 and Appendix E.2. Among the tested settings, m=2 remains our preferred choice due to its lower computational overhead, as m=4 increases training time by approximately 23% without a clear improvement in generation quality.

2. Choice of scheduling signal. We thank the reviewer for raising this point. Indeed, our initial investigation began with simpler signals derived from the critic's prediction loss, before we developed the fake-score drift probe. However, our preliminary experiments suggested that these measurements did not reliably reflect errors in the practical DMD generator update.

Specifically, in controlled frozen-generator experiments, the critic's prediction error on held-out samples showed weak, near-zero, or negative correlations with the DMD-gradient deviation from a well-trained reference. In contrast, temporal fake-score drift exhibited strong Spearman correlations of 0.968–0.980 across seeds under the 14B real-score model.

These observations were an important motivation for developing the fixed-paired drift probe. Rather than relying on loss values that can vary with sample difficulty or diffusion timesteps, we monitor changes in on-policy fake-score predictions under controlled sampling conditions, aiming to capture the temporal nonstationarity of the generator–critic system.

We agree that direct end-to-end comparisons with alternative loss- or gradient-based adaptive controllers would further strengthen this evidence, and appreciate the reviewer highlighting this direction.

Overall, we sincerely appreciate the reviewer's constructive feedback, which has helped us improve this work.

Reviewer_vv4MOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Drift-Gated TTUR, an adaptive optimizer schedule for long-horizon streaming distribution matching distillation (DMD). The authors propose measuring a statistic which they call the fixed-paired probe to measure drift in the current critic (fake score field) network while keeping prompts, latents, diffusion timesteps, and forward noise fixed. When this drift statistic exceeds a threshold calibrated by the median of the first MM non-null drift statistics, the controller skips a scheduled generator update and performs additional critic updates while keeping the underlying DMD objective and inferential procedure unchanged. Experiments use the LongLive pipeline and include VBench and VBench-Long evaluations, threshold ablations, and qualitative comparisons. The authors show that using this interventional procedure, one can obtain improved generation quality over uniform and random schedules with identical generator and critic update counts.

Critical strengths
  1. The proposed method addresses the interaction between an evolving generator distribution and online score estimation by adapting optimizer timing while leaving the underlying DMD objective and inferential procedure fixed. This is intuitively and practically appealing. Although this method is largely a heuristic statistic, I appreciate that the authors attempt to provide some theoretical justification by decomposing the gradient errors into a term governed by the sensitivity of the generator direction and another governed by the critic tracking error and arguing that indeed the proposed heuristic controls the critic tracking error. Additionally, although the local tracking error analysis holds only under stringent assumptions, it is a nice way to illustrate the author’s proposed method.

  2. The comparisons with uniform and random schedules are clearly explained and preserve both generator and critic update counts. On evaluated model checkpoints, Drift-Gated TTUR achieves VBench Total of 81.05, compared with 80.16 for the uniform schedule and 79.58 for the random schedule, providing some initial evidence that drift-conditioned scheduling can improve generation quality. I also find the qualitative examples encouraging.

Critical weaknesses

W1. Many of the experimental results are over multiple inferential runs and only one training run. Although the authors correctly acknowledge that the reported confidence interval measures inference-time variability, the proposed contribution is a training-scheduling heuristic which is dependent on the entire training trajectory. Without independent replications of multiple training runs, it is hard to be sufficiently convinced that the observed improvement represents a reliable benefit of the proposed method.

W2. Additional ablations are required to illustrate robustness across training runs. For example, does the scheduling advantage persist under many different meaningful training configurations such as different initialization or training budgets? Is it highly sensitive to the choice of probe bank? Furthermore, the paper provides a controlled scheduling comparison on VBench, but does not do this for VBench-Long or the qualitative examples. Including such a matched comparison in these evaluations would clarify whether the advantage is attributable to the proposed training heuristic.

Questions for authors

Q1. Does the advantage over the count-matched uniform control persist across independent training seeds, with optimizer counts matched separately for each run? Can the authors provide these results to illustrate the benefit of their method?

Q2. How robust is the improvement to different initializations, training budgets, and probe-bank choices? Can the authors provide VBench-Long results and qualitative comparisons against the count-matched uniform and random controls?

3 comments

W1/Q1. Robustness across independent training seeds

We thank the reviewer for highlighting training-run variability. We conducted two additional independent training-seed experiments, yielding three runs per configuration. Each run compares Drift-Gated TTUR, Fixed 5:1, and count-matched Uniform (U16), using 16 GPUs, batch size 16, and 2,000 iterations. Uniform is constructed separately for each run to match the adaptive run's generator and critic update counts. All checkpoints are evaluated on 946 VBench prompts using seven matched inference seeds.

Training runDrift-GatedFixed 5:1UniformΔ vs. FixedΔ vs. Uniform
Run 1 (original)81.0580.1480.16+0.91+0.89
Run 2 (added)80.4579.9480.28+0.51+0.17
Run 3 (added)81.3580.4679.96+0.89+1.39
Mean ± SD80.95 ± 0.4680.18 ± 0.2680.13 ± 0.16+0.77 ± 0.23+0.82 ± 0.61

Drift-Gated TTUR outperforms both controls in all three independent runs, improving VBench Total by +0.82 ± 0.61 over Uniform and +0.77 ± 0.23 over Fixed 5:1. Each run's score averages seven inference seeds; the final row reports mean ± sample SD across three training runs, not inference-seed uncertainty.

Despite varying improvement magnitudes, the consistently positive results support robustness across training trajectories. We added the summary to Section 5.1 and full results to Appendix G.

W2/Q2. Training checkpoints, probe-bank sensitivity, and VBench-Long

We thank the reviewer for these suggestions. We added checkpoint and probe-bank analyses and extended VBench-Long evaluation to count-matched controls.

Training checkpoints. From 500 to 2,000 iterations, Fixed 5:1 progressively loses motion, whereas Drift-Gated TTUR retains stronger dynamics. The VBench-Long Total gap also widens at later checkpoints (Figure 7, Appendix E.1). These are checkpoint comparisons, not equal-compute comparisons.

Probe-bank sensitivity. We vary bank composition and size with kappa=3 and M=40, recalibrating the threshold for each bank.

Probe configurationVBench TotalThreshold
Default Bank A (m=2)80.930.0927
Independent Bank B (m=2)80.580.0693
Extended Bank A (m=4)79.340.1105

Changing the bank reduces Total by 0.35, indicating some composition sensitivity. Increasing m to four reduces Total by 1.59 and adds approximately 23% training overhead. Larger probe banks are therefore not necessarily better (Appendix E.2).

VBench-Long matched comparisons.

ConfigurationQualitySemanticTotal
Drift-Gated TTUR0.8510.5500.791
Fixed 5:10.8180.5500.764
Uniform (U16)0.8110.5300.755
Random (Rand16)0.8030.5180.746

Our method outperforms Fixed 5:1, Uniform, and Random in VBench-Long Total by 0.027, 0.037, and 0.045, respectively, extending the scheduling advantage to long-horizon evaluation.

For seven matched inference seeds against Fixed 5:1, Total is 0.791 ± 0.002 vs. 0.764 ± 0.006, and Dynamic Degree is 0.746 ± 0.019 vs. 0.505 ± 0.033 (mean ± inference-seed SD). Uniform and Random scores are aggregates without reported seed uncertainty.

Together with W1/Q1, these results support robustness across training runs and checkpoints while identifying probe-bank sensitivity. We incorporated them into Sections 5.1–5.5 and Appendices E/G. Qualitative comparisons against Uniform/Random and alternative pretrained initializations remain untested.

Finally, we hope these additional results address the concerns and would appreciate reconsideration of the rating.

I would like to thank the authors for their detailed and thoughtful rebuttal. The responses have successfully addressed my initial concerns and in light of the responses to the other reviewers’ questions, I believe the revised manuscript is much improved and will increase my score. Good luck!

We sincerely thank you for your positive feedback and willingness to increase your score. We are glad that our responses have addressed your concerns, and we greatly appreciate your time and thoughtful consideration throughout the review process.

Reviewer_E67ZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes Drift-Gated TTUR, an adaptive optimizer schedule for DMD, which adapt the critic / student update ratio according to a probe in training.

When this probe exceeds a threshold, it skips a generator uodate and adds critic recovery steps.

Theoretically, precisely tracking the online distribution genrated by the student is required to ensure the convergence of DMD. Therefore, the motivation is clear, the topic is interatesing and important.

Critical strengths

1, This paper focus on the update ratio in DMD and demonstrate that an adaptive schedule brings better performances. The motivation is clear.

2, Qualitative comparison shows that the method show remarkable improvement compared with prior works. The drift issues is substantially mitigated in this work.

Critical weaknesses

1, The idea of adaptive critic update ratio applies broadly to DMD, but experiments are restricted to long video generation. It remains unclear whether the benefits generalize beyond this setting, e.g., to T2I and T2V.

2, Is there any general setting of probe design, threshold and critic recovery budget? How the setting is affected by the generation task and model size. Is there a best practice that works well on different situations?

3, The paper compares the default 5:1 schedule, count-matched uniform scheduling to adaptive schedule. It should also compare higher fixed ratios, i.e., fixed ratio of maximum ratio in the adaptive schedule as an upper bound setting.

Questions for authors

Please see the weakness

1 comment

W1. Generalization beyond long-horizon video generation

We thank the reviewer for this suggestion. To examine generalization beyond long-horizon generation, we conducted an additional short-horizon T2V experiment (81 frames, 16 fps, without autoregressive rollout). Both configurations share the same initialization, training data, and number of outer iterations, differing only in their update schedules. We evaluate on the official VBench suite (946 prompts).

ModelQualitySemanticTotalDynamic Degree
Wan2.1 teacher84.3065.4680.540.708
Fixed 5:1 TTUR81.8370.8979.640.153
Drift-Gated TTUR83.7669.4580.900.597

The results show that adaptive scheduling also benefits short-horizon T2V generation. Our method improves VBench Total by +1.26 over Fixed TTUR, while substantially preserving motion (Dynamic Degree: 0.597 vs. 0.153). The advantage also persists across the 6k/8k/10k checkpoints (average Total: 80.04 vs. 79.63). We provide side-by-side videos in Appendix C.

These results demonstrate that the benefit extends beyond long-horizon autoregressive rollout. We acknowledge that generalization to T2I and other DMD frameworks remains unverified, as clarified in Appendix C.

W2. Probe design, threshold calibration, and recovery budget

We thank the reviewer for this question. We recommend a small fixed-paired probe bank (m=2), median-based threshold calibration (M=40, kappa=3), and recovery budget K=H=5 as practical starting settings, rather than universally optimal values.

Probe design. We fix prompts, latents, diffusion timesteps, and forward noise while regenerating on-policy samples with the current generator. In our additional ablation, replacing the probe bank changes VBench Total from 80.93 to 80.58; increasing m from 2 to 4 reduces Total to 79.34 with 23% additional training overhead (Appendix E.2).

Threshold. We set the threshold to kappa times the median of the first M valid drift observations and freeze it after calibration. We recalibrate it separately for each run rather than transferring an absolute threshold. For kappa=2/3/4, VBench Total is 80.70/81.05/80.29, supporting kappa=3 among the tested settings (Section 5.4).

Recovery and transfer. We use K=H to allocate one additional baseline cycle of critic updates upon intervention. For new tasks or models, we recommend rebuilding the probe bank, recalibrating the threshold, and monitoring intervention frequency, quality, and overhead. While short-horizon T2V also benefits from our method, systematic sensitivity to K and model size remains untested. We have added these guidelines to Section 4.1.

W3. Higher fixed-ratio baseline

We thank the reviewer for this suggestion. We added Fixed 10:1 TTUR, corresponding to the maximum per-cycle critic budget H+K=10. Since recovery cycles skip generator updates, this is not a literal maximum adaptive ratio.

Under 2,000 outer iterations, Fixed 10:1 performs 200 generator and 2,000 critic updates, versus 400 generator and 2,000 critic updates for Fixed 5:1. All methods are evaluated using three matched inference seeds.

ConfigurationVBench TotalDynamic Degree
Fixed 5:180.460.62
Fixed 10:179.810.45
Drift-Gated TTUR80.930.83

Increasing the fixed ratio to 10:1 reduces Total by 0.65, while our method outperforms it by 1.12 points, consistently across all three seeds.

These results show that a higher fixed update ratio does not necessarily improve performance, while Drift-Gated TTUR achieves better results through adaptive scheduling. Together with the count-matched uniform comparison, this further supports the effectiveness of drift-conditioned update timing.

We have incorporated these results and optimizer counts into Section 5.1 and Appendix D. We hope these new results address the concerns raised and a reconsideration of the assessment.

Comments on the paper

2 comments

We sincerely thank all reviewers for their thoughtful feedback and constructive suggestions. We are encouraged that the reviewers recognized the motivation, simplicity, and practical relevance of Drift-Gated TTUR, as well as the promising improvements in generation quality.

In response to the reviews, we have substantially strengthened both the experimental evaluation and the presentation of our contributions. The revised manuscript includes several new experiments and analyses addressing the main concerns:

  1. Robustness across independent training runs. We conducted two additional independent training replications, yielding three runs per configuration. Drift-Gated TTUR consistently outperforms both Fixed 5:1 and count-matched Uniform scheduling, with average VBench Total improvements of +0.77 and +0.82, respectively (Appendix G).

  2. Stronger scheduling comparisons. We added a Fixed 10:1 baseline and extended the count-matched Uniform and Random comparisons to VBench-Long. These results further support the benefit of drift-conditioned update timing beyond simply increasing critic updates or changing optimizer-update counts (Section 5.1, Appendix D).

  3. Drift dynamics and controller sensitivity. We added three-seed drift comparisons against Fixed TTUR, showing approximately 3.1× lower mean probe drift, along with additional probe-bank and threshold sensitivity analyses. We also clarified the interpretation of the drift signal and provided practical configuration guidelines (Sections 3.2, 4.1, 5.4–5.5, Appendices E–F).

  4. Generalization beyond long-horizon rollout. We conducted an additional short-horizon T2V experiment, improving VBench Total from 79.64 to 80.90 and Dynamic Degree from 0.153 to 0.597. We also expanded the discussion of applicability to other DMD-based frameworks, including Causal Forcing (Appendix C, Section 3.2).

  5. Expanded qualitative evaluation and related work. We added six additional qualitative prompt groups covering diverse scenes (Appendix H) and strengthened the discussion of prior adaptive scheduling methods to clarify our DMD-specific contributions.

Importantly, these additions reinforce our central finding: adapting generator–critic update timing based on observable training dynamics can improve DMD training without modifying the underlying objectives or inference pipeline.

We have incorporated the new results, analyses, and clarifications into the revised manuscript and provided detailed responses to each reviewer. We hope the additional evidence addresses the concerns raised and respectfully invite the reviewers to reconsider their assessments.

Dear Reviewers,

We're the half-way through the discussion phase. The authors have provided their response to your reviews. Please engage with the authors as soon as possible, we need to keep the conversation going!

Many thanks,

AC

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

Related papers

Discussion (0)

Sign in to comment.