acceptodds
Under review as a conference paper at ICLR 2027

Not All Privilege Is Distillable: Calibrating Teacher Corrections for Few-Step Video Generation

Abstract

Privileged distillation uses clean video continuations to supervise a few-step generator on student-visited denoising states. We find that the largest evaluated evidence budget yields lower mean generation scores and motion activity than an intermediate budget, even when teacher and student share the first frame. To examine reference dependence, we introduce Privilege-Induced Teacher Variance (PITV), an offline measure of normalized variation in teacher corrections across screened references at fixed teacher weights and student states. Higher PITV is associated with poorer full-privilege transfer. For training, Mode-Calibrated Privileged Distillation (MCPD) tests support within a single continuation. It compares corrections from a progressively enlarged evidence subset and the full selected budget, using directional agreement and normalized displacement to weight the full correction at each token. Poorly supported targets move toward the teacher prediction conditioned only on the prompt and first frame. On 4-step Wan2.2 and 8-step LongCat-Video, MCPD improves VBench-I2V Total Score over timestep-dependent scalar shrinkage by 0.6–0.9 points and also outperforms adapted DFD. Controls on Wan2.2 show that matching residual energy or mixing the same teacher queries with fixed weights does not recover MCPD's performance.

Primary area:
generative models
Keywords:
Few-step Video Generation, On-Policy Distillation, Privileged Distillation, Video Diffusion Models, Image-to-Video Generation, Video Post-Training

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_Ae5AOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates why additional privileged information does not necessarily improve student performance in few-step video generation. The authors introduce Privilege-Induced Teacher Variance (PITV), an offline metric measuring the variability of teacher corrections across reference continuations, and show that higher PITV correlates with lower distillation gains. To address this, they propose Mode-Calibrated Privileged Distillation (MCPD), which adaptively weights teacher corrections based on directional agreement and normalized displacement. Experiments on Wan2.2 and LongCat-Video demonstrate improvements over several distillation baselines.

Critical strengths

The paper addresses an important limitation of privileged distillation for efficient video, and the motivation for the work is sound.
The methodological contribution is well motivated, and Section 3.3 provides a clear derivation of the calibration mechanism.
The experimental section is detailed and includes comparisons against multiple relevant baselines.

Critical weaknesses

I have almost no experience in this specific field, and this paper is generally far from my domain of expertise. However, although the paper is generally well organized, some parts are difficult to follow without substantial familiarity with privileged distillation and few-step video generation. Several technical concepts are introduced in rapid succession, making it challenging to understand the central intuitions behind PITV and MCPD. A more accessible explanation of the problem and the proposed solution would improve readability.

While MCPD preserves the computational cost of inference, its calibration mechanism requires additional teacher evaluations during training, and although the performance improvements are consistent, their magnitude is relatively modest against the strongest competitors.

Overall, as I said, this paper is very far from my domain of expertise. Therefore, I will carefully read the rebuttal of the authors also to the other reviewers to clarify my assessment and possibly update my score.

Questions for authors

Given the additional training cost introduced by MCPD, could the authors comment on whether the observed performance improvements justify this overhead in practical applications?

3 comments
AuthorsOpenReview ↗

Practical trade-off between post-training cost and video quality

Thank you for asking when the additional training compute is worthwhile. For a checkpoint that will be used repeatedly at a fixed sampling budget, we would choose MCPD when the resulting quality and preference gains matter to the application and the post-training budget can accommodate an additional 23 GPU-hours on Wan2.2 or 40 GPU-hours on LongCat. STEP-SCALAR is a reasonable choice when the post-training budget is tight or frequent retraining is required.

Why one extra teacher prediction? A prompt and an opening frame can allow several plausible camera movements. The teacher sees the recorded continuation, which shows one of them. At the same partly denoised video state produced by the student, STEP-SCALAR uses the difference between the teacher’s predictions with and without continuation evidence and applies a scalar weight for that denoising stage. MCPD makes one additional prediction using a smaller subset of the same continuation. For each spatiotemporal latent position, it retains more of the full correction when the two evidence budgets produce corrections with consistent directions and limited normalized displacement.

Ordinary MCPD training therefore uses one paired continuation and three teacher predictions per supervised student state, compared with two for STEP-SCALAR. PITV and the eight-reference comparisons are offline diagnostics, not additional components of ordinary MCPD training or deployment.

Measured cost and benefit. Table 3 and Appendix F.4 report one timed post-training allocation per configuration on eight H800 GPUs. The I2V gains summarize five paired training runs from a shared initialization per backbone, with quality evaluated at each run’s validation-selected checkpoint.

SettingSTEP-SCALAR GPU-hoursMCPD GPU-hoursExtra GPU-hoursI2V gain over STEP-SCALAR [95% CI]
Wan2.2, 4 steps, LoRA rank 12815217523 (+15.1%)+0.6 [0.38, 0.82]
LongCat, 8 steps, full-parameter22326340 (+17.9%)+0.9 [0.64, 1.16]

Timing includes data loading, forward and backward passes, and optimizer updates. It excludes validation search and checkpoint serialization; these percentages therefore describe matched post-training runs at the selected configurations, not the full model-development cost. On Wan2.2, the stronger four-step PER-STEP-SCALAR schedule in our additional control reduced MCPD’s advantage to +0.40 I2V points [0.20, 0.60].

The automatic-score gains are accompanied by a perceptible difference in the direct blinded comparison. Across the two backbones, 586 of 866 non-tie judgments favored MCPD over STEP-SCALAR: 67.7% [95% CI 62.8, 72.3], with 134 ties among 1,000 judgments. Each backbone contributed 50 condition-matched video pairs from the methods’ Run 1 validation-selected checkpoints. Intervals resample complete pairs within each backbone, retain all ten raters’ judgments for each pair, and are conditional on this rater panel and checkpoints. This comparison establishes preference over STEP-SCALAR; it does not make a human-preference claim about DFD.

Deployment invokes only the student, using its native sampler and the same four or eight denoising steps. No teacher prediction, PITV computation, multi-reference construction, or calibration module is used at deployment. Reusing a checkpoint spreads the additional post-training cost across the videos it produces, although the cumulative training-cost difference remains 23 or 40 GPU-hours. The reported comparison evaluates the specified training protocols rather than establishing which method would be optimal if both were assigned exactly the same total GPU-hour budget.

Readability. We will open Section 3 with the camera-motion example above and explicitly state that the student reaches each noisy state through its own sampling steps. We will relabel Figure 2 to distinguish training-only continuation inputs and the three teacher queries. Before introducing the equations, we will separate the two questions in plain language: PITV measures how a teacher correction varies across possible continuations at fixed states; MCPD measures how strongly one training continuation supports a correction across nested evidence budgets. We will also streamline the notation leading into Section 3.3 while retaining the derivation and estimator details in the appendices.

Dear authors,

I took some time waiting for the other reviewers to follow-up on their review, but I still can't see their new comments. As I already stated previously, my expertise is very far from this domain. However, I trust the judgment form the other reviewers, who responded positively to the paper. I also reviewed the rebuttal form the authors, to both my questions and the questions from the other reviewers. I believe these rebuttals added clarifications and new evidence to the original document.

Overall, I decided to raise my score to 3 -- Weak Accept.

Once again, I would like to repeat that this paper is very far from my expertise. I invite again the AC and the other reviewers to take this into account for their final decisions.

Best regards, Reviewer Ae5A

AuthorsOpenReview ↗

Thank you for the follow-up

Thank you for taking the time to read our responses and reconsider your assessment. We noticed that the rating field in the Official Review still shows 2 on our end. When you have a chance, could you please check whether it reflects the 3 (Weak Accept) you mentioned in your follow-up?

Best regards

Reviewer_n2BROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies privileged on-policy self-distillation for few-step image-to-video generation. The author observed that providing more privileged information (specifically more reference frame) may lead to weaker student after distillation. The author first provides an inference-time metric PITV (Privilege-Induced Teacher Variance) to analyze how the variance of the correction term correlate to the distillation improvement. Then, MCPD (Mode-Calibrated Privileged Distillation) is proposed to determine whether to distill the knowledge from the privileged teacher.

The strengths are:

  • Useful technique with solid empirical analysis. The new metric is useful for image-to-video generation and may inspire other follow-up works.
  • Detailed experiment on energy-matched, fixed-mixture, and shifted-gate controls, two backbones, and five runs with paired intervals. It make the result convincing.

The weaknesses are:

  • The somewhat related idea has been explored by related works, for example: phenomenon of “more privilege information can hurt” is discussed in ADVISOR, Weihs et al., NeurIPS 2021 (not cited). AVSD (NeurIPS 2026) also proposes a similar gating idea of directional agreement x proportionate magnitude. Other papers also discussed similar finding (the first paper is accepted and should be cited):
    1. Rethinking On-Policy Self-Distillation for Thinking Models (ICML 2026)
    2. Is More Privileged Information Better? (arXiv 2608.01589)
    3. Privileged Solutions or Context-Induced Teacher Behavior? (arXiv 2608.09228)
  • Magnitude of gains is somewhat limited. Over the strongest baseline, the gain is 0.6–0.9 points measured on VBench only.
  • Generality: it is unclear whether this approach can be applied to other task such as text-to-video generation
Critical strengths
  • Useful technique with solid empirical analysis. The new metric is useful for image-to-video generation and may inspire other follow-up works.
  • Detailed experiment on energy-matched, fixed-mixture, and shifted-gate controls, two backbones, and five runs with paired intervals. It make the result convincing.
Critical weaknesses
  • The somewhat related idea has been explored by related works, for example: phenomenon of “more privilege information can hurt” is discussed in ADVISOR, Weihs et al., NeurIPS 2021 (not cited). AVSD (NeurIPS 2026) also proposes a similar gating idea of directional agreement x proportionate magnitude. Other papers also discussed similar finding (the first paper is accepted and should be cited):
    1. Rethinking On-Policy Self-Distillation for Thinking Models (ICML 2026)
    2. Is More Privileged Information Better? (arXiv 2608.01589)
    3. Privileged Solutions or Context-Induced Teacher Behavior? (arXiv 2608.09228)
  • Magnitude of gains is somewhat limited. Over the strongest baseline, the gain is 0.6–0.9 points measured on VBench only.
  • Generality: the proposed loss heavily relies on the reference frame format, which is a co-design of loss, model and input data. it is unclear whether this approach can be applied to other task such as text-to-video generation
Questions for authors
  • On the 8-reference PITV set, do tokens with low gate values have higher cross-reference variance? How does MCPD compare with an AVSD-style gate based on agreement across multiple references, evaluated where multiple references are available?
  • Can you more carefully address the difference to the prior related works or conduct ablation experiment in a similar setting if possible?
3 comments
AuthorsOpenReview ↗

Reply 1 of 3 — Relation to ADVISOR, privileged OPSD, and AVSD

Thank you for pointing out these connections. Kaur, Zhao, Ichihara, and AVSD were already cited in Section 2. In the revision, we will expand their comparison with our work, add the missing ADVISOR citation, and update the publication details for Kaur et al. and AVSD to NeurIPS 2026.

The distinctions are as follows:

  • ADVISOR (Weihs et al.) established that privileged expert supervision can create an imitation gap. It uses an auxiliary policy to estimate whether expert behavior is reproducible from student-visible information and adaptively weights imitation and reward-based RL losses. MCPD instead calibrates continuous velocity corrections at fixed diffusion states, using nested evidence budgets from one paired continuation, without an auxiliary policy or RL loss.

  • Kaur et al. showed that privileged-context distillation can harm long-form reasoning and analyzed local changes around high-entropy forks and self-corrections. We study a different manipulation: increasing clean video-continuation evidence and measuring how the resulting teacher corrections vary across possible futures and transfer to a few-step video student.

  • Zhao et al. (PS-OPSD) replace a complete solution with structured problem-space guidance. They therefore change the representation of privilege. MCPD retains the same paired continuation but changes its exposure budget and the token-wise weight of the resulting correction.

  • Ichihara et al. (OP²SD) replace the paired reference with another problem and solution, revealing context-induced teacher behavior. We vary candidate futures under the same prompt and first frame while holding teacher weights and student-visible states fixed. This measures reference dependence, but does not claim to separate every possible context effect.

  • AVSD is the closest gating precedent. It aggregates multiple privileged views in vocabulary-probability space, retains a geometric-consensus update, and adds a residual when cross-view direction and magnitude are consistent. MCPD instead compares two nested budgets from one continuation, operates on continuous velocity vectors, and falls back toward the deployment-conditioned prediction v0v^0, rather than retaining a privileged consensus floor.

Our existing CR-VEL control implements only a single-reference, two-budget velocity-space analogue. On Wan2.2, CR-VEL scored 86.5 and MCPD scored 87.2; the mean paired MCPD−CR-VEL difference was +0.7 across five runs (paired SD 0.19, Table 8). CR-VEL neither reads the eight references nor isolates fallback choice; the next reply separately reports the requested multi-reference comparison.

Our contribution is the combination of three measurements and design choices in few-step video generation: (i) separating teacher reference fidelity from student benefit as the clean temporal-latent budget changes—on Wan2.2, increasing fixed BB from 4 to 8 reduced student I2V from 85.9 to 85.6; (ii) measuring cross-reference correction variation at fixed teacher weights and student states; and (iii) evaluating a calibration rule that requires only one paired continuation during ordinary training. PITV is an offline diagnostic, not an inference-time metric; deployment runs only the student.

AuthorsOpenReview ↗

Reply 2 of 3 — Token-level reference variance and a multi-reference gate comparison

Lower MCPD gates were associated with higher cross-reference variance, including when variance and energy were computed entirely from the other seven references. We evaluated the actual gates on the 184 retained conditions using fixed snapshots, saved states, noise steps, reference ordering, and valid output positions Ω.

For each output position, let g(m) be the gate for reference m, ν the raw cross-reference variance, e the correction energy, and p = ν/(e + ε). We compared (i) the mean gate across eight references with variation across those eight and (ii) each g(m) with variation computed from the other seven. Raw variances in the table are channel-averaged squared-prediction units multiplied by 1,000.

AnalysisRaw ν, low/high gateLow−high [95% CI]Normalized p, low/high gateLow−high [95% CI]
Eight-reference mean gate0.294 / 0.176+0.118 [0.086, 0.151]0.413 / 0.287+0.126 [0.091, 0.161]
Leave-one-reference-out gate0.263 / 0.193+0.070 [0.043, 0.097]0.376 / 0.302+0.074 [0.042, 0.106]

After adjustment for log correction energy and noise step, the partial rank associations between gate and normalized variation were −0.32 [−0.41, −0.23] for the eight-reference analysis and −0.21 [−0.29, −0.12] for leave-one-reference-out. The latter used energy computed from the other seven references as its covariate.

Low/high denotes the bottom/top gate quartiles. We assigned midranks within each condition for the eight-reference analysis and within each condition–reference pair for leave-one-out, keeping boundary ties together. For leave-one-out, we averaged over references within each condition. Conditions were then weighted equally. Zero gates accounted for 18.6% of reference-token observations. We applied no additional token-level energy cutoff. All intervals use 5,000 complete-condition bootstrap resamples, retaining all references, states, and tokens within each condition. These results support token-level empirical alignment under the screened reference distribution Q. The limits on interpreting the gate as a PITV estimator in Appendix B.2 still apply.

We next evaluated an eight-reference, velocity-space adaptation of AVSD’s agreement and proportionality gate. Let δ(m) denote the full correction for reference m, v⁰ the anchor-only prediction, and norm(x) the Euclidean norm of vector x. Both means below are over references 1 through 8 at the same output position.

c = mean of δ(j) over all eight references j.

A = mean of norm(δ(j)) over all eight references j.

r(m) = δ(m) − c.

h(m) = [norm(c) / (A + ε)] × [norm(c) / (norm(c) + norm(r(m)) + ε)].

t(m) = v⁰ + c + h(m) × r(m).

We compared all targets on the 184 conditions using the block interventions in Appendix E.3. Each method intervened on the same block at the same saved student state. We held the negative-branch prediction at that state, the CFG rule, the solver, and subsequent random increments fixed. Later predictions were recomputed along each intervened trajectory. The reference index m cycled deterministically across condition–state pairs. FULL, MCPD, and XR-AVSD-VEL used the same selected m, with no averaging of targets over m. Utility is the change in the Eq. (35) composite score relative to the matched anchor-only intervention (Eq. (36)).

TargetDirect-use utility vs. anchor-only [95% CI]Retained energy vs. FULL
FULL, selected reference−0.02 [−0.06, 0.02]1.00
Single-reference MCPD+0.12 [0.09, 0.15]0.48
Eight-reference consensus (v⁰ + c)+0.06 [0.03, 0.09]0.43
Eight-reference XR-AVSD-VEL+0.09 [0.06, 0.12]0.54

The direct paired differences were XR−consensus = +0.03 [0.01, 0.05] and MCPD−XR = +0.03 [0.01, 0.05]. Adding the gated residual improved on consensus alone. Single-reference MCPD had the highest direct-use utility in this fixed-state comparison.

The normalized squared XR–consensus distance was E[norm(t(m) − (v⁰ + c))²] / E[norm(δ(m))²] = 0.12. Both expectations use the same selected references, masks, and weights as the energy table, with the same FULL-energy denominator. The gated residual therefore remained nonzero under this aggregation.

The experiment evaluates complete targets at fixed states. Retained energy is descriptive, and the comparison leaves open the specific contribution of gate location and the outcome of full multi-reference retraining. Per state, MCPD uses three logical teacher queries for one reference, each eight-reference target uses nine, and the offline all-reference gate analysis uses 17 with a shared anchor. These counts exclude reference generation and screening. XR has access to eight references per condition and a larger query budget than ordinary MCPD training.

AuthorsOpenReview ↗

Reply 3 of 3 — Effect sizes, blinded preference, and text-only transfer

We agree that the absolute aggregate-score gains are moderate, so we report each comparator explicitly rather than describing them as uniformly large:

Paired I2V comparisonWan2.2LongCat
MCPD−STEP-SCALAR+0.6 [0.38, 0.82]+0.9 [0.64, 1.16]
MCPD−DFD†+0.7 [0.46, 0.94]+0.8 [0.53, 1.07]
MCPD−FULL-PRIV PD+1.6 [1.30, 1.90]+2.0 [1.66, 2.34]

Wan2.2 LoRA ranks 64/128/256 also give MCPD−STEP differences of +0.4/+0.6/+0.7, with all three intervals above zero; these results support consistency across tested ranks, not a statistically established rank trend. For completeness, the added four-step PER-STEP-SCALAR scores 86.8, and MCPD retains a +0.40 [0.20, 0.60] paired advantage over it.

The evidence is also not limited to automatic VBench scores. In the direct blinded MCPD–STEP-SCALAR study, each backbone contributed 50 condition-matched video pairs, with ten raters per pair:

BackboneMCPD winsTiesSTEP winsNon-tie MCPD preference [95% CI]
Wan2.22857014566.3% [59.4, 72.6]
LongCat3016413569.0% [62.1, 75.2]

Overall, 586 of 866 non-tie judgments favored MCPD: 67.7% [62.8, 72.3], with 134 ties among 1,000 judgments. Intervals resample complete video pairs within backbone and retain all ten judgments per pair; the pooled analysis is backbone-stratified. They are conditional on this rater panel and the Run 1 checkpoints. This human result concerns STEP-SCALAR specifically, not DFD.

Regarding T2V, Table 1 includes the fixed BASE initialization and models post-trained for shared-first-frame I2V. During MCPD post-training, the student and teacher received the same prompt and first frame, and only the teacher received additional clean continuation evidence. For T2V evaluation, we ran the same checkpoints through their native text-only interfaces using the prompt alone:

BackboneSTEP-SCALARDFD†MCPD
Wan2.2 T2V84.484.584.8
LongCat T2V85.285.385.8

Relative to fixed BASE, MCPD’s T2V gains are +1.4 [1.05, 1.75] and +1.6 [1.21, 1.99]; the +0.4/+0.6 differences relative to STEP-SCALAR are separate comparisons.

These results demonstrate cross-condition transfer after I2V post-training, rather than direct on-policy distillation on text-only student trajectories. The current training implementation assumes that the teacher can receive nested clean continuation slices at the student’s state and produce aligned velocity predictions. Within that assumption, the results cover two backbones’ native text-only generation interfaces; we do not claim unrestricted generality to other tasks or privilege formats.

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

This paper shows that giving the teacher more privileged future frames does not always help a few-step I2V student, and they use PITV to measure how much the teacher's correction depends on the reference. The proposed MCPD compares a smaller-budget correction with the full one at each token and uses a gate to scale the full correction. On wan2.2 and LongCat-Video, MCPD beats scalar shrinkage in terms of VBench-I2V points and also adapted DFD.

Critical strengths
  1. I like the question set up. Figure 3 makes a nice point that a teacher that copies the reference better does not necessarily give us a better student
  2. The experiments look careful. MCPD beats a tuned STEP-SCALAR on both backbones with paired CIs above zero. On Wan2.2, neither energy matching nor fixed-weight mixing of the same teacher predictions recovers the gain, and the ablations show both parts of the gate help.
  3. It's free at inference time, and the extra training cost is modest.
Critical weaknesses
  1. I'm not fully convinced the gate is really tied to PITV. PITV looks at how corrections vary across references, but the gate compares two budgets within a single reference... The authors even give a case in the appendix B, where the gate is close to one while PITV is at its max. So, I cannot tell whether the gate is actually fixing the problem that PITV points to. A simple check I can think of would be: on high-PITV conditions, does MCPD win back more over FULL-PRIV PD? If you show MCPD, FULL-PRIV PD and EMA-SD scores separately within each PITV bin, we can see whether a bigger gap comes from FULL-PRIV PD getting worse, MCPD getting better, or both. For the diagnostic itself, I appreciate that the correlation still holds on the 45 conditions where all references were judged action-compliant, so obvious action mismatch seems unlikely to explain it. Still, the references come from the base multi-step generator rather than the real training videos, and the correlation is not controlled for simple things like correction energy or base-model dynamic degree, which could influence both PITV and the measured transfer gain.
  2. How were beta and gamma picked? MCPD uses (Bmin⁡,Bprobe,Bmax⁡,γ,β)=(1,4,8,1,2)(B_{\min},B_{\text{probe}},B_{\max},\gamma,\beta)=(1,4,8,1,2) in table 4, but I could not find how these values were chosen or how many settings you have tried. STEP-SCALAR, on the other hand, is picked from a 4x4 grid. The ablations tell that each piece of the gate helps, but not whether things hold up for beta around 2. I would also like to know how much of the gap a per-step scalar schedule (K weights instead of just 2) would close. To me, that's the cleanest way to tell the token-wise gate than just having more flexible timestep weighting.
  3. The human study seems to skip the strongest baselines. Against the STEP-SCALAR and DFD, all the evidence is VBench. Neither the human study in appendix, nor the short-update probe compares MCPD with STEP-SCALAR directly. The paired CIs do make the VBench gain clear for these two initializations, but I'm left wondering whether human raters would actually prefer MCPD's videos over STEP-SCALAR's. Besides, I notice that DFD has higher A-Div. than MCPD on both backbones in table 1, so beating DFD may cost a bit diversity...

In summary, I lean to acceptance. My main reservations are whether the gate really connects to PITV, and whether human raters would prefer MCPD over the strong baselines.

Questions for authors
  1. On high-PITV conditions, does MCPD win back more over FULL-PRIV PD? Could you show each method's score within PITV bins on the 184 conditions?
  2. How did you pick the beta and gamma, and how many settings did you try, compared with the 16 for STEP-SCALAR? How does MCPD do against a per-step scalar schedule?
  3. Could you also consider to add a blind human comparison of MCPD vs STEP-SCALAR if accepted?
4 comments
AuthorsOpenReview ↗

Reply 1 of 3 — PITV-stratified student recovery

Thank you for suggesting this comparison. Yes. Across the 184 retained Wan2.2 I2V conditions, MCPD gained more over FULL-PRIV PD in the high-PITV tercile. The high-minus-low difference in this gain was +0.81 points [95% CI 0.55, 1.07].

PITV tercilenEMA-SDFULL-PRIV PDMCPDMCPD−FULL [95% CI]
Low6285.8285.8786.45+0.58 [0.43, 0.73]
Middle6185.3184.9785.80+0.83 [0.64, 1.02]
High6184.8784.1685.55+1.39 [1.17, 1.61]

All methods used the same terciles, defined by the original PITV values. Absolute scores declined from the low to the high tercile for all three methods, but their changes relative to EMA-SD were markedly different. FULL−EMA moved from +0.05 in the low tercile to −0.71 in the high tercile, a high-minus-low change of −0.76. MCPD−EMA moved from +0.63 to +0.68, a change of only +0.05; its high-tercile mean difference was +0.68 [0.49, 0.87]. Thus, the wider MCPD−FULL gap mainly reflects FULL-PRIV PD’s poorer relative performance under high PITV, rather than an increasing MCPD advantage over EMA-SD.

The STEP-SCALAR comparison shows the same direction. In the high tercile, STEP-SCALAR averaged 85.01, compared with 84.87 for EMA-SD and 85.55 for MCPD. MCPD−STEP was +0.54 [0.35, 0.73] in the high tercile and +0.18 [0.04, 0.32] in the low tercile, giving a high-minus-low contrast of +0.36 [0.11, 0.61]. MCPD’s recovery therefore also exceeds that of the validation-selected early/late schedule.

Figure 3(c) reports a raw Spearman correlation of −0.83 between PITV and FULL−EMA. After rank residualization on fixed-state full-correction energy and per-condition frozen-BASE Dynamic Degree, the adjusted association was −0.75 [−0.82, −0.66]. For MCPD−FULL, the raw and adjusted associations were +0.58 [0.44, 0.69] and +0.51 [0.35, 0.64], respectively. These are adjusted associations, not causal estimates.

Scores use the per-video I2V composite in Eq. (33), averaged over five matched run IDs and four generation seeds per condition, with each method evaluated at its own validation-selected checkpoint. The new analyses use 5,000 condition-bootstrap resamples, retaining method pairing within conditions and recomputing group means and contrasts after each resample. The high and low terciles contain distinct conditions and are not treated as paired groups. Intervals are conditional on the evaluated runs and checkpoints. Reply 2 reports official full-suite I2V scores.

These results establish condition-level alignment between PITV and student outcomes, not token-level equivalence between PITV and the gate. PITV measures variation across eight screened continuations from the base multi-step generator at fixed teacher weights and student states. During training, MCPD tests nested-budget agreement for one paired continuation; Appendix B.2 explains why a stable but reference-specific correction can pass. The association is therefore stated under the screened reference distribution Q, without assuming that Q equals the real training-reference distribution.

AuthorsOpenReview ↗

Reply 2 of 3 — Parameter selection, β robustness, and per-step scalar weighting

We reused B=4 from the four-budget BEST-FIXED PD sweep as the MCPD probe cap. We fixed B_min=1 and B_max=8, and chose γ=1 for a linear allocation in normalized noise before rounding down. With these choices fixed, we compared three Wan2.2 MCPD configurations, β∈{1,2,4}. The mean I2V score on the 200 validation conditions selected β=2. We reused (B_min,B_probe,B_max,γ,β)=(1,4,8,1,2) on LongCat without another search. The same validation conditions served for checkpoint selection, and test scores played no role. STEP-SCALAR used a separate 16-pair early/late grid and selected (0.75,0.50).

For the sensitivity analysis, we reused the existing β=2 runs and evaluated five matched runs at each β:

βWan2.2 I2V, five-run meanFraction of full-correction energy retained
186.90.61
287.20.48
487.00.35

Energy ratios were measured on common frozen student states with the other gate inputs fixed. Across the tested values, retained energy decreased from 0.61 to 0.35 while five-run mean I2V ranged from 86.9 to 87.2; STEP-SCALAR scored 86.6. Thus, β changes attenuation substantially over this neighborhood, while the observed performance remains above the scalar baseline’s mean.

For four-step Wan2.2, we also trained PER-STEP-SCALAR with a separate coefficient from {0.25,0.50,0.75,1.0} at each step. Starting from the selected early/late schedule expanded to (0.75,0.75,0.50,0.50), a single coordinate pass evaluated 13 of the 256 possible schedules on the same validation set: one initialization and three alternatives per coordinate. From highest to lowest noise, the selected coefficients were (0.75,0.75,0.50,0.25).

The selected per-step schedule scored 86.8, compared with 86.6 for STEP-SCALAR and 87.2 for MCPD. It recovered 0.20 of the original 0.60-point gap in five-run means, with PER-STEP−STEP = +0.20 [0.02, 0.38]. MCPD retained a +0.40-point [0.20, 0.60] advantage over PER-STEP-SCALAR.

Thus, the tested per-step schedule closes approximately one-third of the original gap. This is a direct evaluation of the selected 13-configuration search, not a claim that all 256 schedules were exhaustively optimized. The remaining gap is not explained by this per-step scalar schedule; the comparison does not uniquely apportion it among token weighting, adaptation to the sample and student state, and the additional teacher query.

AuthorsOpenReview ↗

Reply 3 of 3 — Human preference, short-update learning, and DFD diversity

We conducted a blinded comparison of MCPD and STEP-SCALAR on 50 condition-matched video pairs per backbone. Each condition contributed one pair, and the two backbones used disjoint condition sets. Each method used Run 1’s validation-selected checkpoint. Conditions were sampled independently of VBench differences, PITV, and qualitative outcomes. Prompts, anchors, random inputs, and sampling settings were matched; method identities were concealed, left/right order was randomized, and ties were allowed. All ten raters evaluated every pair.

BackboneMCPD winsTiesSTEP winsMCPD share of non-tie judgments [95% CI]
Wan2.22857014566.3% [59.4, 72.6]
LongCat3016413569.0% [62.1, 75.2]

Overall, 586 of 866 non-tie judgments favored MCPD (67.7% [62.8, 72.3]), with 134 ties among 1,000 judgments. Intervals were obtained by resampling video pairs within each backbone, keeping all ten judgments for a pair together; the pooled interval used backbone-stratified resampling. Individual ratings were therefore not treated as independent replicates. These intervals are conditional on the ten-rater panel and Run 1 checkpoints, and quantify a different source of uncertainty from the main table’s five-run training variation.

Raters preferred MCPD overall and for motion naturalness. Motion-naturalness preference was 64.9% [59.3, 70.1]. Subject-consistency preference was 54.6% [49.5, 59.7]; because this interval includes 50%, the direction of preference on that dimension is not resolved.

We also added STEP-SCALAR to the short-update probe. After five updates on 128 conditions, we evaluated 128 disjoint conditions using the learning-utility composite in Eqs. (38)–(39). Gains over matched anchor-only updates were +0.08 [0.02, 0.14] for STEP-SCALAR and +0.15 [0.08, 0.22] for MCPD. The direct paired MCPD−STEP contrast was +0.07 [0.01, 0.13], bootstrapped over complete held-out evaluation conditions.

DFD had higher mean A-Div. on both backbones: DFD/MCPD = 0.168/0.165 on Wan2.2 and 0.174/0.172 on LongCat. The five-run paired MCPD−DFD differences were −0.003 [−0.006, 0.001] and −0.002 [−0.005, 0.001], respectively. A small cost in time-averaged appearance diversity therefore remains plausible, although these paired runs do not resolve the differences. A-Div. measures variation in temporally averaged appearance features, not temporal-mode diversity, so these results do not support a broader diversity-loss claim.

I thank the authors for the detailed and high-quality response and new experiments. They addressed my main concerns. If you feel these new experiments help with the paper, please consider adding them to the revised paper. A small finding: in table 1, it seems on the full-suite, FULL_PRIV PD is better than EMA-SD, but from your reply 1, EMA-SD is better. I feel like this mainly is due to the composite's weighting or the diagnostic condition set.

Overall, I believe this is a high quality paper, especially authors address my main concerns pretty well. I will maintain my positive assessment to this paper. :)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.