acceptodds
Under review as a conference paper at ICLR 2027

On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators

Abstract

We propose \method, an on-policy self-distillation paradigm for post-training few-step autoregressive (AR) video diffusion models. Existing few-step AR video generators, often obtained through DMD-style distillation from short-clip video teachers, achieve low latency but suffer from error accumulation and weakened motion dynamics during long rollouts. \method incorporates real long-video data into post-training, rather than relying on short-clip video teachers, to further reduce long-horizon degradation and improve motion dynamics while preserving the original few-step inference path. Specifically, we use the DMD-distilled few-step model for both the student and teacher branches. The student follows the exact inference-time rollout, generating each chunk conditioned on its own previously generated KV cache. The teacher is evaluated at the same student-visited denoising states, but uses a cleaner, AR-consistent temporal cache in which older history is replaced with real-video context. The teacher then provides dense trajectory-level supervision to the student. To maintain autoregressive consistency and prevent the teacher from becoming a fully teacher-forced oracle, both branches share an initial real-video prefix, and the teacher retains its most recent model-generated cache chunk. Experiments on Self-Forcing and LongLive demonstrate improved visual quality and motion dynamics, with higher VBench-Long scores. A user study further shows that \method is preferred over the base models in 66.0% of overall-preference judgments (82.5% excluding ties). Visual results, code, and model checkpoints are available through our anonymous https://anonymous.4open.science/w/OPSD-V-2027/project page, https://anonymous.4open.science/r/OPSD-V-Code-2027/code repository, and https://huggingface.co/anonymousopsd-v/OPSD-VHugging Face repository, respectively.

Primary area:
generative models
Keywords:
Autoregressive Video Generators, On-Policy Self-Distillation, Few Step, Post-Training

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_cGMmOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes OPSD-V, a post-training method for few-step autoregressive video generators. A real first chunk initializes the training caches. The student then generates subsequent chunks using its own history. An EMA teacher predicts velocities at the same noisy states, but uses real-video features for an older cached chunk while keeping its latest chunk model-generated. The student matches the teacher at all four denoising steps, starting at chunk 8. Inference retains the original sampler and cache policy.

The authors evaluate the method on Self-Forcing and LongLive using 240 prompts for one-minute generation. They also report ablations and a paired preference study.

Critical strengths
  1. The teacher cache offers a useful way to bring real video into training. The student remains on its own rollout, while the teacher can use a cleaner recent history. In the training-free example in Figure 7, replacing the latest generated chunk with real-video KV introduces a visible boundary artifact. Keeping that chunk generated is therefore a reasonable design choice, though its effect after training still needs testing.

  2. The method improves both starting models under matched inference settings. OPSD-V raises the reported Quality and Dynamic Degree scores on both backbones. On Self-Forcing, aesthetic and imaging quality also improve. The preference study favors OPSD-V overall, although it covers only ten distinct video pairs per backbone. The denoising-step and latest-KV ablations add useful evidence.

Critical weaknesses
  1. The experiments do not yet isolate the benefit of the proposed cache policy. D-OPSD already matches a privileged teacher’s velocities at the student’s own denoising states in text-to-image diffusion. Adapting that idea to a causal video KV cache is worthwhile, but the choice of real and generated history is the key new part here. The cache ablation in Table 2 changes whether the latest teacher-cache chunk comes from the teacher or the student. It does not compare the proposed delayed replacement with an all-real teacher history after training under the same budget. Figure 7 makes a case for the design through a training-free example, but cannot establish which policy produces the better post-trained model. Under the stated attention window, the teacher sees only one recent real-video chunk beyond the shared first chunk and its latest generated chunk, making that comparison particularly relevant.

  2. Less degradation late in the video has not been measured directly. The main metrics average over the whole minute. For LongLive, Quality rises from 0.7867 to 0.7992, but its five components other than Dynamic Degree all fall slightly. Dynamic Degree measures how much motion is present, not how good it looks. The human preference results offer some positive evidence, but do not assess the later part of the video separately. The configured full-length training rollout contains 60 chunks, whereas evaluation generates 81. The final 21 chunks, about 16 seconds, therefore extend beyond the training rollout length.

Questions for authors
  1. Could you train an all-real-history teacher-cache variant using the same student rollout procedure, velocity loss, data, and training budget? How does its full metric breakdown, especially on later video segments, compare with the proposed delayed replacement?

  2. Could you report the existing aesthetic, imaging, subject and background consistency, motion smoothness, and Dynamic Degree scores separately for early, middle, and late segments, with prompt-level uncertainty intervals? In particular, do the final 21 chunks improve on LongLive when Dynamic Degree is reported separately?

5 comments
AuthorsOpenReview ↗

Response to Reviewer cGMm (1/3) — Q1: All-Real History vs. Delayed Replacement

We invite you to view the videos and tables on our anonymous rebuttal page.

Q1. How does an all-real teacher cache compare with delayed replacement?

We thank the reviewer for suggesting this comparison. We trained All-real variants of LongLive and Self-Forcing. These variants also overwrite the latest completed chunk in the teacher’s KV cache with time-aligned ground-truth KV. This post-training comparison extends our training-free diagnostic in Fig. 7. OPSD-V achieved higher Quality and Dynamic Degree than All-real on both backbones.

All-real lowered full-video Quality from 0.7992 to 0.7179 on LongLive and from 0.8158 to 0.7114 on Self-Forcing. Dynamic Degree showed the largest drop over both the full video and the final 21 chunks. Over the final 21 chunks, Semantic scores rose slightly, while VBench-Long Imaging quality fell from 0.7080 to 0.6193 on LongLive (a decrease of 0.0887) and from 0.7103 to 0.6287 on Self-Forcing (a decrease of 0.0816).

VBench-Long scores

Metric ↑LongLive OursLongLive All-realSelf-Forcing OursSelf-Forcing All-real
Quality0.79920.71790.81580.7114
Semantic0.50680.50660.50460.5023
Subject consistency0.97130.90210.96330.8823
Background consistency0.95580.92180.94910.9176
Aesthetic quality0.56680.54770.59700.5460
Imaging quality0.70600.61720.73010.6181
Motion smoothness0.98750.98700.98390.9842
Dynamic Degree0.60770.33170.67150.3205
Object class0.98830.96860.97750.9680
Multiple objects0.91330.86090.90740.8598
Color0.34090.33940.33890.3234
Scene0.35080.41710.35490.4129
Appearance style0.22720.23300.22530.2273
Overall consistency0.22050.22040.22380.2224

Full-video scores follow the paper’s VBench-Long protocol. Segment scores, including the final-21-chunk table below, evaluate each segment as a separate video.

VBench-Long scores (final 21 chunks)

Metric ↑LongLive OursLongLive All-realSelf-Forcing OursSelf-Forcing All-real
Quality0.79610.71550.79330.6945
Semantic0.49700.50240.48930.4989
Subject consistency0.93470.91660.92770.8971
Background consistency0.94100.93560.93490.9257
Aesthetic quality0.57170.55220.55610.5506
Imaging quality0.70800.61930.71030.6287
Motion smoothness0.98660.98510.98430.9817
Dynamic Degree0.63460.28440.64630.1831
Object class0.98940.96870.97750.9648
Multiple objects0.91590.85680.89540.8566
Color0.31110.32630.30160.3204
Scene0.32820.41160.32730.4044
Appearance style0.22760.23360.22270.2295
Overall consistency0.21010.21720.21140.2175

We evaluate chunks 61–81 as a separate video. Quality is the mean of its six components.

Side-by-side video comparisons

AuthorsOpenReview ↗

Response to Reviewer cGMm (2/3) — Q2: LongLive Segment-Level Results

Q2. How does quality change across the rollout and in the final 21 chunks?

We thank the reviewer for suggesting this segment-level analysis. We compare the base models with our OPSD-V checkpoints using the same random seed within each pair. We report six dimensions across the early, middle, and late thirds, as well as chunks 61–81, which extend beyond the 60-chunk training rollout. Each segment is evaluated as a separate video.

From the early third to the final 21 chunks, the mean excluding Dynamic Degree fell by 0.0013 with OPSD-V and 0.0075 without it on LongLive, and by 0.0058 and 0.0132 on Self-Forcing. For both backbones, the paired 95% bootstrap confidence intervals (CIs) for the Imaging quality difference include zero in the early third and lie entirely above zero in the late third and the final 21 chunks.

In the final 21 chunks, the mean excluding Dynamic Degree was higher by +0.00247 (95% CI [−0.00090, +0.00570]) on LongLive and by +0.00587 (95% CI [+0.00238, +0.00937]) on Self-Forcing. Dynamic Degree changed by +0.05342 (95% CI [−0.00112, +0.10876]) and +0.01945 (95% CI [−0.02356, +0.06268]), respectively. On both backbones, OPSD-V had higher Aesthetic quality and Imaging quality point estimates and slightly lower Subject consistency. Background consistency was nearly unchanged.

LongLive

Early · chunks 1–27

Metric ↑BaseOPSD-VΔ95% CI
Aesthetic quality0.58980.5880−0.00181[−0.00632, +0.00272]
Imaging quality0.71540.7147−0.00070[−0.00468, +0.00325]
Subject consistency0.93620.9280−0.00820[−0.01129, −0.00518]
Background consistency0.93690.9299−0.00699[−0.00945, −0.00462]
Motion smoothness0.98890.9881−0.00079[−0.00122, −0.00034]
Dynamic Degree0.51140.5946+0.08325[+0.04428, +0.12458]
Mean excluding Dynamic Degree0.83340.8297−0.00370[−0.00538, −0.00204]

Middle · chunks 28–54

Metric ↑BaseOPSD-VΔ95% CI
Aesthetic quality0.57380.5773+0.00356[−0.00296, +0.00992]
Imaging quality0.70590.7075+0.00152[−0.00433, +0.00732]
Subject consistency0.93590.9320−0.00392[−0.00901, +0.00166]
Background consistency0.94250.9409−0.00164[−0.00430, +0.00108]
Motion smoothness0.98690.9872+0.00032[−0.00028, +0.00096]
Dynamic Degree0.54170.5735+0.03188[−0.01029, +0.07549]
Mean excluding Dynamic Degree0.82900.8290−0.00003[−0.00288, +0.00289]

Late · chunks 55–81

Metric ↑BaseOPSD-VΔ95% CI
Aesthetic quality0.56710.5720+0.00488[−0.00219, +0.01184]
Imaging quality0.70140.7083+0.00688[+0.00078, +0.01306]
Subject consistency0.93410.9307−0.00336[−0.00922, +0.00242]
Background consistency0.94020.9396−0.00056[−0.00382, +0.00273]
Motion smoothness0.98650.9867+0.00023[−0.00052, +0.00103]
Dynamic Degree0.57080.6237+0.05288[+0.00919, +0.09860]
Mean excluding Dynamic Degree0.82590.8275+0.00161[−0.00150, +0.00458]

Final 21 · chunks 61–81

Metric ↑BaseOPSD-VΔ95% CI
Aesthetic quality0.56610.5717+0.00565[−0.00173, +0.01301]
Imaging quality0.70030.7080+0.00768[+0.00075, +0.01456]
Subject consistency0.93570.9347−0.00100[−0.00710, +0.00499]
Background consistency0.94120.9410−0.00018[−0.00339, +0.00292]
Motion smoothness0.98640.9866+0.00022[−0.00059, +0.00105]
Dynamic Degree0.58120.6346+0.05342[−0.00112, +0.10876]
Mean excluding Dynamic Degree0.82590.8284+0.00247[−0.00090, +0.00570]

Δ is OPSD-V minus Base. We compute the 95% confidence intervals by resampling prompt pairs within each benchmark and weighting the two benchmarks equally. The mean excluding Dynamic Degree gives equal weight to the other five scores.

Quality and motion by segment

AuthorsOpenReview ↗

Response to Reviewer cGMm (3/3) — Q2: Self-Forcing Segment-Level Results and Ethics

Q2 continued. Self-Forcing

Early · chunks 1–27

Metric ↑BaseOPSD-VΔ95% CI
Aesthetic quality0.57830.5804+0.00209[−0.00207, +0.00639]
Imaging quality0.71810.7189+0.00076[−0.00315, +0.00456]
Subject consistency0.93010.9240−0.00611[−0.00905, −0.00314]
Background consistency0.93720.9334−0.00377[−0.00586, −0.00167]
Motion smoothness0.98620.9852−0.00098[−0.00153, −0.00044]
Dynamic Degree0.63640.6513+0.01491[−0.01392, +0.04391]
Mean excluding Dynamic Degree0.83000.8284−0.00160[−0.00326, +0.00013]

Middle · chunks 28–54

Metric ↑BaseOPSD-VΔ95% CI
Aesthetic quality0.55050.5637+0.01311[+0.00705, +0.01927]
Imaging quality0.70320.7124+0.00916[+0.00268, +0.01577]
Subject consistency0.92950.9256−0.00388[−0.00786, +0.00023]
Background consistency0.93670.9344−0.00237[−0.00507, +0.00031]
Motion smoothness0.98540.9848−0.00065[−0.00134, +0.00007]
Dynamic Degree0.63970.6413+0.00170[−0.03551, +0.03926]
Mean excluding Dynamic Degree0.82110.8242+0.00307[+0.00041, +0.00585]

Late · chunks 55–81

Metric ↑BaseOPSD-VΔ95% CI
Aesthetic quality0.53280.5566+0.02380[+0.01574, +0.03219]
Imaging quality0.69880.7104+0.01153[+0.00279, +0.02054]
Subject consistency0.92830.9229−0.00541[−0.01052, −0.00052]
Background consistency0.93200.9335+0.00148[−0.00118, +0.00411]
Motion smoothness0.98490.9843−0.00064[−0.00166, +0.00031]
Dynamic Degree0.64070.6633+0.02264[−0.01784, +0.06211]
Mean excluding Dynamic Degree0.81540.8215+0.00615[+0.00295, +0.00950]

Final 21 · chunks 61–81

Metric ↑BaseOPSD-VΔ95% CI
Aesthetic quality0.53180.5561+0.02422[+0.01569, +0.03291]
Imaging quality0.69950.7103+0.01073[+0.00121, +0.02050]
Subject consistency0.93280.9277−0.00504[−0.01058, −0.00001]
Background consistency0.93480.9349+0.00014[−0.00285, +0.00304]
Motion smoothness0.98500.9843−0.00072[−0.00178, +0.00028]
Dynamic Degree0.62680.6463+0.01945[−0.02356, +0.06268]
Mean excluding Dynamic Degree0.81680.8226+0.00587[+0.00238, +0.00937]

Δ is OPSD-V minus Base. We compute the 95% confidence intervals by resampling prompt pairs within each benchmark and weighting the two benchmarks equally. The mean excluding Dynamic Degree gives equal weight to the other five scores.

Quality and motion by segment


What are the data-use and user-study ethics procedures?

All 3,800 training clips were sourced from YouTube and are licensed under Creative Commons Attribution (CC BY).

We did not seek institutional ethics review before conducting the preference study. All 10 participants consented to participate, and we report only aggregate results.

Data provenance and participant consent

Reviewer_7yiMOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors present OPSD-V, a post-training method designed to reduce error accumulation over time in Self-Forcing models. By leveraging privileged information from long videos in the training dataset, OPSD-V aims to mitigate quality degradation during long-horizon generation. The proposed method does not require any changes to the base model architecture, few-step sampler, or inference-time caching mechanism.

Critical strengths
  • OPSD-V is quite practical post-training methodology for few-step video models. Base few step models are distilled from 5-step teacher generator during Self-Forcing training, hence it can only model short duration of dynamics. With the recent advances in Self-Forcing variants, now these generators can generate hours of videos, though with a repetitive content. OPSD-V uses real long video dataset to supervise both error accumulation over time, and it also teaches the base model long range dynamics. Having a simple training recipe makes the contribution more accessible in the field.
Critical weaknesses
  • In Section 4.3, the teacher replaces older generated chunks with real-video chunks while keeping the most recent student-generated chunk. However, the student's generated video may follow a different motion or semantic trajectory from the real video. This could create inconsistencies in the teacher's temporal context, potentially leading to misleading supervision. Although the authors acknowledge this potential mismatch, they do not investigate its impact or provide ablations on how much real-video context should be used.
Questions for authors

How does varying the proportion of real-video and student-generated chunks in the teacher cache affect the model's performance? Is there an optimal balance between the two, and how sensitive is OPSD-V to this choice?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 7yiM — Real-History Replacement and Teacher–Student Alignment

We invite you to view the videos and tables on our anonymous rebuttal page.

How sensitive is OPSD-V to the proportion of real history?

We thank the reviewer for raising this question. The teacher’s attention window spans 12 latent frames divided into four three-frame chunks: the fixed first-chunk sink, one older history chunk, the most recently completed chunk, and the chunk currently being denoised. To evaluate how replacing generated history with time-aligned ground-truth (GT) KV affects performance, we tested two alternative policies alongside our default design:

  • 100% GT (Ours): Always replaces the older history chunk with time-aligned GT KV whenever available, while retaining the teacher’s own latest generated chunk.
  • 50% GT: Replaces the older history chunk with probability 0.5 when available, while still retaining the teacher’s latest generated chunk.
  • All-real: Overwrites every generated chunk’s KV with GT KV immediately upon completion, so neither history slot holds generated content.

We evaluated all three policies on LongLive. Our default 100% GT policy achieved the highest Imaging quality over the final 21 chunks (0.7080). Lowering the replacement probability to 50% reduced the score to 0.6994, likely because the older slot then holds generated history half the time, which lets errors accumulate in the teacher’s cache. All-real scored the lowest at 0.6193. Replacing the teacher’s latest generated chunk breaks continuity at chunk boundaries and makes the teacher’s targets less reliable.

This boundary artifact is visible in a training-free diagnostic that applies All-real to the original LongLive model (Appendix C.5, Fig. 7). In the motorcyclist example, the next chunk attends to real frames that no longer match the preceding generated frames, so the rider’s pose and the background change abruptly at the chunk boundary.

On LongLive, these results favor refreshing the older history slot with time-aligned GT KV whenever available while retaining the teacher’s latest generated chunk. We have not yet tested probabilities between 50% and 100%. The next section explains how the teacher’s predictions stay aligned with the student’s trajectory during training and degrade less over long rollouts.

LongLive Imaging quality

Metric100% GT (Ours)50% GTAll-real
Final-21-chunk Imaging quality ↑0.70800.69940.6193

How are the teacher and student aligned?

The teacher and student share the same real first chunk as their attention sink, so both branches start from the same scene. Each branch also retains its own most recently generated chunk in its KV cache, which preserves continuity between consecutive chunks.

OPSD-V keeps supervision on-policy by evaluating both branches on the same noisy student latent at the same denoising timestep:

  • Step 1: Both branches receive the same initial Gaussian noise for the current chunk.
  • Steps 2–4: Both branches receive the noisy latent obtained by applying the sampler transition to the student’s clean prediction from the preceding step.

The teacher evaluates this shared latent using its own KV cache (with the older slot refreshed from real video), while the student alone determines the denoising trajectory of the current chunk.

The diagnostic on the linked page reproduces these training-time inputs using the original LongLive weights. Across these examples, the student and teacher predictions follow the same scene progression, while the teacher preserves sharper details and shows fewer artifacts in later chunks.

Matched videos and cache comparison

Reviewer_qc5bOpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a method to address the quality degradation problem in long-video rollout. The key idea is a on-policy rollout similar to self-forcing style while performing self distillation by using the same model's EMA version. The key contribution is the use of clean history for teacher models to provide better supervision during rollout. Empirical results show the effectiveness of the proposed method.

Critical strengths
  1. the major contribution is the clean history used by the teacher model for better supervision, the model achieved improved results after self-distillation
  2. the proposed method introduced minimal overhead over the original training schema of baseline methods such as self-forcing or longlive
Critical weaknesses
  1. the clean history is not the ground truth for student rollouts, therefore it creates a gap between student and teacher rollout. e.g. in the demos such as "further baseline comparison" in the attached website, the videos are less consistent compared to baseline methods which is potentially the result of this gap.
  2. I think the key reason behind this improvement is not studied. E.g, the clean history can be thought of as more attention sinks. How does this affect the velocity of teacher models. Why cannot we use a native AR teacher model introduced in causal-forcing to achieve better results?
  3. Similar baseline methods are not included. There are various baseline methods beyond LongLive that achieve strong results beyond 30s or 60s.
Questions for authors
  1. The proposed method is more empirical than trying to systematically solve the problem. The underlying mechanism is not well studied with missing baseline long rollout methods and ablation studies on other teacher design choices.
  2. How about the results with prompt switching?
5 comments
AuthorsOpenReview ↗

Response to Reviewer qc5b (1/3) — Visual Discontinuity, Attention Sinks, and Teacher Velocity

We invite you to view the videos and tables on our anonymous rebuttal page.

Where does the discontinuity come from?

The reviewer points to the “further baseline comparison” demos on the attached website as examples of reduced consistency. However, the original Self-Forcing model already shows this discontinuity. The cited OPSD-V model was post-trained from it, and with the same prompt and random seed, Self-Forcing produces the same abrupt shift in viewpoint and rooftop alignment.

Similar discontinuities appear in long rollouts of other Wan2.1-1.3B backbones, such as Causal Forcing and Reward Forcing. One possible cause is that the rollout runs past the events described in the prompt, so the later content is underspecified. This base-model behavior should be checked before the artifact is attributed to OPSD-V. The linked comparisons show the base and OPSD-V versions of other backbones on the same prompt, alongside paired examples on additional prompts.

The reviewer also raises a valid concern that the teacher’s and student’s rollouts may become misaligned. Such a mismatch produces a different kind of discontinuity, which we analyze with a training-free All-real diagnostic in Appendix C.5 (Fig. 7): replacing the teacher’s latest generated chunk with ground-truth (GT) KV introduces a visible jump at the chunk boundary. OPSD-V avoids this by retaining the teacher’s own latest generated chunk and evaluating the teacher at the student’s denoising states throughout training. In our teacher velocity analysis, the decoded teacher and student predictions follow the same scene progression.

All-real comparison · Teacher velocity analysis and decoded predictions

What role does real history play in OPSD-V’s improvements?

Attention sink

The review does not define “more attention sinks” or state a testable prediction. We address the two readings we consider most likely: whether the replaced history slot acts as a sink, and whether OPSD-V uses more sinks at inference.

What is an attention sink? StreamingLLM demonstrated that initial tokens can attract disproportionately high attention in pretrained language models. LongLive showed that after streaming long tuning, retaining the first chunk as an attention sink improves long-range consistency. Subsequent chunks attend to this first chunk throughout the rollout; because it is generated before autoregressive errors accumulate and its cached content remains fixed during inference, it serves as a stable anchor for long video generation. Rolling Forcing similarly retains the KV of the first chunk (three latent frames) as a persistent anchor, while Reward Forcing extends this with EMA-Sink, which maintains a dedicated sink updated via exponential moving average. These video sinks add no parameters. The sink effect is an attention bias that the network learns on its own during rollout training, because the sink slot is present at every step. (StreamingLLM, LongLive, Rolling Forcing, Reward Forcing)

In contrast, OPSD-V adds no cache slot. During training, it overwrites the older rolling-history slot in the teacher’s KV cache with time-aligned GT KV. No network learns an attention bias toward this slot: the teacher, an EMA copy of the student, only provides targets, and the student’s history slots hold only its own generated chunks. The replaced slot thus remains ordinary history whose content advances with the rollout. Moreover, at inference, the base models and their OPSD-V counterparts on LongLive and Self-Forcing all use the same cache with a single sink (the first chunk). Multiple sinks therefore play no role in our results.

How real history shapes the teacher’s predictions

We initialized both branches with the original LongLive weights and queried them at the same student denoising states. The teacher’s cache held real history, while the student continued its own rollout. Across 50 training videos sampled proportionally by data source, the mean student–teacher velocity MSE was higher after 40 seconds than during the first 20 seconds at each of the four denoising steps. In the linked video examples, both decoded predictions follow the same scene progression, while the teacher preserves sharper visual details as the student’s rollout degrades. At the states the student actually visits, the teacher thus provides a sharper target.

Velocity plots and decoded videos

AuthorsOpenReview ↗

Response to Reviewer qc5b (2/3) — Causal Forcing’s Native AR Teacher

How does a native autoregressive teacher compare?

The reviewer’s proposed substitution conflates on-policy self-distillation (OPSD) with on-policy distillation (OPD) from an external teacher. Specifically, OPSD-V matches the velocity predictions of an EMA copy of the same four-step model at the student’s four fixed denoising states. By contrast, Causal Forcing’s teacher-forcing checkpoint is a multi-step AR diffusion model. It can predict a velocity at each of these states, but matching velocities pointwise does not reproduce the transitions of its multi-step sampler. Causal Forcing itself reports abrupt transitions when this checkpoint is sampled directly in four steps, requiring causal ODE distillation to adapt it to few-step generation (Causal Forcing, Appendix C.2 and Fig. 7).

Distilling a separate multi-step teacher is a different problem. For instance, π\pi-Flow uses intermediate integration substeps to imitate the teacher’s trajectory, and DiffusionOPD adds an explicit transition-matching objective along the student’s rollout (π-Flow · DiffusionOPD).

We nevertheless tested the suggestion by training a student with Causal Forcing’s multi-step teacher-forcing checkpoint as the teacher. Both the student and teacher shared the same real first chunk. Following OPSD-V, we queried the teacher at the student’s four denoising states and matched their velocity predictions via MSE. Under matched student inference settings, the resulting student degraded severely in both quantitative metrics and visual quality. On our benchmarks, we evaluated frame quality using MUSIQ (the no-reference metric underlying VBench’s Imaging quality score). Over the first 5 seconds, the student trained with the multi-step teacher scored 0.2471, compared to 0.6979 for OPSD-V. Over the final 21 chunks, it scored 0.1867, compared to 0.6734 for OPSD-V. Because this substitution changes the teacher’s architecture, guidance, and history construction at once, the result does not isolate any single factor.

Teacher comparison and videos

Beyond the choice of teacher model, we also ablated two alternative teacher-cache designs: 50% GT, which uses GT KV for the older history chunk with probability 0.5, and All-real, which also uses GT KV for the latest completed chunk. Both ablations are detailed in our responses to Reviewers 7yiM and cGMm.

50% GT and All-real comparison · All-real full metrics

AuthorsOpenReview ↗

Response to Reviewer qc5b (3/3) — Additional Backbones and Prompt Switching

How does OPSD-V perform on additional long-video baselines?

The review requests additional long-video baselines but does not name a method or the claim a comparison should test.

OPSD-V is a post-training stage for existing few-step causal generators, so we measure its effect by comparing each generator before and after post-training. We selected Self-Forcing as a standard few-step baseline and LongLive as a generator already trained on long rollouts. Within each backbone, we compared OPSD-V with its base checkpoint, flow-matching supervised fine-tuning (SFT), and Astrolabe reinforcement learning under matched inference settings.

Interpreting the reviewer’s request as a question of generalization to additional DMD-based causal generators, we further evaluated OPSD-V on Causal Forcing and Reward Forcing. Each pair uses identical prompts, random seeds, four-step schedules, and cache configurations. With OPSD-V, Causal Forcing improves across all three metrics, while Reward Forcing improves in Quality and Dynamic Degree with a slightly lower Semantic score (Causal Forcing · Reward Forcing):

ModelQualityDynamic DegreeSemantic
Causal Forcing0.81470.92150.4872
Causal Forcing + OPSD-V0.81800.93850.4969
Reward Forcing0.79400.68750.5110
Reward Forcing + OPSD-V0.80250.69880.5093

All scores are reported on a 0–1 scale. We evaluate text-only videos generated with seed 1 across 120 matched prompts each from MovieGenBench and MeiBench, weighting both benchmarks equally.

Evaluation details and matched videos

What happens when the prompt changes during generation?

We thank the reviewer for asking about dynamic prompt switching. LongLive natively supports switching prompts mid-generation. With the same prompt sequences and switching schedules, LongLive + OPSD-V retains this capability. The examples below show our results, with more dynamic motion than LongLive.

We use the official LongLive v1.0 prompt sequences and match all inference settings and random seeds within each video pair (LongLive interactive generation).

Prompt-switching videos

Comments on the paper

4 comments
AuthorsOpenReview ↗

Response to the Area Chair, Reviewers, and Readers (1/2)

Dear Area Chair, Reviewers, and Readers,

Our submission received initial ratings of 3, 3, and 1. We contest Reviewer qc5b’s rating of 1 (clear rejection), and are sharing our objection publicly so that the Area Chair, the reviewers, and other readers can examine the evidence and join the discussion. Reviewer qc5b reported a confidence of 4 and rated presentation, informativeness, and soundness as 3, 2, and 2, respectively; the soundness rating of 2 reads “I have minor or fixable concerns about the paper’s soundness.” The reviewer also acknowledged that our method improves performance with minimal additional training overhead. Nothing in the review identifies a technical flaw warranting a clear rejection; instead, several criticisms reflect limited familiarity with standard methods and concepts in long-video generation. Full results and side-by-side video comparisons are available in our individual responses and on the rebuttal page.

We thank Reviewer qc5b for two constructive questions regarding how real-video history affects the teacher’s velocity predictions and how OPSD-V behaves when the prompt changes during generation. We conducted additional experiments to address both points. For experiment details, see the teacher velocity analysis and the prompt-switching experiments.

The remaining criticisms do not justify a clear rejection

“More empirical than trying to systematically solve the problem.” Figure 1 presents a controlled, training-free cache intervention. Holding the original LongLive weights, the initial real chunk, and the sampler fixed, we replace only the KV entries corresponding to older history. This single intervention visibly improves continuation quality and directly motivates our teacher-cache design. The review identifies neither a technical error in our formulation nor a missing control that would overturn any central claim. Dismissing a method as “empirical” on this basis substitutes subjective judgment for technical evaluation—an irresponsible basis for a rating of 1. What exactly does the reviewer mean by “systematically solve the problem”? For experiment details, see the training-free cache intervention in Figure 1.

The cited visual discontinuity. The OPSD-V model shown in the “further baseline comparison” demo was post-trained from the original Self-Forcing checkpoint. With the same prompt and random seed, the original Self-Forcing model exhibits the same abrupt shift in viewpoint and rooftop alignment, meaning this discontinuity already exists in the base model. Similar jumps also appear in Causal Forcing and Reward Forcing. One possible explanation is that the prompt underspecifies the later portion of a long rollout. A reviewer familiar with long-video generation should recognize this base-model behavior and verify the matched baseline before attributing the artifact to OPSD-V. For experiment details, see the matched video comparisons for the cited discontinuity.

Teacher–student history mismatch is indeed worth isolating, which is what Appendix C.5 (Figure 7) and the “Replacing the Latest History Chunk” section of our demo do. In the training-free All-real diagnostic, overwriting the latest generated chunk’s KV causes a distinct discontinuity in the rider’s pose and background at the chunk boundary. This observation is why OPSD-V retains the teacher’s own latest generated chunk. For experiment details, see the training-free latest-chunk replacement comparison.

AuthorsOpenReview ↗

Response to the Area Chair, Reviewers, and Readers (2/2)

“More attention sinks.” The review never defines “sink” here and offers no falsifiable prediction, forcing us to guess what hypothesis is being proposed. Raising a mechanism-level objection this vaguely is irresponsible. OPSD-V overwrites an existing rolling history slot with time-aligned real-video KV, so the content of this slot advances with the rollout. No network learns an attention bias toward it: the teacher, an EMA copy of the student, only provides targets, and the student’s history slots hold only its own generated chunks. By contrast, a LongLive-style attention sink relies on a learned attention bias toward a cache slot whose contents remain fixed during inference. Moreover, during text-only inference, the base models and their OPSD-V counterparts use the same single-sink KV-cache configuration. We suspect that the ambiguity in the reviewer’s question stems from a misunderstanding of what an attention sink is in this setting. For experiment details, see the attention-sink configuration and teacher velocity analysis.

Causal Forcing’s native AR teacher. Treating an external multi-step teacher as a drop-in replacement for the EMA teacher overlooks the distinction between on-policy self-distillation (OPSD) and on-policy distillation (OPD) from an external teacher, suggesting the reviewer may be unfamiliar with this distinction. Moreover, the review never specifies how a multi-step teacher is supposed to supervise a four-step student, even though Causal Forcing itself reports abrupt transitions when its teacher checkpoint is sampled directly in four steps. Even so, we ran the suggested experiment using direct velocity matching at the four denoising states visited by the student. Under matched inference settings, the student trained with the external teacher scored 0.2471 on MUSIQ over the first 5 seconds (compared with 0.6979 for OPSD-V), and 0.1867 over the final 21 chunks (compared with 0.6734 for OPSD-V). For experiment details, see the native autoregressive teacher comparison.

Missing long-video baselines. Citing “missing baseline long rollout methods” as grounds for rejection without naming a single method or identifying what hypothesis the comparison should test provides no actionable guidance and is irresponsible. OPSD-V is a post-training method; accordingly, within each backbone, we compare against the base model, SFT, and Astrolabe RL under matched inference settings. Given the lack of any named baseline, we also evaluated whether OPSD-V transfers to Causal Forcing and Reward Forcing by testing each backbone before and after post-training. Adding OPSD-V to Causal Forcing improves all three reported metrics. On Reward Forcing, Quality increases from 0.7940 to 0.8025 and Dynamic Degree increases from 0.6875 to 0.6988, while Semantic decreases slightly from 0.5110 to 0.5093. For experiment details, see the Causal Forcing and Reward Forcing comparisons.

Request to the Area Chair and Reviewers

We ask the Area Chair and Reviewers 7yiM and cGMm to judge whether the technical evidence supports a score of 1. Reviewers 7yiM and cGMm identified concrete gaps and proposed testable controls, which we addressed through cache ablations and segment-level evaluations, and we thank them for their rigor. We are releasing our code, demos, and model weights, and anonymous links to all three are already provided in our submission. Using real-video history to construct teacher targets while preserving few-step inference opens a practical direction for post-training long-video generators. For experiment details, see the real-history proportion ablation, the All-real comparison, and the segment-level evaluation.

AuthorsOpenReview ↗

A Public Request to Reviewer qc5b: Which Technical Flaw Justifies a Clear Rejection?

We ask Reviewer qc5b to answer our questions publicly and specifically: Why list “missing baseline long rollout methods” as grounds for rejection without naming a single baseline to compare against? What does “more attention sinks” actually mean here, and what empirical result would confirm or falsify it? Above all, which concrete technical flaw justifies a clear rejection? If the reviewer cannot answer these questions, a score of 1 should not stand.

Frankly, the questions raised in the review do not read like those of someone familiar with this domain—least of all a reviewer claiming a confidence of 4 (“I have substantial direct research experience on this topic”). Even setting the technical misunderstandings aside, the review levels its criticisms in an ambiguous, highly subjective, and irresponsible manner. In both technical substance and professional rigor, this review falls well short of the standard expected at ICLR.

We have made our claims easy to verify. Our code, model checkpoints, and detailed demo videos are publicly available, and the comparisons in our rebuttal use matched prompts, random seeds, and inference settings. We stand behind our core idea, our method, and our results, and we welcome anyone to put them to the test.

Peer review works only when reviewers are held accountable for their scores. A reviewer who is unwilling to engage with the evidence should decline the assignment rather than hand out an unsubstantiated rejection.

Area_Chair_9iDXOpenReview ↗

reviewer discussion

Dear Reviewers,

The authors have submitted their rebuttal. Please read their response carefully and engage in the discussion as soon as possible. Let us know whether the rebuttal addresses your concerns and update your final rating accordingly.

Thank you for your time and service to ICLR.

Best regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.