acceptodds
Under review as a conference paper at ICLR 2027

History Dropout for Long-Horizon Autoregressive Video Generation

Abstract

Autoregressive video generators use previously generated content to maintain continuity, but this history can also carry errors into later predictions. We introduce History Dropout (), a history regularization method for long-horizon video generation. During training, varies historical value masks across denoising steps while retaining all keys for attention normalization. Inverse-retention scaling preserves the expected attention output for fixed, mask-independent inputs. At inference, each attention head independently samples historical key-value tokens, with sink and current-chunk tokens always retained. The two operations serve different purposes: training perturbs historical contributions without changing the softmax denominator, whereas inference uses shorter attention inputs. Applied to Self Forcing, Causal Forcing, and LongLive, improves visual quality and dynamics at both 60 and 240 seconds, while inference latency decreases from 0.72 to 0.64 seconds on a single NVIDIA H20 GPU. Please refer to https://anonymous.4open.science/w/hidrop-supplementary-2E19/ this link for our supplementary video.

Primary area:
generative models
Keywords:
Video Generation

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_M6T6OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces History Dropout (HiDrop), a history regularization method for long-horizon autoregressive video generation. During training, it applies step-dependent dropout to historical attention values while retaining all keys for softmax normalization and scaling surviving historical values by the inverse retention probability. During inference, it instead independently samples historical key-value tokens per attention head and attention call, while retaining sink and current-chunk tokens. The method leaves the underlying rollout objective unchanged. Experiments on Self Forcing, Causal Forcing, and LongLive report improvements in aesthetic quality, imaging quality, and dynamic degree at 60 and 240 seconds.

Critical strengths
  1. A small, practical intervention. There are no extra learned parameters or auxiliary losses, making the method relatively easy to integrate with existing rollout training.

  2. Improvements across multiple baselines. Table 1 reports improvements in aesthetic quality, imaging quality, and dynamic degree across all three evaluated baselines.

Critical weaknesses
  1. The training-inference mismatch requires stronger justification and controls. Training uses historical value dropout with full-key normalization and inverse-retention scaling, whereas inference removes both keys and values, renormalizes over the selected keys, and uses independently sampled per-head masks. These operations differ in both attention normalization and mask sharing. The authors do not compare against using the same per-head KV dropout during both training and inference under matched settings. Thus it remains unclear whether the asymmetric design is preferable to a simpler design that uses the same selection rule at both stages. The authors need to provide stronger justification for this design.

  2. There is no direct quantitative evaluation of drift over time. The paper's motivation concerns error propagation through generated history, but the quantitative experiments primarily report holistic VBench scores for long videos. These results do not directly establish that quality or identity degrades more slowly within a rollout, or that historical errors propagate less strongly. The drifting metric used in Longlive and Rolling Forcing can be used.

Questions for authors
  1. Could the authors compare the proposed value-dropout training/KV-selection inference design against the same per-head KV selection used during both training and inference, with matched retention rates, inference settings, and training budgets? Does the asymmetric design offer an advantage, and what explains that advantage?

  2. Could the authors provide quantitative drift measurements on the same prompt rollouts, for example by tracking quality, identity, color, and motion across consecutive temporal windows? Alternatively, could controlled perturbations of generated history demonstrate that HiDrop reduces the error accumulation?

4 comments

We appreciate the reviewer’s insightful questions and provide our responses below.

Q1: Compare HiDrop with per-head KV dropout used in both training and inference, and explain the benefit of using different operators at the two stages.

We conduct an ablation on different configurations, showcasing that: (1) Using per-head KV dropout during both training and inference (Config. A) yields lower imaging quality and dynamics than HiDrop at both 60s and 240s; (2) Our asymmetric design balances generation quality, training time, and inference efficiency. Please refer to Appendix C.6 (Table 8 and Figure 8) in the revised manuscript for the quantitative and qualitative comparisons.

Table 8. Training configurations and results on Self Forcing. Mask sampling and value dropout refer to training; all configurations use per-head KV selection at inference. Value dropout includes inverse-retention scaling. Gen. step denotes time per generator training step.

HorizonConfig.Mask samplingValue dropoutAQ ↑BC ↑DD ↑IQ ↑MS ↑SC ↑TF ↑Gen. step (s) ↓
60sAPer-head×0.5840.9530.6570.6970.975‾\underline{0.975}0.962‾\underline{0.962}0.982‾\underline{0.982}131‾\underline{131}
60sBPer-head✓0.594‾\underline{0.594}0.9530.784‾\underline{0.784}0.7090.9790.9660.985132
60sOursPer-step✓0.5950.9530.7850.708‾\underline{0.708}0.9790.9660.985121
240sAPer-head×0.548‾\underline{0.548}0.9520.6750.7070.973‾\underline{0.973}0.961‾\underline{0.961}0.974‾\underline{0.974}131‾\underline{131}
240sBPer-head✓0.5600.958‾\underline{0.958}0.7920.711‾\underline{0.711}0.9790.9650.979132
240sOursPer-step✓0.5600.9590.790‾\underline{0.790}0.7140.9790.9650.979121

The following comparisons explain our training and inference choices.

  1. Training operator: A vs. B and Ours. With per-head training masks, B improves visual quality and dynamics relative to A, and achieves scores comparable to Ours. This combining Table 2 supports value dropout as the training operator.

  2. Training efficiency: B vs. Ours. Compared with B, Per-step masks reduce generator training-step time from 132s to 121s, an 8.3% reduction, with comparable generation quality. This is because independent per-head masks increase the complexity of the Self Forcing training graph. We therefore use per-step masks during training.

  3. Inference efficiency. KV selection shortens attention inputs, reducing steady-state chunk latency from 0.72s to 0.64s (Table 1). We observe no clear improvement in generation quality from value dropout at inference, while chunk latency increases from 0.64s to 0.72s. We therefore retain per-head KV selection at inference while using value dropout during training.

Changes in the paper: Appendix C.6 and Table 8 compare the three configurations and report generator training-step time. Figure 8 presents the corresponding qualitative results.

Q2: Can quantitative drift measurements on matched rollouts, or controlled perturbations of generated history, demonstrate that HiDrop reduces error accumulation?

HiDrop reduces mean appearance and color drift and better preserves motion dynamics across all three models at 60s and 240s. It also improves the model's ability to continue generation from imperfect history. Please refer to Appendix C.5 (Figures 6–7) and Appendix C.7 (Table 9) in the revised manuscript for the temporal-drift and shared-history continuation experiments.

  1. Temporal drift. (1) Settings: We evaluate Self Forcing, Causal Forcing, and LongLive using 326 prompts for 60s videos (VBench-Long) and 128 prompts for 240s videos (MovieGen). Each rollout is divided into non-overlapping 5s windows. AQ and IQ drops are measured relative to the first window; subject and background drift use DINO and CLIP similarity to the first frame; color drift compares normalized hue histograms with the first window. DD change tracks the proportion of dynamic clips relative to the initial window. The curves include pointwise 95% confidence intervals from 2,000 bootstrap resamples of matched baseline–HiDrop video pairs. (2) Results: Averaged over subsequent windows of the 240s videos, HiDrop reduces subject drift by 26–40% and color drift by 26–29%, with smaller AQ drops across all three models. It also better preserves dynamics: the mean change in dynamic proportion from the initial window improves from −19.78 to −8.49 percentage points for Causal Forcing and from −12.27 to +1.30 percentage points for LongLive. Please refer to Figures 6 and 7 in the revised manuscript for the full temporal curves.

  2. Continuation from imperfect history. To assess robustness to history perturbations, we use a 10s baseline-generated prefix as a proxy for perturbed ground-truth history. The baseline and HiDrop then each continue generation to 60s from the same historical KV cache and subsequent noise state. This compares their continuation quality under identical imperfect conditioning. VBench evaluates the complete videos, including the shared prefix. HiDrop improves AQ, IQ, and DD across all three backbones (Appendix C.7, Table 9).

Table 9. 60s generation from a shared baseline-generated prefix.

BackboneContinuationAQ ↑BC ↑DD ↑IQ ↑MS ↑SC ↑TF ↑
Self ForcingBaseline0.5820.9530.6930.7030.9790.9650.981
Self ForcingBaseline + Ours0.5980.9540.7650.7060.9800.9660.982
Causal ForcingBaseline0.5620.9580.4510.7020.9860.9640.982
Causal ForcingBaseline + Ours0.6020.9600.5530.7090.9860.9700.983
LongLiveBaseline0.5730.9510.7090.6940.9750.9600.982
LongLiveBaseline + Ours0.5950.9580.7730.6980.9780.9680.982

Changes in the paper: Appendix C.5 and Figures 6–7 report temporal drift. Appendix C.7 and Table 9 report continuation from a shared baseline-generated history.

Thank the authors for the detailed rebuttal and the additional experiments. The new ablation studies provide a clearer justification for the asymmetric training-inference design, showing the trade-offs between generation quality and computational efficiency. The temporal drift analysis and shared-history continuation experiments provide more evidence supporting the effectiveness of HiDrop in mitigating error accumulation. I believe my major concerns have been addressed, and I will raise my score accordingly.

AuthorsOpenReview ↗

Thank you for your constructive feedback

Dear Reviewer M6T6,

Thank you for carefully considering our rebuttal and additional experiments, and for updating your assessment. We are glad that these additions have addressed your major concerns. Your constructive suggestions helped us clarify the training–inference design and strengthen our evaluation of temporal drift and error accumulation.

We sincerely appreciate your time, thoughtful feedback, and support.

Best regards, The Authors

Reviewer_X2oNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper propose a history dropout for consistent streaming video generation. They drop Value tokens of attention during training and Key and Value during inference per attention head during inference. While the methods seems to be incremental and the dropout is a rather naive idea, the demostration shown in website seems to be consistent.

Critical strengths
  • The videos on the website mostly feature consistent camera work and stable characters, with fewer artifacts.
  • The method seems to have good compatibility with different types of streaming models.
Critical weaknesses
  • Dropping out history tokens is not a novel idea at all. What's more important, I don't see the connections between ‘The history may carry errors for later generation’ and ‘random dropout history’. Randomly dropping out history may also decrease correct information in history.
  • The evaluation results show a very subtle improvement.
Questions for authors
  • I doubt the visual result is more static, which is not a good sign. And the performance gain may come mainly from the static collapse. I wonder if the authors could provides results on multi-prompt videos.
  • Could the authors please clarify the core innovation between their method and previous methods e.g.History-Guided Video Diffusion? Or, more specifically, could you also help us understand the intuition or the observation that could lead to the methods in the paper? I would like to say a more complete analysis of the method.
6 comments

We appreciate the reviewer’s insightful questions and provide our responses below.

W1 & Q2(b) Clearer analysis on motivation and innovation of HiDrop.

(a) Preliminary findings (motivation of HiDrop)

  1. Observation: Errors persist despite self-rollout training. Autoregressive generation exhibits persistent subject distortion, texture degradation, and appearance drift over long rollouts, even with self-rollout training (Figure 1; Appendix C.5).

  2. Intuition: Dropout may reduce overreliance on imperfect history. A video generation model is capable of generating high-quality frames without history, but may still produce poorer frames when conditioned on imperfect history, allowing errors to accumulate. This suggests that models may overfit to historical information during self-rollout training. Inspired by dropout’s role in reducing overfitting, we explore whether history dropout can regularize this dependence and improve generation from imperfect history.

  3. Experimental development: Step-dependent masks and value rescaling make dropout effective. We first tested a naive dropout strategy by removing a fixed subset of historical KV tokens; quality and dynamics declined (Table 2). We then combined step-dependent masks with rescaled value dropout, improving quality and dynamics (Table 2). Varying masks lets omitted information return; retaining all keys and rescaling surviving values avoids the normalization bias of KV removal and preserves expected attention output for fixed, mask-independent inputs. Finally, selecting fewer historical KV tokens enables faster inference (Tables 1–2).

(b) Contribution (innovation of HiDrop)

Our contribution has two parts:

  • Regularizing self-forcing frameworks by token-level dropout. To our knowledge, HiDrop is the first to regularize self-forcing frameworks by applying token-level dropout to the attention values of self-generated history. HiDrop is simple to implement and readily integrates into existing self-rollout training frameworks, improving visual quality and motion dynamics (Table 1).
  • Making history dropout effective for training. Naively removing historical KV tokens during training reduces visual quality and dynamics. We introduce step-dependent mask sampling to reduce persistent reliance on specific historical features and rescaled value dropout to stabilize training. Together, these choices enable dropout to improve visual quality and dynamics (Table 2).

Q2(a) — Comparison with HGV: How does HiDrop differ from History-Guided Video Diffusion?

(a) Theoretical difference

History-Guided Video Diffusion (HGV) is the most closely related work to ours. We summarize the key differences in the table below.

AspectHistory-Guided Video DiffusionHiDrop (Ours)
ObjectiveHistory guidanceHistory regularization
Training settingDiffusion ForcingSelf-rollout training
DesignFrame-level noise maskingToken-level value dropout; step-dependent masks; value rescaling
Effect on historyNoise-corrupted contentIntact cache; step-wise re-access

HGV combines predictions conditioned on differently noise-masked histories to guide sampling, while HiDrop regularizes self-rollout training through token-level value dropout. HiDrop keeps the KV cache intact and varies masks across denoising steps, so masked tokens can contribute again later.

(b) Experimental comparison

To further illustrate the effect of our method, we compare both methods on 60s and 240s generation. The HGV-style adaptation underperforms HiDrop, potentially because it obscures the full history during training, whereas HiDrop retains access to it.

VBench comparison on Self Forcing at 60s and 240s.

DurationMethodAQ ↑BC ↑DD ↑IQ ↑MS ↑SC ↑TF ↑
60sHGV-style history noise0.5710.9500.6740.6950.9720.9650.978
60sHiDrop (ours)0.5950.9530.7850.7080.9790.9660.985
240sHGV-style history noise0.5440.9560.7210.6980.9730.9620.964
240sHiDrop (ours)0.5600.9590.7900.7140.9790.9650.979

W2: The reported improvements appear modest.

HiDrop consistently improves visual quality, motion dynamics, and temporal stability across all three backbones. (1) Visual quality and dynamics: HiDrop achieves substantial gains in these two central aspects of long-video generation at both 60s and 240s—for example, +12.3% AQ and +20.3% DD for Causal Forcing at 240s—while other metrics remain comparable or improve slightly (Table 1). (2) Temporal stability: Our added temporal-drift evaluation tracks subject-appearance and color deviations over time, complementing aggregate video-level scores with an assessment of stability over long rollouts. HiDrop reduces mean subject appearance drift by 26–40% and color drift by 26–29% across all three backbones at 240s (Appendix C.5, Figures 6–7).

Key 240s results:

EvaluationBackboneKey resultEvidence
Temporal drift (new)All threeMean subject drift 26–40% lower; mean color drift 26–29% lowerAppendix C.5, Figures 6–7
Visual quality and dynamicsCausal ForcingAQ +12.3%; DD +20.3%Table 1
DynamicsSelf ForcingDD +10.8%Table 1
Human preferenceSelf Forcing / Causal Forcing / LongLive95% / 96% / 92% preference for HiDrop, respectivelySection 4.3, Table 3

Q1: Assess whether the quality gains arise from static collapse, and provide multi-prompt generation results.

(a) HiDrop consistantly preserves dynamics better than basealine.

  • HiDrop improves DD by 7.9–25.3% relative to the corresponding baselines across all three backbones at 60s and 240s (Table 1).
  • On 240s rollouts, time-averaged DD changes relative to the initial window improve from −9.7 to −3.9 percentage points for Self Forcing, −19.8 to −8.5 for Causal Forcing, and −12.3 to +1.3 for LongLive (Appendix C.5, Figures 6–7).

(b) Multi-prompt generation results.

HiDrop also improves visual quality and dynamics in 60s multi-prompt generation. We use LongLive's multi-prompt dataset and generation code, following its official training configuration. AQ increases from 0.579 to 0.618, IQ from 0.698 to 0.716, and DD from 0.472 to 0.628, together with improvements in background consistency, subject consistency, and motion smoothness. Please refer to Appendix C.8 (Table 10 and Figure 9) in the revised manuscript for evaluation results.

Multi-prompt video results are provided at the top of our anonymous supplementary material https://anonymous.4open.science/w/hidrop-supplementary-2E19/.

Table 10. VBench scores for 60s multi-prompt generation with LongLive.

MethodAQ ↑BC ↑DD ↑IQ ↑MS ↑SC ↑TF ↑
LongLive0.5790.9570.4720.6980.9800.9680.991
LongLive + Ours0.6180.9650.6280.7160.9860.9780.991

Changes in the paper: Appendix C.5 (Figures 6–7) report the temporal evolution of dynamic degree. Appendix C.8, Table 10, and Figure 9 present the multi-prompt evaluation and examples of responses to prompt changes.

Dear Reviewer X2oN,

Thank you again for your thoughtful review. As the discussion period is underway, we wanted to kindly follow up on our rebuttal and the new results added to the revised manuscript:

  • Multi-prompt generation & static collapse: We added the multi-prompt evaluation you requested (Appendix C.8 (Table 10 and Figure 9), and supplementary videos) along with temporal analyses of motion dynamics (Appendix C.5).
  • Comparison with HGV: We clarified the conceptual distinctions from History-Guided Video Diffusion and included an HGV-style history-noise comparison.

We would greatly appreciate it if you could let us know whether these additions address your main concerns, or if any points would benefit from further clarification. Thank you very much for your time and consideration.

Best regards, The Authors

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

This paper proposes History Dropout (HiDrop) for long-horizon autoregressive video generation. During training, HiDrop varies historical value masks across denoising steps while retaining all keys and applying inverse-retention scaling. At inference, it independently samples historical key-value tokens for each attention head while preserving sink and current-chunk tokens. The authors characterize the training operator’s local expectation property and evaluate the method on Self Forcing, Causal Forcing, and LongLive. The paper reports improved visual quality and dynamics at 60 and 240 seconds, favorable human preferences, and an 11.1% reduction in steady-state chunk latency on an H20 GPU.

Critical strengths
  • HiDrop addresses a practical limitation of autoregressive video generation through a simple modification that adds neither learned parameters nor auxiliary losses. Its compatibility with three existing training frameworks makes the approach potentially useful.

  • Table 1 reports consistent improvements in aesthetic quality, imaging quality, and dynamic degree across all three baselines at both long horizons. The human preference study provides complementary evidence, although its video sample is small.

  • The paper clearly distinguishes training-time value masking from inference-time KV selection. I appreciate the explicit limits placed on the expectation identity and the inclusion of the training/inference factorial comparison in Table 4.

Critical weaknesses
  • The specific design choices are not sufficiently isolated. The biggest concern for me is whether the experiments establish the advantage of the proposed masking schedule and normalization choices. Table 2 tests fixed versus varying masks under direct KV removal, whereas the final training method uses value dropout with full-key normalization. Therefore, this comparison does not establish that step-dependent masks are preferable under the actual training operator. Similarly, normalization and inverse-retention scaling are introduced together, and independent head sampling lacks a shared-mask control at the same token budget. Equation (7) correctly establishes a local expectation identity, but preserving the original attention output is not itself evidence of better generation, particularly when the history contains errors. Appendix C.1 acknowledges these missing controls. Comparisons that hold the value-dropout operator fixed while changing mask timing, and hold the inference budget fixed while changing head sharing, would clarify which choices provide the benefit over simpler variants.

  • Experimental reporting limits confidence in the improvements. Appendix B identifies the prompt sources but does not specify the numbers of evaluated prompts, videos per prompt, or random seeds. It also refers to baseline defaults without recording important settings such as model size/checkpoints, training duration, and numerical cache-window lengths. The frame sampling and aggregation procedure used to evaluate four-minute videos is unclear. These details matter when interpreting relatively small differences: in Table 4, adding training to inference selection changes 240s IQ from 0.712 to 0.714, although the DD improvement is larger, from 0.734 to 0.790. Please provide the concrete configurations and paired uncertainty estimates across prompts and generation seeds. For the human study, per-video preferences and uncertainty accounting for repeated ratings of the same 10 pairs would make the reported preference rates easier to assess.

  • The connection to reduced error propagation remains largely untested. The motivation concerns errors repeatedly carried through generated history, but the quantitative evaluation reports aggregate quality and consistency scores. Such improvements do not distinguish slower error accumulation from a general change in how strongly the model uses historical context. The different prompt sets at 60s and 240s also prevent a controlled comparison of degradation across those horizons. The authors explicitly acknowledge these limitations in Appendix B.3, so I do not interpret the paper as proving rollout-level error correction. Nevertheless, this leaves an important part of the motivation unsupported. Time-resolved quality and identity measurements on matched prompts, or a controlled experiment perturbing the supplied history and measuring subsequent degradation, would help establish when and why selective history access is useful.

  • The evaluation does not sufficiently address whether the generated motion follows the prompt. Higher dynamic degree and visual consistency do not establish that the requested action or camera movement is preserved. A concrete example is Figure 1(c), which presents “Stable Camera” as an improvement, while the corresponding supplementary prompt requests that the camera arc around the subject’s shoulders. This makes it important to distinguish unwanted drift from intended camera motion. The seven reported metrics and the human preference question about visual quality and dynamics do not separately resolve this issue. A focused evaluation of action and camera-instruction adherence would clarify whether reducing historical conditioning introduces a trade-off that the current scores miss.

Questions for authors

See weaknesses.

8 comments

We appreciate the reviewer’s careful assessment and provide our responses below.

W1: Isolate the effects of mask timing, normalization, value scaling, and head sharing.

We conduct an ablation study showing that: (1) Using per-head KV dropout during both training and inference (Config. A) yields lower imaging quality and dynamics than rescaled value-dropout training at both 60s and 240s; (2) Sharing training masks across heads retains comparable generation quality while reducing training time. Please refer to Appendix C.6 (Table 8 and Figure 8) in the revised manuscript for the quantitative and qualitative comparisons.

Table 8. Training configurations and results on Self Forcing. Mask sampling and value dropout refer to training; all configurations use per-head KV selection at inference. Gen. step denotes time per generator training step.

HorizonConfig.Mask samplingValue dropoutAQ ↑BC ↑DD ↑IQ ↑MS ↑SC ↑TF ↑Gen. step (s) ↓
60sAPer-head×0.5840.9530.6570.6970.975‾\underline{0.975}0.962‾\underline{0.962}0.982‾\underline{0.982}131‾\underline{131}
60sBPer-head✓0.594‾\underline{0.594}0.9530.784‾\underline{0.784}0.7090.9790.9660.985132
60sOursPer-step✓0.5950.9530.7850.708‾\underline{0.708}0.9790.9660.985121
240sAPer-head×0.548‾\underline{0.548}0.9520.6750.7070.973‾\underline{0.973}0.961‾\underline{0.961}0.974‾\underline{0.974}131‾\underline{131}
240sBPer-head✓0.5600.958‾\underline{0.958}0.7920.711‾\underline{0.711}0.9790.9650.979132
240sOursPer-step✓0.5600.9590.790‾\underline{0.790}0.7140.9790.9650.979121

Inference configurations and results on Self Forcing. Mask sampling refers to inference. Lat. denotes steady-state chunk latency.

HorizonConfig.Mask samplingAQ ↑BC ↑DD ↑IQ ↑MS ↑SC ↑TF ↑Lat. (s) ↓
60sCPer-step0.594‾\underline{0.594}0.9530.784‾\underline{0.784}0.7090.9790.9660.984‾\underline{0.984}0.64
60sOursPer-head0.5950.9530.7850.708‾\underline{0.708}0.9790.9660.9850.64
240sCPer-step0.559‾\underline{0.559}0.958‾\underline{0.958}0.789‾\underline{0.789}0.712‾\underline{0.712}0.9790.9650.9790.64
240sOursPer-head0.5600.9590.7900.7140.9790.9650.9790.64

The following comparisons clarify which design choices these results support.

  1. Training operator: A vs. B and Ours. With per-head training masks, B improves visual quality and dynamics relative to A, and achieves scores comparable to Ours. This combining Table 2 supports value dropout as the training operator.

  2. Training efficiency: B vs. Ours. Compared with B, Per-step masks reduce generator training-step time from 132s to 121s, an 8.3% reduction, with comparable generation quality. This is because independent per-head masks increase the complexity of the Self Forcing training graph. We therefore use per-step masks during training.

  3. Inference efficiency. KV selection shortens attention inputs, reducing steady-state chunk latency from 0.72s to 0.64s (Table 1). We observe no clear improvement in generation quality from value dropout at inference, while chunk latency increases from 0.64s to 0.72s. We therefore retain per-head KV selection at inference while using value dropout during training.

  4. Head sharing at inference: C vs. Ours. Shared masks (C) and independent per-head masks (Ours) yield comparable generation qualit with the same latency. Per-head sampling is simpler to integrate into the inference code, and we observe slight improvements in visual quality and dynamics in our experiments. This may be because, for a distilled model, moderately increasing the randomness of dropout masks helps improve generation quality. We therefore use per-head masks at inference.

Changes in the paper: Appendix C.6 and Table 8 compare the three configurations. Figure 8 presents the corresponding qualitative results.

W2: Provide concrete experimental settings and uncertainty estimates, including the sampling unit in the human study.

We summarize each requested detail below.

Reviewer questionKey details
Prompt sources and counts?326 VBench-Long prompts at 60s; 128 MovieGen prompts at 240s (Appendix C.5).
Videos per prompt?5 at 60s; 1 at 240s.
Generation seeds?Base seed 0; seeds jj at 60s (j=0,…,4j=0,\ldots,4).
Model sizes and checkpoints?Wan2.1-T2V-1.3B backbone.
Training duration?equal to 64 h800 GPU Hours.
Cache-window lengths?Default evaluation configuration: 12 latent frames = 3 sink + 6 FIFO + 3 current; chunk size 3; RoPE training range 21.
Other inference settings?Default keep ratio 0.5.
Frame sampling for four-minute videos?48 clips ×\times 5s/clip
Aggregation procedure?clip scores averaged per original video, then across valid videos; subject/background consistency combine within-clip and clip-to-clip scores; flicker uses static filtering.
Paired uncertainty across prompts?pointwise 95% CIs; 2,000 prompt-bootstrap resamples, preserving matched video pairs and full trajectories (Figures 6–7).
Human-study sampling unit and sample size?Video pair: 10 pairs per backbone; 30 pairs total; 20 evaluators (Section 4.3). SF: 95% [92%, 98%]; CF: 96% [93%, 99%]; LongLive: 92% [90%, 95%].

W3: Measure degradation within rollouts and continuation from imperfect history.

HiDrop reduces mean appearance and color drift and better preserves motion dynamics across all three models at 60s and 240s. It also improves the model's ability to continue generation from imperfect history. Please refer to Appendix C.5 (Figures 6–7) and Appendix C.7 (Table 9) in the revised manuscript for the temporal-drift and shared-history continuation experiments.

  1. Temporal drift. (1) Settings: We evaluate Self Forcing, Causal Forcing, and LongLive using 326 prompts for 60s videos (VBench-Long) and 128 prompts for 240s videos (MovieGen). Baseline–HiDrop comparisons are matched within each horizon, and degradation is measured within individual rollouts. Each rollout is divided into non-overlapping 5s windows. AQ and IQ drops are measured relative to the first window; subject and background drift use DINO and CLIP similarity to the first frame; color drift compares normalized hue histograms with the first window. DD change tracks the proportion of dynamic clips relative to the initial window. The curves include pointwise 95% confidence intervals from 2,000 bootstrap resamples of matched baseline–HiDrop video pairs. (2) Results: Averaged over subsequent windows of the 240s videos, HiDrop reduces subject drift by 26–40% and color drift by 26–29%, with smaller AQ drops across all three models. It also better preserves dynamics: the mean change in dynamic proportion from the initial window improves from −19.78 to −8.49 percentage points for Causal Forcing and from −12.27 to +1.30 percentage points for LongLive. Please refer to Figures 6 and 7 in the revised manuscript for the full temporal curves.

  2. Continuation from imperfect history. To assess continuation from imperfect history, we use a 10s baseline-generated prefix. The baseline and HiDrop then each continue generation to 60s from the same historical KV cache and subsequent noise state. This compares their continuation quality under identical imperfect conditioning. VBench evaluates the complete videos, including the shared prefix. HiDrop improves AQ, IQ, and DD across all three backbones (Appendix C.7, Table 9).

Table 9. 60s generation from a shared baseline-generated prefix.

BackboneContinuationAQ ↑BC ↑DD ↑IQ ↑MS ↑SC ↑TF ↑
Self ForcingBaseline0.5820.9530.6930.7030.9790.9650.981
Self ForcingBaseline + Ours0.5980.9540.7650.7060.9800.9660.982
Causal ForcingBaseline0.5620.9580.4510.7020.9860.9640.982
Causal ForcingBaseline + Ours0.6020.9600.5530.7090.9860.9700.983
LongLiveBaseline0.5730.9510.7090.6940.9750.9600.982
LongLiveBaseline + Ours0.5950.9580.7730.6980.9780.9680.982

Changes in the paper: Appendix C.5 and Figures 6–7 report temporal drift. Appendix C.7 and Table 9 report continuation from a shared baseline-generated history.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.