acceptodds
Withdrawn from ICLR 2027

TRAC: Trajectory-aware Reuse and Adaptive Correction for Efficient Autoregressive Video Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Autoregressive (AR) video generation enables scalable long-video synthesis by sequentially generating fixed-length chunks, but incurs substantial computational cost from repeated denoising and attention over growing historical contexts. Recent cache reuse methods exploit temporal redundancy to reduce this cost, yet are mainly designed for single-trajectory generation, making them less suitable for sequentially coupled AR trajectories where approximation errors can accumulate and propagate across chunks. To address this challenge, we propose TRAC, a Trajectory-aware Reuse and Adaptive Correction framework for efficient and robust AR video generation. TRAC integrates three complementary components: (1) Robust Cumulative Scheduling (RCS), which selects cache reuse schedules by accounting for cumulative rollout errors and their variation across chunks and prompts; (2) Autoregressive Trajectory-aware Guidance Scheduling (ATGS), which coordinates classifier-free guidance refreshes along the global AR trajectory; and (3) Spectral Structure Correction (SSC), which adaptively restores low-frequency structural information to mitigate long-term error accumulation. By combining these innovations, TRAC substantially improves the efficiency-quality trade-off of AR video generation. For example, on VBench with SkyReels-V2, it achieves a 6.03× speedup over the vanilla model with only ∼ 0.6% VBench score degradation, while consistently outperforming existing AR acceleration methods in both latency and generation quality.

Primary area:
generative models
Keywords:
Autoregressive video generation, cache reuse, denoising scheduling, training-free acceleration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_LQXTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TRAC is a training-free framework for accelerating AR video generation. The main idea is to make cache reuse across denoising steps more aware of the generation trajectory. TRAC combines an offline schedule search for cache reuse, CFG residual reuse across denoising steps, and a structural correction mechanism to reduce computation while limiting error accumulation across autoregressive chunks. Experiments on two video generation models show that TRAC can achieve speedups with relatively small quality degradation.

Critical strengths

1. TRAC achieves speedups while keeping the quality degradation relatively small. For example, it achieves around 6× speedup on SkyReels-V2 with only a small drop in VBench.

2. The idea of using different reuse patterns for different denoising steps is reasonable, and RCS, ATGS, and SSC work together in a fairly coherent way. The ablation studies also make it relatively easy to see what each component contributes.

Critical weaknesses

1. I am not convinced that the main ideas are sufficiently novel. RCS, CFG residual reuse, and structural correction all seem closely related to existing ideas in cache reuse, adaptive scheduling, and error correction. Overall, the method feels more like combining and adapting existing techniques to AR video generation.

2. TRAC requires an offline calibration stage to determine the reuse schedule, but it is not clear whether or how this cost is accounted for in the reported speedups. Since the baselines do not appear to require the same offline schedule search, I think the paper should report the calibration cost and explain how it should be accounted for in the evaluation.

Questions for authors
  1. I would appreciate more clarification on the novelty of the proposed approach. In particular, it would be helpful if the authors could more clearly explain what the key new insights are behind RCS, CFG residual reuse, and structural correction, and how they differ from existing approaches to cache reuse, adaptive scheduling, and error correction.

  2. TRAC requires an offline calibration stage to determine the reuse schedule, but the cost of this stage is not clearly reported in the paper or reflected in the reported speedups. Since the baselines do not appear to require a similar offline schedule search, it would be helpful to provide more details on the calibration cost, such as the required computation, time, number of calibration samples, and model evaluations. It would also be helpful to clarify whether this cost is excluded from the reported latency and how it would be amortized in a practical setting.

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

This paper addresses cache error propagation in autoregressive video generation, integrating risk-aware denoising scheduling RCS, global CFG refresh scheduling ATGS, and segment-start-guided low-frequency correction SSC. A 6.03× speedup is reported on SkyReels-V2 with an approximately 0.53 drop in the overall VBench score; a 1.67× speedup is achieved on FramePack-F1.

Critical strengths
  • The method separately handles conditional branch reuse, CFG reuse and structural drift, and allocates denser computation to the first segment, which aligns with the characteristic that historical errors may affect subsequent generation.
Critical weaknesses
  • Equation (7) calculates the standard deviation, CVaR and terminal mismatch of the cumulative error over the entire path; these quantities generally cannot be decomposed into the sum of independent costs for each edge. However, Appendix B.1 directly uses scalar edge costs for dynamic programming, without defining their equivalence relationship or approximation error with the original objective. Therefore, it cannot be confirmed that the original objective is solved correctly based on this approach.
  • RCS gains depend on cache strength, as removing it doubles latency (266.9s to 572.9s, Fig. 5). Missing comparisons with MeanCache under identical NFE and initial budgets, and weight sensitivity analysis cannot substitute individual ablations of standard deviation, CVaR, and endpoint terms.
Questions for authors

As mentioned in the weaknesses. If the optimization targets in equal 7 a substitute objective, please clarify its definition. Moreover, after fixing the NFE and the first-segment budget, what are the advantages of RCS over MeanCache, and what are the respective contributions of each risk item to the mean error and tail error?

Reviewer_D21rOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TRAC, a training-free framework for accelerating autoregressive video generation through three components: Robust Cumulative Scheduling (RCS), Autoregressive Trajectory-aware Guidance Scheduling (ATGS), and Spectral Structure Correction (SSC). RCS optimizes cache reuse schedules using cumulative rollout errors and cross-prompt risk statistics; ATGS allocates classifier-free guidance (CFG) refreshes along the autoregressive execution trajectory; and SSC introduces a low-frequency structural correction mechanism using the first chunk as an anchor. The authors evaluate TRAC on SkyReels-V2 and FramePack-F1.

Critical strengths
  1. Training-free acceleration. TRAC does not require additional model training and could be useful in practical inference pipelines where retraining is expensive.

  2. A coherent optimization framework. The paper combines cache scheduling, CFG residual reuse, and structural correction within a single inference-time pipeline. The cumulative-risk objective in RCS and the budgeted scheduling formulation in ATGS provide explicit mechanisms for allocating computation.

Critical weaknesses
  1. Limited novelty relative to recent autoregressive video acceleration methods. The paper's main motivation and several technical ideas overlap substantially with recent work. ARCache (CVPR 2026) explicitly addresses error accumulation in autoregressive video diffusion through history-guided cache scheduling and residual correction. FreqForcing (2026) uses spectral self-anchoring to suppress low-frequency drift and preserve long-horizon visual stability. MotionCache (ICML 2026) adaptively allocates computation according to motion characteristics. These works address the same central challenges motivating TRAC: accumulated errors across autoregressive segments, adaptive allocation of expensive computation, and preservation of global structure during rollout. Although the implementations are not identical, the conceptual differences need to be articulated and experimentally established much more rigorously. In particular, SSC appears closely related to the structural anchoring principle in FreqForcing. Using pooled attention to extract a low-frequency correction and modulating its strength according to motion may be a useful implementation variant, but the paper does not sufficiently demonstrate why this constitutes a substantial methodological advance over existing spectral anchoring and correction strategies. Similarly, RCS extends offline cache-schedule optimization with cumulative error and risk-aware statistics, while ATGS applies budgeted scheduling to CFG residual reuse along the autoregressive path. The paper needs to clarify which aspects are genuinely new, which are adaptations of existing scheduling principles, and why the resulting combination offers advantages beyond a straightforward integration of established techniques. The related-work discussion and experimental comparisons should be strengthened to reflect these developments, particularly ARCache and FreqForcing. Without more direct comparisons and clearer technical differentiation, the contribution appears incremental.

  2. The qualitative evidence does not convincingly demonstrate a perceptual advantage. Figures 6–13 provide sampled-frame comparisons across several prompts. However, the differences between TRAC and the baselines are generally subtle, and TRAC does not consistently exhibit an obvious visual advantage over vanilla inference or the competing acceleration methods, while the score in Table 1-3 really improves.

  3. More importantly, the paper does not provide readily inspectable generated videos that would allow readers to evaluate. Static frame grids are inadequate for validating claims about autoregressive trajectory robustness.

  4. The paper should provide complete video comparisons using matched prompts and random seeds, with representative cases where the proposed method clearly improves over the baselines. Failure cases should also be included.

  5. The quantitative evaluation does not adequately establish the claimed advantage over recent methods. The paper does report full standard VBench results and VBench-long, so the concern is not the complete absence of VBench. Rather, the aggregate scores do not provide sufficiently convincing evidence of a meaningful perceptual improvement.

  6. On SkyReels-V2, TRAC achieves 82.444 compared with 82.975 for vanilla inference. Its score is only 0.464 points higher than FlowCache-fast (81.981), despite the reported speedup being almost identical (6.028× versus 5.988×). The practical significance of this relatively small difference is unclear without per-dimension results, repeated runs, uncertainty estimates, or stronger perceptual evidence. Moreover, Table 3 reports a higher aggregate score for FreqForcing (83.460) than for TRAC (82.925) on the 94-prompt subset, although FreqForcing incurs substantially higher latency. This supports a potential efficiency advantage, but does not establish that TRAC achieves superior generation quality.

  7. The paper should provide a more complete comparison of the quality-speed Pareto frontier against the most relevant recent methods, including ARCache and FreqForcing, under matched backbones, prompts, seeds, resolutions, video lengths, and evaluation protocols. Per-dimension VBench results, representative videos, and uncertainty estimates would make the comparison more informative. The authors should also explain the differences between the reported VBench scores in Tables 1 and 3, including the different evaluation-set sizes and configurations.

  8. The long-video evaluation is insufficient for a method motivated by cumulative autoregressive errors. Long-horizon error accumulation is a central motivation for TRAC, and the proposed RCS and SSC components are explicitly intended to mitigate degradation along the autoregressive trajectory. Nevertheless, the reported evaluation uses 177 frames for SkyReels-V2 and 145 generated frames for FramePack-F1, corresponding to only a small number of autoregressive sections in the reported configurations. These experiments do not adequately establish whether the method remains effective over genuinely long rollouts. In particular, the paper does not demonstrate how quality, identity consistency, color stability, motion, and structural coherence evolve as the generated video becomes substantially longer.

Questions for authors
  1. Novelty relative to ARCache and FreqForcing: What is the precise technical distinction between SSC and the spectral anchoring mechanism in FreqForcing? How does TRAC differ from ARCache's history-guided scheduling and residual correction beyond the specific implementation choices? Can the authors provide matched-budget comparisons and component-level analyses that isolate these differences?

  2. Long-horizon effectiveness: Can the authors provide results at substantially longer generation horizons, including at least 30-second and 60-second settings where supported by the backbones, and ideally minute-scale rollouts? How do temporal consistency, visual quality, and accumulated error evolve with video length?

  3. Quality-speed trade-off: Given that FreqForcing obtains a higher reported aggregate score on the 94-prompt subset, what evidence demonstrates that TRAC offers a better overall operating point rather than simply a faster but slightly lower-quality alternative? Are the differences statistically or perceptually meaningful?

  4. Contribution of individual components: Since ATGS is inactive in the reported FramePack-F1 configuration with CFG = 1, how much of the overall improvement generalizes beyond the SkyReels-V2 setting? Can the authors provide additional evaluations that independently validate the three components under appropriate configurations?

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

Related papers

Discussion (0)

Sign in to comment.