Source-linked AI summary
TRAC: Trajectory-aware Reuse and Adaptive Correction for Efficient Autoregressive Video Generation
Jiaxing Song, Weiqi Yan, You Huang, Mingte Qiu, Huazhong Liu, Xiaofeng Zhu, Yunshan Zhong
TL;DR
AR video generation couples sequential chunk-level denoising trajectories, while existing acceleration methods mainly target single-trajectory generation. TRAC combines cumulative cache scheduling, global-trajectory CFG scheduling, and spectral structure correction. Across SkyReels-V2 and FramePack-F1, it achieves the highest inference efficiency and best generation quality compared with existing methods.
Problem
Existing acceleration methods mainly target single-trajectory generation, whereas AR video couples chunk-level trajectories and approximation errors can accumulate and propagate.
Source: §1 · p. 2Method
TRAC is a training-free framework combining robust cumulative scheduling, autoregressive trajectory-aware guidance scheduling, and spectral structure correction.
Source: §3.2.1 · p. 5, §3.2.2 · p. 5, §3.2.3 · p. 7Results
Across SkyReels-V2 and FramePack-F1, TRAC achieves the highest inference efficiency and best generation quality compared with existing methods.
Source: §6 · p. 10Takeaways & Limitations
On SkyReels-V2, TRAC accelerates long-video generation while maintaining high visual fidelity and temporal consistency.
Source: §1 · p. 2Generalization to emerging autoregressive video models remains an open question.
Source: §5 · p. 9
Abstract
from arXiv · showhide
In this paper, we present trajectory-aware reuse and adaptive correction (TRAC), a training-free framework for efficient autoregressive (AR) video generation. Existing acceleration methods mainly target single-trajectory generation with bidirectional attention. AR video generation, by contrast, sequentially couples chunk-level denoising trajectories. Consequently, approximation errors accumulate and propagate through the generation process. TRAC addresses this challenge with three components, including robust cumulative scheduling (RCS), autoregressive trajectory-aware guidance scheduling (ATGS), and spectral structure correction (SSC). RCS selects cache reuse schedules by cumulative rollout error and cross-chunk/prompt variation. ATGS coordinates CFG refreshes along the global AR trajectory. SSC restores low-frequency structure of the first chunk to correct long-term structural loss. Experiments on SkyReels-V2 and FramePack-F1 show that, compared with existing methods, TRAC achieves both the highest inference efficiency and the best generation quality for AR video generation.
1 INTRODUCTION
AR video generation reduces long-video generation to sequential chunk-level denoising, but repeated computation and coupled chunk trajectories make efficient generation challenging. TRAC combines three mechanisms to address this challenge and reports strong speed and quality results.
1 INTRODUCTION
- Each new chunk repeatedly attends to growing historical context and undergoes heavy denoising, while CFG requires conditional and unconditional predictions at guided steps.These costs leave long-video generation far from practical real-time deployment.Source: §1 · p. 1
- Cache-reuse accelerators selectively run the full model at some denoising steps and reuse cached information at others; training-free methods avoid extra training or changes to the base generator.Examples include cached Jacobian-vector-product velocity approximations and methods that refresh the guidance residual only at selected steps.Source: §1 · pp. 1–2
- Unlike single-trajectory methods, AR generation couples chunk-level trajectories, so earlier reuse errors can propagate; TRAC addresses this with RCS, ATGS, and SSC.RCS considers cumulative rollout error and cross-chunk/prompt risk, ATGS calibrates CFG refreshes over the global AR trajectory, and SSC restores low-frequency structure using the first chunk as a reference.Source: §1 · p. 2
- TRAC achieves a 6.03× speedup over the vanilla model with only ∼0.6% quality degradation (82.44 vs. 82.97) on SkyReels-V2.It also scores 0.36 higher on VBench than FlowCache-slow while reaching 6.03× vs. 4.97× speedup, and improves the score by 0.64 over MotionCache while accelerating generation by nearly 29%.Source: §1 · p. 2
2 RELATED WORK
Related work spans autoregressive chunked video generation, cache-based acceleration, and frequency-domain approaches to temporal consistency and structural drift.
Autoregressive Video Generation
- Autoregressive video approaches generate temporally ordered chunks, with later denoising trajectories relying on previously generated content as historical context.The passage describes recent long-video diffusion systems as underscoring the viability of autoregressive and diffusion-forcing approaches.Source: §2 · pp. 2–3
Diffusion Caching and Denoising Scheduling
- Training-free cache methods reuse slowly varying intermediate features across denoising steps, while CFG reuse reduces overhead by exploiting temporal redundancy in guidance residuals.The passage also describes chunk-wise cache policies and offline trajectory-map scheduling.Source: §2 · p. 3
Frequency-domain Stabilization for Video Generation
- Frequency-domain methods address temporal consistency and structural drift through spectral blending, multi-band fusion, or low-frequency anchoring.The passage identifies FreqForcing's Spectral Self-Anchoring as combining a motion-and-detail branch with an anchor branch for low-frequency structure.Source: §2 · p. 3
3 METHOD
TRAC combines cache scheduling, global-path CFG refresh scheduling, and spectral correction to address reuse errors across sequential AR chunk trajectories. Its components allocate reuse according to cumulative risk, coordinate refreshes globally, and restore low-frequency structure from the first chunk.
3.1 PRELIMINARIES AND NOTATION
- AR generation denoises fixed-length chunks sequentially using prior chunks as context; cache schedules choose full-model steps and reuse cached information between them.CFG reuse refreshes the guidance residual at selected denoising steps and reuses the most recent cached residual at other steps.Source: §3.1 · pp. 3–4
3.2 TRAJECTORY-AWARE REUSE AND ADAPTIVE CORRECTION
- TRAC combines robust cumulative scheduling (RCS), autoregressive trajectory-aware guidance scheduling (ATGS), and spectral structure correction (SSC) for efficient, robust AR video generation.RCS schedules model reuse, ATGS schedules CFG refreshes, and SSC corrects low-frequency structure.Source: §3.2 · p. 4
3.2.1 ROBUST CUMULATIVE SCHEDULING
- RCS selects schedules by cumulative rollout error and cross-prompt/chunk risk, using a denser first-chunk schedule because its errors affect more later chunks.RCS achieves lower cumulative rollout error than MeanCache (Gao et al., 2026).Source: §3.2.1 · pp. 4–5, Figure 2 · p. 5
3.2.2 AUTOREGRESSIVE TRAJECTORY-AWARE GUIDANCE SCHEDULING
- ATGS treats CFG reuse as path planning over the global AR trajectory, placing refreshes at high-error positions rather than scheduling each chunk independently.Its budgeted schedule minimizes total path error and uses dynamic programming; Figure 3 compares residual error and cumulative reuse cost under the same budget.Source: §3.2.2 · pp. 5–6, Figure 3 · p. 5
3.2.3 SPECTRAL STRUCTURE CORRECTION
- SSC uses first-chunk key/value caches in a low-resolution attention path, then extracts a low-frequency discrepancy rather than replacing the current attention output.Pooling suppresses token-level fluctuations, while the low-pass correction leaves high-frequency detail and short-range motion governed by the original output.Source: §3.2.3 · p. 7
- SSC scales correction by local motion: low-motion chunks receive stronger first-chunk structure, while high-motion chunks receive less, slowing low-frequency error growth.RCS sets full-model evaluation steps, ATGS sets CFG refreshes, and SSC operates orthogonally to both schedules.Source: §3.2.3 · p. 7, Figure 4 · p. 7
4 EXPERIMENTS
Experiments on SkyReels-V2 and FramePack-F1 evaluate TRAC against cache-acceleration baselines and report strong efficiency–quality trade-offs. Component ablations show distinct contributions from RCS, ATGS, and SSC.
4 EXPERIMENTS
- The evaluation compares SkyReels-V2-1.3B-540P and FramePack-F1 with training-free cache baselines under matched prompts, seeds, video lengths, resolutions, and input conditions.SkyReels-V2 is assessed with VBench, PSNR, SSIM, and LPIPS; FramePack-F1 uses VBench++ Imaging Quality, Quality, I2V, and Total scores, with end-to-end latency and speedup also reported.Source: §4 · pp. 7–8
4.1 RESULTS
- 6.0283× speedup accompanies a SkyReels-V2 latency reduction from 1608.6470 to 266.8501 seconds; TRAC scores 82.4440 on VBench, about 0.6% below Vanilla.On FramePack-F1, latency falls from 360.7360 to 216.2220 seconds, with the highest speedup (1.6680×) and Quality score (0.8205); its Total score of 0.8882 is 0.0002 below Vanilla.Source: Table 1 · p. 8, Table 2 · p. 8, §4.1 · p. 8
- 6.03× speedup at 266.8s gives TRAC the strongest efficiency among the additional SkyReels-V2 baselines, while FreqForcing scores 83.29 versus TRAC’s 82.78 on VBench Total.The comparison uses 94 matched VBench prompts; FreqForcing requires 583.4s, while FasterCache and MeanCache trail TRAC in speed and generation metrics.Source: §4.1 · pp. 8–9, Table 3 · p. 9
4.2 ABLATION STUDY
- Removing SSC lowers latency from 266.9 to 259.1 seconds but lowers VBench from 82.44 to 81.81; removing ATGS raises latency to 291.5 seconds with VBench at 82.39.Removing RCS raises latency to 572.9 seconds despite a VBench score of 83.24, showing the components have complementary roles.Source: §4.2 · p. 9
5 LIMITATIONS AND FUTURE WORK
The paper identifies open questions about TRAC’s generalization to emerging autoregressive video models and proposes combining it with other acceleration techniques as future work.
5 LIMITATIONS AND FUTURE WORK
- TRAC has been evaluated on two widely used autoregressive backbones, and its generalization to emerging AR video models remains open.The authors say addressing this question would require computational resources not currently available to them.Source: §5 · p. 9
- Combining TRAC with quantization, distillation, or sparse attention may yield additional efficiency gains, but is left for future work.The paper describes these acceleration techniques as orthogonal to TRAC.Source: §5 · p. 9
6 CONCLUSION
TRAC is a training-free framework for efficient autoregressive video generation that targets error accumulation across sequential chunk-level trajectories. Experiments on SkyReels-V2 and FramePack-F1 report the highest inference efficiency and best generation quality compared with existing methods.
6 CONCLUSION
- TRAC combines robust cumulative scheduling (RCS), autoregressive trajectory-aware guidance scheduling (ATGS), and spectral structure correction (SSC) for autoregressive video generation.The framework is training-free and addresses approximation errors that accumulate and propagate through the generation process.Source: §6 · p. 9
- RCS selects cache-reuse schedules using cumulative rollout error and cross-chunk and prompt variation, while ATGS coordinates CFG refreshes along the global AR trajectory.These components target reuse scheduling and guidance refreshes across coupled chunk-level trajectories.Source: §6 · p. 9
- SSC restores low-frequency structure from the first chunk to correct long-term structural loss.Source: §6 · pp. 9–10
- Experiments on SkyReels-V2 and FramePack-F1 report that TRAC achieves the highest inference efficiency and best generation quality compared with existing methods.Source: §6 · p. 10
AI USE STATEMENT
The authors report using generative AI tools for language polishing and code-writing assistance, not experimental execution or research evidence.
AI USE STATEMENT
- Generative AI tools were used for language polishing and code-writing assistance, not to replace experiments, generate data, or supply unverified conclusions or literature evidence.The authors state that they reviewed the generated content and take responsibility for the final work.Source: AI USE STATEMENT · p. 10
REPRODUCIBILITY STATEMENT
The paper reports reproducibility measures including released anonymized code, fixed post-calibration schedules, and identical evaluation conditions.
REPRODUCIBILITY STATEMENT
- Schedules are fixed after calibration for deterministic execution with zero online scoring overhead, and experiments use fixed seeds and identical prompts, resolutions, and video lengths.Anonymized source code is provided in the supplementary material; method and settings are described in Sections 3 and 4, with further implementation details in Appendix A.Source: REPRODUCIBILITY STATEMENT · p. 10
A IMPLEMENTATION DETAILS
The appendix specifies backbone-specific execution settings and describes budgeted dynamic programming for RCS and ATGS, followed by a unified calibration and inference procedure for TRAC.
A IMPLEMENTATION DETAILS
- PyTorch experiments run on NVIDIA A800 GPUs at batch size 1, with backbone-specific RCS, ATGS, and SSC configurations.SkyReels-V2 uses 177 frames, 50 denoising steps per chunk, 30 first-chunk and 10 later-chunk full-model passes; FramePack-F1 uses 145 frames and 54 full-model passes across four sections. ATGS is inactive for FramePack-F1 because cfg = 1; SSC is enabled.Source: A IMPLEMENTATION DETAILS · p. 14
B.1 RCS DYNAMIC PROGRAMMING
- RCS assigns calibrated scalar scores to feasible reuse spans and uses budgeted dynamic programming to select minimum-score schedules by backtracking.The solver initializes at the fixed warm-up prefix; it computes Sbase under budget BD and the denser first-chunk Sfirst under expanded budget Bf.Source: §B.1 · p. 15
B.2 ATGS DYNAMIC PROGRAMMING
- ATGS treats CFG refresh placement as fixed-budget path planning, minimizing accumulated calibrated reuse costs and backtracking to recover the refresh schedule.The state tracks minimum accumulated CFG reuse cost for a path ending at refresh step j with exactly b refresh allocations; boundary events can be enforced as mandatory nodes.Source: §B.2 · pp. 15–16
B.3 UNIFIED ALGORITHM FOR TRAC
- Offline calibration produces RCS schedules, the ATGS refresh schedule, and a fixed SSC branch; inference applies them chunk by chunk.The first chunk uses Sfirst and later chunks use Sbase; full-evaluation steps may use SSC, while reuse steps skip SSC, and CFG residuals are refreshed at scheduled steps or when no compatible cached residual exists.Source: §B.3 · pp. 16–17
C QUALITATIVE RESULTS
The qualitative results section presents matched-prompt SkyReels-V2 frame comparisons across eight prompts and five generation methods.
C QUALITATIVE RESULTS
- Eight matched-prompt comparisons show Vanilla, TeaCache-slow, FlowCache-slow, MotionCache, and TRAC under the same prompt and seed.Each panel shows four sampled frames; the prompts include pandas discussing an academic paper or attending calculus class, Shanghai in oil-painting style, a Cyberpunk Beijing robot, a corgi playing drums, a nighttime freeway, and two Mars scenes.Source: C QUALITATIVE RESULTS · p. 17, Figure 6 · p. 18, Figure 7 · p. 18
- The remaining matched SkyReels-V2 prompts depict a busy freeway at night, a sunrise time lapse on Mars, and a spaceship landing on Mars.These appear alongside the other qualitative comparisons, with four sampled frames per panel under matched prompts and seeds.Source: C QUALITATIVE RESULTS · p. 17, Figure 11 · p. 20, Figure 12 · p. 21
D ADDITIONAL ABLATION STUDIES
The ablations examine SSC correction strength and RCS objective weights on a matched 94-prompt SkyReels-V2 subset; λA = 0.40 gives the highest Total score.
D ADDITIONAL ABLATION STUDIES
- λA = 0.40 gives the highest Total score in the SSC correction-strength ablation.Table 4 reports Quality, Semantic, and Total metrics; the remaining settings follow the default configuration.Source: D ADDITIONAL ABLATION STUDIES · p. 22, Table 4 · p. 22
- The RCS weight sweeps vary η, ρ, and λq independently while fixing the other two weights at their default values.The tables specify fixed settings of ρ = 0.01 and λq = 0.003 for the η sweep, η = 0.03 and λq = 0.003 for the ρ sweep, and η = 0.03 and ρ = 0.01 for the λq sweep.Source: D ADDITIONAL ABLATION STUDIES · p. 22, Table 5 · p. 22, Table 6 · p. 22
- Some RCS weight settings select the same calibrated schedule after discrete path search and therefore produce identical evaluation scores.The ablations use the matched 94-prompt SkyReels-V2 subset unless otherwise stated.Source: D ADDITIONAL ABLATION STUDIES · p. 22