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

cs.CVOct 2, 2026

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

Authors: Jiaxing Song, Weiqi Yan, You Huang, Mingte Qiu, Huazhong Liu, Xiaofeng Zhu, Yunshan Zhong

Organizations: School of Computer Science and Technology, Hainan University · MAC Lab, Department of Artificial Intelligence, School of Informatics, Xiamen University · Zhejiang Normal University, Jinhua, China

Abstract

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.

Figures & tables

Figure 1: Overview of the proposed trajectory-aware reuse and adaptive correction (TRAC). TRAC comprises three complementary mechanisms for AR video generation, including robust cumulative scheduling (RCS), autoregressive trajectory-aware guidance scheduling (ATGS), and spectral structure correction (SSC).

Figure 2: Cumulative rollout error (vertical) versus frame index (horizontal), with error accumulated from frame‑wise output RMSE. Curves compare MeanCache and our RCS, averaged over 30 prompts randomly sampled from VBench using SkyReels‑V2‑1.3B‑540P. Shaded regions indicate the standard error of the mean, and vanilla outputs serve as the reference.

Figure 3: The horizontal axis shows the guided denoising step index along the global AR path. The left and right vertical axes plot the CFG residual error (purple) and cumulative reuse cost, respectively. ATGS (Ours), Gate, and Uniform are three CFG refresh schedules compared under the same budget. Results are on SkyReels‑V2‑1.3B‑540P with 30 prompts from VBench.

Figure 4: Low-frequency spectral-energy error (normalized by the first frame). Purple: cache-based rollout w/o SSC; green: same rollout with SSC enabled. SSC reduces low-frequency error growth under identical rollout settings. Results on SkyReels‑V2‑1.3B‑540P with 30 VBench prompts.

Method	Latency (s) $\downarrow$	Speedup $\uparrow$	VBench $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Vanilla	1608.6470	1.0000 $\times$	82.9749	–	–	–
TeaCache-slow	814.0000	1.9762 $\times$	81.6700	18.9600	0.6501	0.2472
TeaCache-fast	686.0000	2.3450 $\times$	80.0600	18.3900	0.6121	0.3063
FlowCache-slow	323.4470	4.9734 $\times$	82.0870	17.2600	0.6370	0.3145
FlowCache-fast	268.6330	5.9883 $\times$	81.9805	16.2437	0.6088	0.3597
MotionCache	343.3827	4.6847 $\times$	81.8000	17.1731	0.6301	0.3267

Table 1: Comparison on SkyReels-V2-1.3B-540P ( Chen et al., 2025 ) .

Method	Latency (s) $\downarrow$	Speedup $\uparrow$	Imaging Quality $\uparrow$	Quality $\uparrow$	I2V $\uparrow$	Total $\uparrow$
Vanilla	360.7360	1.0000 $\times$	0.7055	0.8203	0.9565	0.8884
TeaCache-fast	228.8670	1.5760 $\times$	0.7108	0.8125	0.9557	0.8841
TeaCache-slow	254.3220	1.4180 $\times$	0.7048	0.8123	0.9580	0.8851
FlowCache-fast	241.3940	1.4940 $\times$	0.7073	0.8117	0.9609	0.8863
FlowCache-slow	280.2290	1.2870 $\times$	0.6804	0.7780	0.9030	0.8405
MotionCache	253.9000	1.4210 $\times$	0.7049	0.8077	0.9582	0.8830

Table 2: Comparison on FramePack-F1 ( Zhang et al., 2026 ) .

Method	Latency (s) $\downarrow$	Speedup $\uparrow$	Quality $\uparrow$	Semantic $\uparrow$	Total $\uparrow$
FreqForcing	583.4457	2.7571 $\times$	83.4603	82.5981	83.2879
FasterCache	458.6105	3.5077 $\times$	80.8739	77.8126	80.2617
MeanCache	405.6705	3.9654 $\times$	67.2080	61.6877	66.1040
TRAC	266.8461	6.0284 $\times$	82.9253	82.1763	82.7755

Table 3: Additional baseline comparison on SkyReels-V2.

Figure 5: Component ablation on SkyReels-V2. The scatter plot compares latency and VBench total after removing one component at a time.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Qualitative comparison on SkyReels-V2. Prompt: “Two pandas discussing an academic paper.”

Figure 7: Qualitative comparison on SkyReels-V2. Prompt: “A confused panda in calculus class.”

Figure 8: Qualitative comparison on SkyReels-V2. Prompt: “The Bund Shanghai, oil painting.”

Figure 9: Qualitative comparison on SkyReels-V2. Prompt: “A giant robot in Cyberpunk Beijing.”

Figure 10: Qualitative comparison on SkyReels-V2. Prompt: “A corgi is playing drum kit.”

Figure 11: Qualitative comparison on SkyReels-V2. Prompt: “Busy freeway at night.”

Figure 12: Qualitative comparison on SkyReels-V2. Prompt: “Time lapse of sunrise on Mars.”

Figure 13: Qualitative comparison on SkyReels-V2. Prompt: “Spaceship landing on Mars.”

$\lambda_{A}$	Quality $\uparrow$	Semantic $\uparrow$	Total $\uparrow$
0.20	82.4367	81.7599	82.3013
0.30	82.7358	81.7433	82.5373
0.40	82.9253	82.1763	82.7755
0.50	82.0764	82.4972	82.1606
0.60	82.0731	83.4813	82.3547

Appendix

Table 4: Ablation studies on

\lambda_{A}

$\eta$	Quality $\uparrow$	Semantic $\uparrow$	Total $\uparrow$
0.03	82.9253	82.1763	82.7755
3.00	82.9524	81.9728	82.7565
10.00	82.9309	81.9741	82.7395

Appendix

Table 5: Sensitivity to the RCS weight

\eta

, with

\rho=0.01

and

\lambda_{q}=0.003

fixed.

$\rho$	Quality $\uparrow$	Semantic $\uparrow$	Total $\uparrow$
0.01	82.9253	82.1763	82.7755
3.00	82.9524	81.9728	82.7565
20.00	82.9524	81.9728	82.7565

Appendix

Table 6: Sensitivity to the RCS weight

\rho

, with

\eta=0.03

and

\lambda_{q}=0.003

fixed.

$\lambda_{q}$	Quality $\uparrow$	Semantic $\uparrow$	Total $\uparrow$
0.003	82.9253	82.1763	82.7755
0.30	82.6695	82.7072	82.6770
1.00	82.6695	82.7072	82.6770

Appendix

Table 7: Sensitivity to the RCS weight

\lambda_{q}

, with

\eta=0.03

and

\rho=0.01

fixed.

Explore similar work

Not All Frames Deserve Full Computation: Accelerating Autoregressive Video Generation via Selective Computation and Predictive Extrapolation
Apr 3, 2026Hanshuai Cui, Zhiqing Tang, Zhi Yao +3Autoregressive Video Generation Diffusion Model Inference Acceleration
School of Artificial Intelligence, Beijing Normal University, Beijing, China · Institute of Artificial Intelligence and Future Networks, Beijing Normal University, Zhuhai, China · Shenzhen University of Advanced Technology, Shenzhen, China
Motion-Aware Caching for Efficient Autoregressive Video Generation
May 3, 2026Jing Xu, Yuexiao Ma, Xuzhe Zheng +7Video Diffusion Models KV-Cache Management
1Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, P.R. China · 2ByteDance · 3Max Planck Institute for Intelligent Systems, 4ELLIS Institute Tbingen, 5Tbingen AI Center
DCARL: A Divide-and-Conquer Framework for Autoregressive Long-Trajectory Video Generation
Mar 25, 2026Junyi Ouyang, Wenbin Teng, Gonglin Chen +2Long-Horizon Video Generation Autoregressive Video Generation
Institute for Creative Technologies · University of Southern California

Apr 3, 2026J/K move · Enter open · S save

Not All Frames Deserve Full Computation: Accelerating Autoregressive Video Generation via Selective Computation and Predictive Extrapolation

Hanshuai Cui, Zhiqing Tang, Zhi Yao, Fanshuai Meng+2

School of Artificial Intelligence, Beijing Normal University, Beijing, China · Institute of Artificial Intelligence and Future Networks, Beijing Normal University, Zhuhai, China · Shenzhen University of Advanced Technology, Shenzhen, China

Autoregressive Video Generation Diffusion Model Inference Acceleration Video Diffusion Models

Open paper AskPDF arXiv

Autoregressive (AR) video diffusion models enable long-form video generation but remain expensive due to repeated multi-step denoising. Existing training-free acceleration methods rely on binary cache-or-recompute decisions, overlooking intermediate cases where direct reuse is too coarse yet full recomputation is unnecessary. Moreover, asynchronous AR schedules assign different noise levels to co-generated frames, yet existing methods process the entire valid interval uniformly. To address these AR-specific inefficiencies, we present SCOPE, a training-free framework for efficient AR video diffusion. SCOPE introduces a tri-modal scheduler over cache, predict, and recompute, where prediction via noise-level Taylor extrapolation fills the gap between reuse and recomputation with explicit stability controls backed by error propagation analysis. It further introduces selective computation that restricts execution to the active frame interval. On MAGI-1 and SkyReels-V2, SCOPE achieves up to 4.73x speedup while maintaining quality comparable to the original output, outperforming all training-free baselines.

Abstract

Explore similar work

Not All Frames Deserve Full Computation: Accelerating Autoregressive Video Generation via Selective Computation and Predictive Extrapolation

Motion-Aware Caching for Efficient Autoregressive Video Generation

DCARL: A Divide-and-Conquer Framework for Autoregressive Long-Trajectory Video Generation