Papers
Topics
Authors
Recent
Search
2000 character limit reached

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

Published 3 Apr 2026 in cs.CV | (2604.02979v1)

Abstract: 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.

Summary

  • The paper introduces SCOPE, a tri-modal scheduler that uses predictive extrapolation and selective computation to reduce redundant denoising steps.
  • The approach achieves up to 4.73× speedup on SkyReels-V2 and 2.55× on MAGI-1 while maintaining video quality comparable to full computation.
  • By leveraging Taylor series-based predictors and dynamic frame selection, SCOPE enables near-lossless acceleration for autoregressive video diffusion.

Accelerating Autoregressive Video Diffusion with Selective Computation and Predictive Extrapolation

Introduction

Autoregressive (AR) video diffusion models are instrumental for generating long-form, high-quality videos. However, their inference cost is prohibitive due to repeated transformer forward passes over numerous denoising steps and AR units. Existing training-free acceleration methods predominantly depend on binary cache-or-recompute paradigms, which fail to exploit intermediate cases where direct reuse is suboptimal and full recomputation is unnecessary. Additionally, asynchronous AR schedules introduce spatial redundancy by uniformly processing all frames within a valid interval, regardless of their denoising activity, resulting in wasted computation. This paper introduces SCOPE (Selective COmputation with Predictive Extrapolation), a training-free framework that addresses both temporal and spatial inefficiencies in AR video diffusion by incorporating predictive extrapolation and selective computation.

Methodology

Tri-Modal Scheduler: Cache, Predict, and Recompute

SCOPE extends the conventional binary cache-or-recompute policy to a tri-modal scheduler by integrating a Predict mode. When step-level similarity is insufficient for direct reuse and yet does not require complete recomputation, SCOPE leverages Taylor series-based velocity extrapolation along the denoising trajectory. This enables the scheduler to interpolate features between cache reuse and full recomputation, with stability controls anchored in rigorous error propagation analysis. The discrepancy score, computed from lightweight pre-forward features, governs the selection among the three operational modes via thresholding, ensuring robust decisions and restricting drift from prediction-induced errors.

Predictive Extrapolation

The Predict mode forecasts denoising velocities using Taylor expansion in noise-level coordinates. Both first-order and second-order predictors are employed, ensuring local smoothness of the denoising trajectory and explicit estimation of velocity progression. Error propagation is bounded using Taylor remainder terms, with stability ensured by discrepancy decay, bounded skip-length, and mandatory periodic recomputation. This mechanism substantially reduces redundant computation during intermediate denoising steps without degrading output fidelity.

Selective Computation in Asynchronous AR Schedules

Spatial redundancy in asynchronous AR schedules arises when model forwards are uniformly applied across a full-frame interval, even to frames that have converged or are yet to initiate meaningful denoising. SCOPE resolves this by dynamically identifying the active frame subset at each iteration, restricting computation exclusively to those frames actively advancing their denoising state. The active interval is computed directly from the scheduler step matrix, and selective execution is orthogonally compatible with the Predict mode, compounding temporal and spatial efficiency gains.

Experimental Results

SCOPE is benchmarked on MAGI-1 and SkyReels-V2, two state-of-the-art AR video diffusion frameworks. On SkyReels-V2, SCOPE achieves up to 4.73× speedup with LPIPS and VBench quality metrics nearly indistinguishable from the original pipeline. On MAGI-1, speedup reaches 2.55× at comparable quality. SCOPE consistently outperforms all training-free baselines, including FlowCache, TeaCache, and TaylorSeers, in both efficiency and fidelity. The tri-modal policy outstrips binary reuse schemes, with experimental ablations confirming that selective computation independently contributes to speedup and that Taylor-based prediction delivers robust error control.

A detailed sensitivity analysis demonstrates that SCOPE's performance is robust to threshold selection for cache and prediction modes, and that bounded skip-length is an effective trade-off knob for balancing speed and quality depending on the AR backbone architecture. SCOPE's scalability improves with increasing video length, and peak GPU memory usage remains competitive with other caching-based baselines.

A notable claim, substantiated by experiment, is that same-runtime original pipelines with step count reduction are both slower and substantially lower in quality relative to SCOPE, confirming that trajectory-preserving acceleration via extrapolation is preferable to naive reduction in denoising depth.

Implications and Future Directions

SCOPE's tri-modal scheduler introduces a new axis of flexibility into training-free diffusion acceleration, enabling fine-grained trade-offs between quality and efficiency and reconciling local predictability with stability constraints. Its spatial branch, selective computation, adapts naturally to the heterogeneous frame-wise structure of modern AR schedules. Practically, SCOPE offers a near-lossless acceleration method for high-resolution, long-form video generation, making AR diffusion viable for real-time and resource-constrained applications.

Theoretically, SCOPE's error propagation analysis sets a foundation for further exploration of predictive extrapolation, risk-aware scheduling, and hybrid caching strategies in generative modeling. Future research may extend SCOPE via automated predictor selection, threshold adaptation, or integration with orthogonal acceleration techniques at the attention or token level. Further work could enhance transferability across architectures, reduce the reliance on manual threshold calibration, and explore learned extrapolation for greater generality and reliability.

Conclusion

SCOPE advances training-free acceleration for AR video diffusion by combining a tri-modal scheduler with selective computation, resolving both temporal and spatial redundancy. Experimental results on MAGI-1 and SkyReels-V2 demonstrate substantial speedups without quality degradation, validating SCOPE's superiority over existing baselines. The framework is robust, scalable, and compatible with contemporary AR diffusion paradigms, and its methodological innovations suggest promising future directions for efficient generative modeling in both research and industry (2604.02979).

Paper to Video (Beta)

No one has generated a video about this paper yet.

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Open Problems

We found no open problems mentioned in this paper.

Collections

Sign up for free to add this paper to one or more collections.