acceptodds
Withdrawn from ICLR 2027

AlignFlow: Multi-Axis Cache Scheduling for Efficient Continuous Video Editing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Diffusion transformers now dominate video generation, and training-free feature caching is the standard way to cut their denoising cost, yet every existing cache is designed and evaluated for a single generation call. Real editing is not a single call: users issue a sequence of cumulative edits, each conditioned on the previous output. This discards the redundancy that exists across consecutive edits and makes approximation error compound recursively, so single-turn fidelity cannot characterize a session. Extending acceleration across edits is not a matter of adding a second cache. Raw cross-edit reuse is intermittent and, used alone, needs far more exact evaluations than a step-axis cache; finer token-level reuse traded a latency reduction for a memory increase in our diagnostic; and the edited region is precisely where past information is invalid. We present AlignFlow, a training-free controller that predicts the classifier-free-guided output and schedules reuse jointly along the step, region, and edit axes. Damped first-order extrapolation of the guided residual supplies the acceleration backbone; a progress-aligned bank of the previous turn's exact outputs supplies a conservatively mixed cross-edit source, which a dilated latent mask forbids inside the edited region; and a closed-loop calibrator admits each source only when its measured per-region, per-phase reliability clears a threshold, otherwise falling back to an exact step. On five-turn editing chains with VACE Wan2.1-1.3B, AlignFlow reaches a speedup at dB PSNR, dB above the strongest accelerated baseline at lower latency, and transfers to real video and a second architecture. Session-average gains exceed final-turn gains, so we report per-turn fidelity and drift together rather than claiming that recursive drift is removed.

Primary area:
generative models
Keywords:
Video Diffusion Acceleration, Training-Free Feature Caching, Continuous Multi-Turn Video Editing, Diffusion Transformers, Classifier-Free Guidance

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_gtgLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Training-free acceleration for DiT video editing where edits are applied cumulatively, each turn conditioning on the previous accelerated output. AlignFlow predicts the CFG-guided output and skips whole steps using damped first-order extrapolation across steps (backbone), a previous-turn output bank mixed at α=0.05 outside the edited region, and a per-region/phase reliability gate with exact fallback. Reports 1.965× at 27.23 dB vs. the unaccelerated session on 100 five-turn Wan2.1 chains (EasyCache 1.736×/18.26 dB), fastest but not most faithful on LOVEU/DAVIS, 1.633× on MiniMax-H3, plus an axis ablation and a ten-turn drift study.

Critical strengths
  1. Session-level framing with per-turn fidelity to the unaccelerated session is new to the caching literature and useful beyond this method.

  2. Candid reporting of where baselines win, unseparated ablation mechanisms, and unresolved drift.

Critical weaknesses
  1. The novel edit axis contributes least and is mostly inactive. Step extrapolation and region routing are known techniques and alone give 1.939×/25.57 dB (Table 5). The edit source at α=0.05 changes latency by −0.30% yet adds 2.03 dB, which the authors cannot attribute to the draft vs. altered admission. On DAVIS it is selected 294/19/0/0/0 times across turns; in the ten-turn study the staleness cap never fires. In the motivating masked real-footage setting the headline component is effectively off.

  2. Single operating point per baseline. One fixed threshold each vs. ~30 pilot-tuned parameters for AlignFlow (Table 12). Baselines being both slower and ~9 dB worse on Wan100 is abnormal for a speed–fidelity tradeoff; on DAVIS the methods are on a similar frontier (TeaCache-fast 1.535×/32.93 vs. 1.784×/32.28). T0 protocol is also unclear (Sec. 5.1 vs. Fig. 3).

  3. Drift is not solved and the metric conflates divergence with degradation. 17.2 dB at T10, guard recovers 0.07 dB. PSNR to the exact session measures trajectory divergence, not quality; no Exact VBench-8 on Wan100, no per-turn no-reference metric.

Questions for authors
  1. Was T0 accelerated for baselines and/or AlignFlow on Wan100? Report Table 1 with a shared exact T0.
  2. Sweep each baseline's threshold (≥3 points) and place AlignFlow on the curve.
  3. Does the +2.03 dB survive with admission decisions frozen to step+region?
  4. Report Exact's VBench-8 on Wan100 and per-turn no-reference quality for the ten-turn study.
Reviewer_htz3OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a new training-free framework for speeding up multi-turn video editing called AlignFlow. It skips certain DiT evaluations by predicting their outputs from earlier denoising steps. In addition, it reuses information from the previous editing turn outside of the region being changed and falls back to the full DiT when its predictions are considered unreliable. Experiments show roughly 2x speedup and better similarity to full inference than the tested caching baselines.

Critical strengths
  1. The paper addresses the practical and relatively underexplored problem of accelerating multiple turn video editing, where approximation errors can accumulate because each edited output becomes the input to the next turn.

  2. The method is training-free and achieves roughly 2x speedup while remaining closer to full inference than the previous caching baselines. The paper also provides useful ablations showing that most of the speedup comes from prediction across denoising steps, while the edit and region components mainly help preserve fidelity.

Critical weaknesses
  1. The contribution is quite limited (given the extensive prior work on caching for DiT acceleration), overly heuristic, and may have diminishing practical relevance. Most of the speedup comes from extrapolating DiT outputs between denoising steps, while the edit and region components mainly serve as safeguards. The method combines many fixed schedules, thresholds, and mixing weights, with limited evidence that they generalize beyond the tested settings. Moreover, it is evaluated with 20-30 denoising steps, whereas step-distilled models can operate in only 1-4 denoising steps, leaving little computation to skip and making the practical usefulness of the proposed method unclear.

  2. The evaluation does not establish meaningful editing quality. The main metrics measure similarity to the unaccelerated model rather than whether the requested edits are correct or whether unedited content is preserved. A high PSNR can simply mean that the method reproduces the full model’s own drift over repeated edits. This concern is particularly important because the paper provides no human evaluation.

Questions for authors
  1. How does the method perform with step-distilled video editing models using only one to four denoising steps?
  2. Can you please provide a human evaluation comparing the proposed approach against relevant baselines across multiple editing turns?
  3. Can you provide any visual results (as videos) to assess the quality of the presented framework? The rolled-out frames are not a full representation of a video since temporal artifacts are less visible.
  4. How sensitive are the results to the main controller hyperparameters, and were any of them retuned for MiniMax-H3 or the evaluation datasets?
  5. Can the authors provide quantitative or human evaluation of instruction following and preservation of unedited content, preferably separately for edited and preserved regions?
Reviewer_oR2eOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposed a cache scheduler for diffusion transformers targeting editing tasks. The proposed scheduler exploits the redundancy across subsequent editing operations, thus enabling caching across multiple denoising trajectories. The authors demonstrated the effectiveness of this strategy through experiments.

Critical strengths
  1. The idea of leveraging redundancy across multiple editing operations is sound and clear, and is of practical significance as it aligns with real use cases.
  2. Experimental results show that the proposed method consistently achieves promising efficiency.
  3. Detailed analysis over cache axis is given, and is supported by experimental results.
  4. Limitations (e.g., not fully resolve recursive drift) are properly acknowledged in the paper.
Critical weaknesses
  1. The claim that the assumptions of existing methods are broken in editing tasks (line 041) is not accurate. Existing methods only assume that features within a single generation call can be reused. The fact that features can be reused across multiple generation calls does not break their assumption. Therefore, the motivation for the proposed design is not correctly stated.
  2. Some parts of the paper are not easy to follow and lack a reasonable amount of explanation and context. For instance, the meaning of "anchor" is not sufficiently introduced in the main text. Readers who are not very familiar with the context need extra effort to understand this term and its role in the method.
  3. The preliminary section (lines 113–124) is not clear enough. The descriptions of concepts such as exact step and cached step are not sufficiently formal and clear, which makes the subsequent introduction to method design vague and requires extra effort to understand.
  4. The evaluation covers two dimensions, speed and fidelity, and the experimental results are actually promising. However, the presentation does not directly show the trade-off between the two dimensions. A 2D plot showing the trade-off between speed and fidelity would make the results more intuitive.
Questions for authors
  1. (W1) The claim in line 041 needs clarification. Please clarify if the proposed method is making an assumption to capture a not-before-exploited property of editing tasks, or if the proposed method indeed targets a case where existing methods break (if so, please explain how are they broken).
  2. (W2) The presentation of some key concepts needs improvement. Please address the following questions: What exactly does "anchor" refer to in this work? If it is a common terminology used in cache schedulers, please properly cite its origin.
  3. (W3) The preliminary section (lines 113–124) needs clarification. Please address the following questions: What is the exact definition of an "exact step" and a "cached step"? How are they distinguished in the diffusion process, and what is computed differently in each case? Can the authors provide a formal definition for both concepts (for instance, clearly state how the CFG output y in Eq. 1 are computed in these steps)?

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

Related papers

Discussion (0)

Sign in to comment.