acceptodds
Under review as a conference paper at ICLR 2027

TaylorGAP: Progressive Anchor Scheduling and Geometry-Aware Taylor Forecasting for Training-Free Diffusion Acceleration

Abstract

Feature caching accelerates diffusion and flow-matching models by forecasting intermediate features at timesteps that skip a backbone evaluation. TaylorSeer estimates feature derivatives from cached full evaluations, but its fixed refresh interval commits to a full-length forecast block while only the minimum history is available; for a second-order predictor, increasing the offset from one to five enlarges the horizon-dependent factor in the Taylor remainder bound by . The subsequent transition from unit-spaced startup anchors to the wider refresh interval also creates a non-uniform anchor grid. On such a grid, adjacent lower-order factors have order-dependent effective centers whose separation generally differs from the latest anchor interval, so normalizing the recursive higher-order differences by that interval leaves those factors persistently mis-scaled. Based on these observations, we propose TaylorGAP, a training-free method that couples progressive anchor scheduling with grid-aware Taylor forecasting. Its progressive schedule uses anchor intervals to delay long extrapolations while preserving a fixed budget of ten full evaluations in a 50-step trajectory. Its geometry-aware recursion normalizes each order by the corresponding effective-center distance; on the progressive grid, this yields the closed-form coefficient . The correction is applied only at offsets one and two for stability. At a matched NFE budget, TaylorGAP improves consistently over equal-NFE TaylorSeer for orders two through four on both FLUX.1-dev and HunyuanVideo. Relative to the corresponding baseline, it reduces LPIPS by 28.69%–32.35% on images and 41.00%–43.00% on videos, raises PSNR by 18.28%–19.21% and 26.17%–28.64%, and raises SSIM by 12.23%–13.77% and 13.54%–15.29%. TaylorGAP thus provides a principled, training-free connection between cache scheduling and feature forecasting.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Diffusion Model Acceleration, Feature Caching, Taylor Forecasting, Training-Free Inference

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_TYVHOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TaylorGAP, a training-free diffusion acceleration method extending TaylorSeer. It introduces a progressive anchor schedule that gradually increases the intervals between full model evaluations and a geometry-aware correction for higher-order finite-difference estimates on non-uniform anchor grids. The correction is applied only to short prediction offsets. Experiments on FLUX.1-dev and HunyuanVideo show improvements over TaylorSeer at matched numbers of full model evaluations, particularly in PSNR, SSIM, and LPIPS. Additional experiments on FLUX.1-dev at 40 and 60 sampling steps demonstrate that these improvements persist across different trajectory lengths.

Critical strengths

The paper reports consistent improvements over TaylorSeer in PSNR, SSIM, and LPIPS across Taylor orders two through four on both image and video generation tasks. The comparisons use matched full-evaluation budgets, making the reported fidelity improvements meaningful. The additional experiments at 40 and 60 sampling steps also provide evidence that the improvements are not limited to a single trajectory length. The paper identifies a scaling issue in higher-order finite-difference estimates when anchor intervals are non-uniform and proposes a geometry-aware correction. The mathematical derivation and quadratic example provide a reasonable numerical motivation for this modification.

Critical weaknesses
  • The methodological novelty appears limited compared to TaylorSeer. The proposed method retains the same Taylor-series forecasting framework, with the main modifications being a progressive anchor schedule and a geometry-aware correction for higher-order finite differences. While these modifications are reasonable, the paper does not evaluate their individual contributions. The ablation study varies the number of sampling steps without separately testing the two components. It is therefore unclear how much of the reported improvement comes from the schedule alone and how much the geometry-aware correction adds.

  • The experimental comparisons are mostly against TaylorSeer, with FORA included only for image generation. The method is evaluated on FLUX.1-dev and HunyuanVideo, while the additional 40- and 60-step experiments are limited to FLUX.1-dev. The paper also does not report uncertainty estimates for its performance metrics. These limitations make it difficult to assess the method's competitiveness against other approaches and its generalizability beyond the evaluated settings.

  • The presentation could be clearer. The abstract and introduction contain detailed equations and numerical examples before the main idea is clearly explained, making the paper harder to follow. The main text is also equation-heavy, which further affects readability. In addition, the submission does not include the AI use statement required by the ICLR 2027 author guidelines.

Questions for authors

Q1. Could the authors provide matched-NFE ablations for progressive anchor scheduling alone, geometry-aware correction alone, and their combination? This would help clarify how much each component contributes to the reported improvements.

Q2. Have the authors tested alternative progressive schedules or different ranges for applying the geometry-aware correction? The paper provides a motivation for these design choices, but does not compare the specific intervals and the cutoff with reasonable alternatives.

Q3. Could the authors include comparisons with additional feature-caching methods at similar computational costs? It would also be useful to report variability or uncertainty for the relatively small improvements in ImageReward and VBench.

Reviewer_9cceOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

TaylorGAP proposes a training-free acceleration scheme that combines progressive anchor scheduling with geometric-aware Taylor prediction, addressing the extrapolation error caused by fixed intervals in existing feature caching methods and geometric scaling errors due to non-uniform grids.

Critical strengths

The relative fidelity improvement of NFE to TaylorSeer is stable. This is true, but not enough to support acceptance.

Critical weaknesses
  1. Lack of comprehensive comparison:

    a. The comparison method is only TaylorSeer. It is concerning whether it is a specific improvement to the method.

    b. There is no ablation for "only change scheduling" or "only change differencing"

  2. The design of the method is relatively direct and simple, but there are too many formulas; it lacks visualization, and readability is poor.

  3. Hallucination reference: HunyuanVideo in Section 3.1 is miscited as "Sun et al., 2024", which is an LLM paper.

Questions for authors
  1. Add ablation studies for each module.
  2. Include more comparisons of Token Cache Methods, such as TeaCache and HiCache.
  3. Improve the paper to enhance readability.
Reviewer_fPC1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TaylorGAP, a training-free inference acceleration method for diffusion and flow-matching models based on improved feature caching. Main Contributions

  1. Identifying the coupling between scheduling and prediction: The intervals between full network evaluations affect both the extrapolation horizon and the scaling of higher-order Taylor finite differences. Fixed-interval methods can introduce long-horizon extrapolation and scale mismatches during the startup phase.
  2. Introducing two jointly designed components: The intervals between full evaluations increase progressively as (1,2,3,...), while higher-order finite-difference factors are corrected to account for nonuniform spacing. The correction is applied only to the first two steps after each anchor to limit error amplification. Experimental Validation
  3. Tasks: The method is evaluated on FLUX.1-dev for image generation and HunyuanVideo for video generation, using 200 prompts and 4,720 videos, respectively.
  4. Quality: Compared with TaylorSeer under the same computational budget, TaylorGAP reduces LPIPS by approximately 29%–32% for images and 41%–43% for videos, producing outputs closer to those obtained with full computation. The improvements also hold under 40-step and 60-step sampling settings.
Critical strengths

1.The paper connects cache scheduling with Taylor prediction and uses second-order error analysis to justify the scale correction, providing a concrete mathematical basis for the design. 2.LPIPS decreases by approximately 29%–32% for images and 41%–43% for videos compared with TaylorSeer. Improvements across sampling lengths further support the method’s effectiveness.

Critical weaknesses
  1. The experiments do not isolate scheduling, geometric correction, leaving the independent contribution of the core correction mechanism unclear.
  2. Comparisons primarily involve TaylorSeer and FORA. Without stronger, closely related baselines, the method’s incremental value and quality–latency advantages remain insufficiently established.
Questions for authors
  1. Could you provide additional ablation studies—for example, to establish whether geometric correction provides independent benefits beyond progressive scheduling?
  2. Have you tried alternative progressive scheduling strategies?
  3. Does TaylorGAP retain its advantages over closely related methods, such as CG-Taylor, DPCache, and TC-Padé?
  4. The references for evaluation and metrics are incomplete; in addition, the way the tables are presented may be misleading.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.