acceptodds
Under review as a conference paper at ICLR 2027

PoSTCache: From Forecasting to Postdictive Smoothing for Revisable Diffusion Caching

Abstract

Diffusion and flow matching models achieve impressive visual generation quality but remain computationally expensive due to iterative sampling. Caching-based acceleration provides a practical training-free solution by reusing or predicting intermediate features. However, existing methods largely enforce a forward-only information flow: approximations produced for skipped model evaluations are immediately committed to the sampling trajectory and cannot benefit from feedback generated by later full-computation steps. We propose PoSTCache (Postdictive Smoothing Trajectory Cache), a training-free and plug-and-play framework that reformulates diffusion caching as revisable trajectory inference. PoSTCache first uses an existing caching strategy to construct forward proposals for skipped steps. At a later full-computation step, the discrepancy between the refreshed prediction and its forward proposal yields a correction signal, which PoSTCache propagates backward to refine and reintegrate the preceding predictions. Viewing full-computation steps as sources of corrective feedback also changes how they should be selected. We therefore introduce the Posterior Variance Reduction (PVR) strategy, which estimates cross-timestep dependencies from offline denoising trajectories and selects anchors according to their expected reduction in surrogate trajectory uncertainty rather than local approximation error alone. Experiments on representative image and video generation models show that PoSTCache consistently improves the quality–efficiency trade-off over representative caching methods.

Primary area:
generative models
Keywords:
Diffusion Models, Flow Matching, Generative Model Acceleration, Diffusion Caching, Postdictive Smoothing, Efficient Inference

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_arfVOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PoSTCache, a training-free caching framework that uses feedback from later full-computation steps to retrospectively correct skipped predictions and reintegrate the sampling trajectory without additional model evaluations. It also introduces Posterior Variance Reduction (PVR), an offline strategy that selects full-computation timesteps using cross-timestep covariance statistics under a fixed budget. Experiments on image and video generation models report improved reconstruction fidelity at comparable or higher acceleration, with ablations examining the contributions of retrospective correction and timestep selection

Critical strengths
  1. Simple and efficient design. PoSTCache uses feedback from existing full-computation anchors to correct earlier cached predictions. Correction and reintegration are lightweight and require no additional model evaluations at skipped steps.

  2. Broad experimental coverage. Experiments cover both image and video generation models, with fixed-budget ablations supporting the effectiveness of retrospective correction and timestep selection.

Critical weaknesses
  1. Limited novelty. The default correction is essentially linear interpolation between full-computation outputs, followed by reintegration. PVR introduces a new criterion for offline timestep selection, an already well-studied approach. While useful, these contributions appear relatively incremental.

  2. Unclear performance advantage, particularly for video generation. The reported results suggest that PoSTCache may not offer the best quality–speed trade-off among existing methods. For example, D2Cache [1] appears to preserve video quality better at comparable speedups, yet is not included in the comparisons. A direct comparison with this baseline would help clarify the practical advantage of PoSTCache.

[1] Enhuai Liu, Yunke Wang, Changming Sun, and Chang Xu. “D2Cache: Second-Order Delta Caching for Higher Video Diffusion Acceleration.” CVPR 2026.

Questions for authors

Could the authors compare PoSTCache with D2Cache [1] on Wan 2.1 and OpenSora at comparable speedups? Reporting the VBench drop relative to full inference, alongside reconstruction metrics, would help clarify whether PoSTCache offers an advantage over this baseline in video generation.

1 comment

Thank you for the further clarification and additional results. Most of my concerns was addressed. I will adjust my score accordingly.

Reviewer_gT79OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PoSTCache, a training-free framework for accelerating diffusion models via caching. Unlike prior forward-only methods, PoSTCache retrospectively revises skipped predictions using the discrepancy between a later full computation and its forward proposal (the anchor innovation), then re-integrates the corrected trajectory. The paper also proposes PVR, an offline anchor selection method that minimizes surrogate trajectory uncertainty under a fixed computation budget. Theoretical derivations support both components, and experiments on multiple text-to-image and text-to-video models show improved reconstruction fidelity over existing caching baselines.

Critical strengths
  • The paper presents postdictive smoothing as a novel alternative to forward-only caching, with a clear theoretical motivation from Bayesian smoothing.
  • Experiments across image and video models show consistent improvements over baselines, with negligible online overhead and compatibility with different forward caching strategies.
Critical weaknesses
  • Weak justification for the scalar gain: the replacement of the derived covariance-dependent gain with a relative-position scalar gain relies only on the empirical analysis in Sec. 3.2. The sample size of this analysis is not reported, and the “96% retention” is measured against the oracle scalar gain, not the full gain. It therefore does not show that the simplification preserves most of the achievable correction.
  • Smoothing contributes less than PVR: in Table 3, PVR alone outperforms MagCache with smoothing, and smoothing barely improves DiCache (PSNR 23.61 → 23.66). Most of the gain appears to come from anchor scheduling rather than from the central postdictive correction.
Questions for authors
  • The offline calibration cost is excluded from latency, yet it must be repeated for each model and sampling configuration. Is the comparison with calibration-free baselines fully fair?
  • How many samples were used in the Sec. 3.2 analysis?
1 comment

Dear authors,

Thank you for the detailed rebuttal for my concerns. Most of my concerns have been resolved by the rebuttal. Accordingly, I will maintain my score of 3.

Best regards,

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

This paper studies caching-based acceleration for diffusion-based models. The motivation lies in that existing caching methods are predominantly forward-only. When later full model evaluations reveal accumulated errors, the preceding cache is not reliable. PoSTCache treats these predictions as revisable proposals. At a later full-computation anchor, it computes the discrepancy between the refreshed prediction and its forward proposal. Then it propagates a correction backward to the preceding skipped predictions, and reintegrates the interval using the original numerical solver. The paper additionally proposes Posterior Variance Reduction (PVR) strategy, which estimates cross-timestep dependencies from offline complete denoising trajectories and selects anchors according to their expected reduction in surrogate trajectory uncertainty rather than local approximation error alone. Extensive experiments confirm the effectiveness of PoSTCache.

Critical strengths
  1. This paper is well-written and easy to follow.
  2. The core claim makes sense: in diffusion sampling, errors naturally accumulate and utilizing some heuristic methods to update the cache is reasonable.
Critical weaknesses
  1. According to Figure 3 and related texts, PoSTCache uses a scalar gain with a linearly increasing relative-position weight to achieve the corrected cache. However, diffusion is widely seen as a non-linear condition-dependent process. It is hard to judge whether this heuristic formulation really align with the exact correction direction.
  2. PVR uses covariance estimated from a small number of offline complete trajectories as a proxy for cache proposal error. However, covariance itself stands for the correlation among states in a trajectory. It does not directly have connection with the estimation error. At this time, it is hard to judge whether this choice is reasonable and correct.
Questions for authors
  1. Please see questions.
  2. I wonder whether the computation cost of the offline trajectories has been included in the latency. What if the default configuration, e.g., the scheduler or CFG setting, is changed? Does PoSTCache still requires extra sampling?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.