acceptodds
Under review as a conference paper at ICLR 2027

PRICE: Propagation-Risk-Informed Dynamic Diffusion Caching under Exact Compute Budgets

Abstract

Diffusion caching accelerates sampling by reusing features across denoising steps. Given a target compute budget, a caching policy must decide which steps should use full computation and which can reuse the cache. Existing adaptive caching methods often make this decision based on local signals, such as the difference between adjacent steps. However, the goal of caching is to preserve the final output of full computation, and a small local error does not always lead to a small final error. Our single-step caching experiments show that similar local errors can have very different final effects depending on where they occur in the denoising process. Motivated by this observation, we propose PRICE (Propagation-Risk-Informed Caching under Exact budgets), a training-free method that estimates this final effect from the local discrepancy, solver step size, and downstream propagation. At each step, it compares the estimated risk with offline reference risks for the remaining steps and adjusts the threshold to the remaining reuse quota. On FLUX.1-dev, Qwen-Image, and Wan2.1, PRICE achieves 5.24x, 6.87x, and 3.49x acceleration, respectively, while better preserving full computation outputs than state-of-the-art caching baselines and maintaining competitive perceptual quality.

Primary area:
generative models
Keywords:
Generative models, Diffusion caching, Inference acceleration, Training-free

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_46HZOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PRICE, a training-free dynamic caching policy for diffusion Transformers under a strict per-sample compute budget. Its starting point is a mismatch that is observed but not systematically addressed: the magnitude of a local cache error does not determine the eventual impact of that cache action on the final output. The paper derives a discrete error-propagation bound that decomposes the terminal contribution of a cache error into local perturbation magnitude, solver-step scaling, and downstream propagation. Because the quantities in that bound (the true perturbation E_i and the per-timestep Lipschitz constant Λ_i) are unavailable online, the method builds a causal risk proxy from an offline timestep-dependent sensitivity estimate L_i and the current displacement from the most recent full-computation anchor.

Critical strengths
  • PRICE addresses a real and important mismatch between local cache error and final-output impact.

  • The exact-budget controller is simple, causal, and cleanly constrained.

  • The risk-component ablations are thorough in their combination.

Critical weaknesses
  • The offline risk quantiles are estimated under a uniform caching schedule, whereas the deployment policy changes the distributions of states and anchors. The paper has not validated whether these quantiles remain applicable during adaptive rollouts or how sensitive the final performance is to the risk-bank collection policy.

  • The connection between the theoretical bound and the practical risk estimator is insufficiently established. The theorem uses true local perturbations and fixed-time Lipschitz constants, whereas the deployed estimator uses state distances and adjacent-timestep trajectory secants as proxies. Although the ablations support the score’s practical utility, they do not directly verify whether it accurately estimates local caching errors or correctly ranks the terminal impact of individual caching actions.

  • The reproduced ERTACache results on FLUX.1-dev are markedly inconsistent with the other reported results, and the paper provides no explanation.

  • The authors are encouraged to include a focused analysis of GOC[1] and HiCache[2] in the related work section. Discussing these representative studies would help trace the development of diffusion feature caching and clarify the technical positioning of this submission.

[1]. Qiu, Junxiang, et al. "Accelerating diffusion transformer via gradient-optimized cache." 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025.

[2]. Feng, Liang, et al. "HiCache: A plug-in scaled-Hermite upgrade for Taylor-style cache-then-forecast diffusion acceleration." International Conference on Learning Representations. Vol. 2026. 2026.

Questions for authors

See the Weaknesses

Reviewer_88hfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies feature caching for diffusion transformers under an exact per-sample budget of BB full evaluations. Observing that local cache error and final-output deviation rank reuse positions differently, the authors derive a discrete error-propagation bound and define a risk score Ri=Si∣hi∣wiR_i=S_i|h_i|w_i. Here SiS_i scales the state displacement since the last full step by an offline local sensitivity LiL_i, ∣hi∣|h_i| is the solver step size, and wiw_i is an offline downstream propagation weight. Online, PRICE reuses the cache when RiR_i falls below the quantile of an offline risk bank over the remaining steps, taken at the remaining reuse fraction, while feasibility rules enforce the budget exactly. Experiments on FLUX.1-dev, Qwen-Image, and Wan2.1 report better LPIPS, SSIM, and PSNR relative to 50-step outputs than seven caching baselines at matched NFE.

Critical strengths
  1. A training-free online controller that meets the budget exactly. The suffix-quantile threshold with feasibility rules guarantees exactly BB full evaluations per sample, without training or online search and at negligible overhead, which yields predictable latency. It consistently outperforms a global threshold built from the same risk bank at all three FLUX budgets (e.g., LPIPS 0.3928→0.2869 at B=9B=9), making it the best-supported contribution of the paper.

  2. Consistent fidelity gains across architectures. The experiments cover two image DiTs and one video DiT with Euler and UniPC solvers and compare against seven recent caching methods at matched NFE. PRICE achieves the best LPIPS, SSIM, and PSNR in every reported setting, which is the main empirical case for acceptance.

Critical weaknesses
  1. The central risk score is not validated as a terminal-error predictor. Theorem 3.1 is a standard discrete Grönwall bound, while the implemented LjL_j is an adjacent-timestep secant that mixes the explicit tt-dependence of vθv_\theta with spatial sensitivity, not a Lipschitz estimate. Hence Li∣hi∣wiL_i|h_i|w_i is a fixed per-step weight, and Ri≤τiR_i\le\tau_i reduces to a TeaCache-like displacement-since-refresh test with a time- and quota-dependent threshold. In Table 9, adding wiw_i or ∣hi∣|h_i| alone to SiS_i worsens LPIPS at every budget (e.g., 0.1712→0.2627 at B=15B=15), so the two factors appear to offset each other rather than validate the decomposition. Rank correlations between RiR_i and the measured terminal error (Figures 2 and 4), and a variant weighted by the measured propagation profile, are needed.

  2. The state-of-the-art claim, the main empirical contribution, is weakly supported. On FLUX, the margins over the best baseline are small (LPIPS 0.0788 vs. 0.0794 at B=21B=21) and are reported without variance or paired significance tests. The large Qwen-Image gain highlighted in the introduction is measured against ERTACache, which is no better than vanilla 7-step sampling (LPIPS 0.5324 vs. 0.5377), while SeaCache, SenCache, MagCache, and BudCache are omitted. ERTACache on FLUX is even worse than vanilla 21-step sampling (0.3839 vs. 0.3298), suggesting misconfiguration. Uniform and random schedules using the same residual reuse are also missing, so the policy's contribution cannot be separated from the caching mechanism; these baselines and confidence intervals should be added.

Questions for authors

Clarification

  1. How much do PRICE's selected schedules vary across prompts, and how often are the forced-full and forced-reuse rules triggered?
  2. Is the global threshold in Table 3(b) constructed as F^−1(m/n)\hat F^{-1}(m/n) (Section 3.3) or tuned to match the mean NFE (Section 4.4), and does it also use the feasibility rules?
  3. Were ERTACache on FLUX and TaylorSeer on Qwen-Image run with their officially recommended settings, given that both fall below vanilla step reduction at some budgets?
  4. Where do the 32 image calibration prompts and the 80 prompts in Figure 4 (right) come from, and were design choices such as the median secant and the forced-full steps selected on the DrawBench evaluation set?
  5. How does PRICE fundamentally differ from BAG (Zhao et al., 2026a), which is cited only for its prompt protocol but tackles the same exact-NFE online caching problem under a nearly identical setup and reports LPIPS 0.0773 on FLUX at B=20B=20 versus 0.0788 for PRICE at B=21B=21?
Reviewer_3XULOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

PRICE is a training-free method that accelerates diffusion models by dynamically reusing cached features under a fixed compute budget. It estimates how caching errors affect the final output to decide which steps need full computation. Experiments on image and video models show improved output fidelity at comparable inference cost.

Critical strengths
  1. Online decisions require inexpensive state comparisons and precomputed statistics, with no additional full model evaluations or policy training.

  2. Preserving the original model’s output under an explicit computation budget is a useful deployment goal.

  3. The single-step caching experiments and equal-magnitude interventions support the argument that local error alone does not determine terminal impact.

Critical weaknesses
  1. Novelty needs a more precise comparison with the closest work.

  2. Several improvements are too small to assess without uncertainty estimates.

  3. Better reconstruction does not consistently produce better generation quality.

Questions for authors

Overall, I think this paper address an important problem in the realm of Diffusion Caching, and the idea is interesting. I have a few comments and questions as follow.

For the presentation of this paper, I think the words in the figures are too small to view.

  1. How accurately does the risk score predict final-output error? Please report correlations or ranking accuracy between RiR_i and measured terminal error, compared with local-error estimates.

  2. What is the main advantage over closely related methods? Please clarify the distinction from SenCache, ERTACache, and BAG, particularly whether gains come from propagation modeling, budget allocation, or avoiding training.

  3. Does uniform-schedule calibration remain reliable during adaptive inference? PRICE’s decisions change cache anchors and subsequent risks. Please quantify this mismatch and how often budget constraints force otherwise undesirable decisions.

  4. Are the improvements statistically significant? Several FLUX and Wan gains are small.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.