acceptodds
Under review as a conference paper at ICLR 2027

EDCache: Endpoint-Driven Timestep Caching for Diffusion Transformers

Abstract

Diffusion transformers excel at image and video generation, but iterative denoising incurs high inference costs. Existing caching methods often use local proxy signals to decide what to reuse. However, under matched compute, we find that importance-based token selection does not consistently outperform random selection and incurs additional overhead. We therefore shift our focus from token-level to timestep-level caching: reusing whole-timestep outputs can cause trajectory drift, while the position of reuse strongly affects final generation quality. To address trajectory drift and refresh timing in whole-timestep reuse, we propose EDCache, a two-stage method combining trajectory correction and timestep scheduling. Stage 1 uses recent full-computation outputs to correct drift during reuse without offline calibration. Stage 2 constructs schedule costs from model-specific single-position endpoint-damage profiles and uses dynamic programming to select full-computation timesteps under a fixed budget. Experiments on DiT-XL/2, FLUX.1-dev, and Wan2.1 show that EDCache improves generation fidelity at high denoising speedups. On DiT-XL/2, EDCache achieves more than 5× denoising speedup with FID 3.64, outperforming all evaluated competing methods in the 2.5–5× speedup range across all five reported distribution metrics.

Primary area:
generative models
Keywords:
Diffusion Transformers, Inference Acceleration, Timestep Caching, Trajectory Correction, Endpoint-Driven Scheduling

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_49WPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EDCache, a two-stage whole-step output reuse method for accelerating diffusion transformers. Motivated by the observation that importance-based token selection does not consistently outperform random selection at matched compute, the paper shifts from token-level to timestep-level caching and uses controlled single-position interventions to show that endpoint damage varies sharply with the cache-injection position and depends on the model–sampler configuration. Stage 1 corrects cache-induced trajectory drift by extrapolating from the two most recent full-compute anchors without offline calibration, and Stage 2 constructs interval costs from a measured single-position endpoint-damage profile and selects full-computation timesteps by dynamic programming under a fixed budget.

Critical strengths
  • The controlled single-position interventions show that endpoint damage depends on the cache-injection position and the model–sampler configuration, providing direct guidance for timestep-level caching.

  • Stage 2 achieves lower paired LPIPS than all evaluated caching baselines on FLUX.1-dev and Wan2.1 across all four budgets, with the measured profile reused across budgets.

Critical weaknesses
  • The Stage 2 formulation, which selects full-computation timesteps by dynamic programming over calibrated costs, follows the same line as DPCache, and the position-dependent endpoint damage was already reported by OmniCache [1], which is not cited. The benefit of the endpoint-damage-based cost should be demonstrated through a controlled comparison with DPCache.

  • The DiT-XL/2 experiments omit the most direct timestep-level methods (DPCache, MagCache, TeaCache), and the Wan2.1 experiments do not include a comparison with DPCache, so the advantage of Stage 2 over the strongest timestep-level baselines remains unverified.

<Reference> [1] Omnicache: A trajectory-oriented global perspective on training-free cache reuse for diffusion transformer models, ICCV, 2025.
Questions for authors

Q1. How were the exponents (1.5, 1.8) and the temporal tilt (−0.1) in Eq. 6, the coefficient λ, and the cached-span limit h (2–11 across models and budgets) chosen, and how sensitive are the results to these values? Since the Stage 2 schedules for FLUX.1-dev and Wan2.1 compute most early steps in full (Table 5), how does Stage 2 compare with a simple schedule following the same pattern?

Q2. The profile is measured without Stage 1 repair, whereas repair is applied during final inference. Is this mismatch the reason why Stage 2 yields worse FID and sFID than Stage 1 on DiT-XL/2 at every budget?

Q3. Appendix D.2 states that measuring the profile requires 6,000 probes for DiT-XL/2 and 640 probes each for FLUX.1-dev and Wan2.1, with a substantial one-time generation cost, but the actual GPU time is not reported. What is the GPU time of this profile measurement, and how is this cost justified relative to DPCache, which uses only a small calibration set?

Reviewer_ahVgOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the problem of caching for diffusion transformers. The authors propose to combine: 1) a trajectory correction mechanism that extrapolates denoiser outputs from previous full computations, and 2) an endpoint-driven scheduling strategy that selects which timesteps to recompute.

The scheduling strategy estimates the impact of caching individual timesteps on the final generated output and uses dynamic programming to optimize the caching schedule.

Critical strengths
  • The paper addresses the relevant caching problem.
  • I appreciate the evaluation across three different models (DiT-XL/2, FLUX.1-dev, and Wan2.1), covering both image and video generation.
  • The ablation studies separating trajectory correction and scheduling are useful.
Critical weaknesses

Several closely related approaches already investigate error-aware caching and timestep scheduling, including DPCache, BudCache, and SenCache (CVPR 2026). SenCache is particularly relevant but is neither cited nor compared against. The main contribution appears to be the combination of trajectory extrapolation with endpoint-driven schedule optimization. The authors should better clarify the novelty of each component compared with existing approaches. Also the experiment raises several questions. See below.

Questions for authors
  1. How does EDCache differ from BudCache and DPCache? Why is BudCache not included in the quantitative comparisons? SenCache (CVPR 2026) is also relevant but not discussed.
  2. How well does the proposed scheduling cost correlate with the actual endpoint error of complete caching schedules? Could the authors report this correlation across different schedules at matched budgets?
  3. Equation 6 introduces some parameters. How sensitive are the results to these choices?
  4. How costly is calibration? Does changing the sampler require a new profile?
Reviewer_JzumOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies training-free acceleration of diffusion transformers through whole step reuse, in which the denoiser runs at only B of the 50 sampling steps and the other steps reuse earlier outputs, a choice motivated by the finding that importance -based token selection shows no consistent advantage over random selection at matched compute. The authors also find that caching a single step changes the final sample by an amount that depends strongly on the step's position, and they report that this damage profile shifts with the model and sampler but keeps its shape across samples. EDCache fills skipped steps by secant extrapolation from the two most recent full outputs, with a fixed coefficient and no extra network calls, and picks the B full steps by dynamic programming over a hand-set cost derived from the offline measured profile, allowing at most h consecutive cached steps. Experiments cover DiT-XL/2 on ImageNet-256, evaluated with FID-50K, sFID, IS, precision and recall, and FLUX.1-dev and Wan2.1-T2V-1.3B, evaluated with LPIPS, SSIM and PSNR against their own 500 step outputs. Compared with native few-step sampling and a model-specific subset of caching methods, the authors report a better trade-off between speed and fidelity for B from 25 down to 10, its about 2× to 5× speedup, including FID 3.64 on DiT at about 5×.

Critical strengths

S1. The single step intervention is a simple and useful diagnostic. Caching one step at a time and measuring the effect on the final sample is more direct than the local proxies that TeaCache or MagCache rely on. The profile also shifts between DiT–DDIM, DiT–DDPM and FLUX. That is a practical warning against carrying a schedule from one model to another by step index.

S2. The fidelity numbers at high speedup are strong. At roughly 5× speedup, Stage 2 reaches: LPIPS 0.219 on FLUX (DPCache 0.239, MagCache 0.285); LPIPS 0.329 on Wan (MagCache 0.409); FID 3.64 on DiT (10-step DDIM: 12.26). The FLUX schedule also holds up on 100 T2I-CompBench prompts that were not used for calibration (Table 8). There is no online overhead and latency is fixed, which helps in deployment. These results are the main reason my score is not lower.

Critical weaknesses

W1. for two of the three samplers, whole step reuse is few step sampling under another name, and the paper does not compare against methods built for that problem.

With FLUX's Euler sampler, or DDIM with η=0\eta = 0 on DiT, applying the same denoiser output over several small steps gives the same result as taking one large step. Caching without repair therefore amounts to BB step sampling on a non-uniform grid. The paper's own numbers fit this: on FLUX at B=15B = 15, the uniform schedule without correction reaches LPIPS 0.403 in table 12, close to 0.418 for native 15 step sampling in table 2, although the two use different prompt sets. Read this way, stage 2 is a timestep schedule search and stage 1 is close to a second order Adams-Bashforth step. Yet the only equal NFE references are the native samplers. DPM-Solver++ and UniPC are not run on DiT or FLUX, and AYS, GITS and OSS are neither compared nor cited. BudCache, which also searches for a schedule under a fixed budget, is cited but not compared. I cannot tell whether EDCache beats the strongest methods at the same number of network calls, and this is the main reason for my score.

W2. could not find direct evidence that the measured damage profile is what makes the schedules work.

Each probe caches one step, reuses an output that is one step old, starts from an unperturbed trajectory and has repair turned off. At deployment, reused outputs can be up to hh steps old (h=11h = 11 for FLUX B10), errors accumulate, and repair is on. A hand set cost,

w(t,q)=dˉ(t)1.5⋅(t−q)1.8⋅exp⁡(−0.1 τ),w(t,q) = \bar{d}(t)^{1.5} \cdot (t-q)^{1.8} \cdot \exp(-0.1\,\tau),

is meant to bridge this gap, but the paper never checks how well it tracks the LPIPS of a complete schedule. Two simple controls are also missing: a constant profile, and a rule that places full steps densely early and sparsely later. Table 5 makes me suspect the second would come close. On FLUX and Wan the profile is near zero after about step 10, so hh ends up fixing the late positions: the last two gaps in FLUX B10 hold exactly 11 cached steps each, and every late gap in Wan B10 holds 7. The DiT results point the same way. At B=15B = 15 the DP schedule does slightly worse than the uniform one with repair (FID 2.69 vs 2.60, sFID 5.60 vs 5.05). The 12-point DiT profile also skips the largest peak in Fig. 1c, near steps 46 to 48, and the B10 schedule caches every step from 45 to 49. Since endpoint-driven scheduling is the paper's central claim, this gap weighs heavily in my assessment.

Questions for authors

Q1. Please add equal-NFE baselines at B = 10 and B = 15, ideally with the skipped steps filled the same way across methods: DPM-Solver++(2M) and UniPC as solvers, a schedule from AYS, GITS or OSS, and BudCache. Since whole-step reuse without repair is B-step sampling on a non-uniform grid under Euler and DDIM (η=0\eta = 0), these are the comparisons that decide whether EDCache helps. If it still leads, I would raise my score.

Q2. Please rerun the schedule search with (a) a constant profile in place of dˉ(t)\bar{d}(t), and (b) a fixed rule that runs the first k steps in full and spaces the rest geometrically, keeping the cost form and hh unchanged. On FLUX and Wan the measured profile is near zero after about step 10, so I suspect the second control would come close. Please report the gap to Stage 2 on all three models.

Q3. How well does the proxy cost w(t,q)=dˉ(t)1.5⋅(t−q)1.8⋅exp⁡(−0.1 τ)w(t,q) = \bar{d}(t)^{1.5} \cdot (t-q)^{1.8} \cdot \exp(-0.1\,\tau) track reality? The probes measure a single step with one-step-old outputs and no repair, while deployment runs with outputs up to hh steps old and repair on, and the paper never checks the gap. Reporting the Spearman correlation between J^\hat{J} and the endpoint LPIPS over a set of sampled schedules would settle W2.

Q4. fig 1c puts the largest damage near steps 46 to 48, but the 12-point profile used for scheduling does not cover that region, and the B10 schedule then caches all of steps 45 to 49. Were the two figures measured under the same setting, and does the schedule change if the profile is measured at all 50 positions?

Reviewer_rSb2OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

EDCache accelerates diffusion transformers through output extrapolation and offline timestep scheduling. It uses the two most recent full evaluations to predict outputs at skipped steps. To choose which steps to compute, it measures how caching at individual positions affects the final output, then uses dynamic programming to select a schedule under a fixed budget. The experiments cover DiT, FLUX, and Wan.

Critical strengths

A Simple Method with Promising Results.** The method is simple and achieves promising results. On FLUX and Wan in particular, changing the placement of full evaluations substantially improves fidelity to the original outputs at almost the same inference time.

Reusable Calibration and Useful Evaluation.** A calibrated sensitivity profile can be reused across computation budgets for the same model and sampler, which makes the approach more practical. The evaluation covers image and video generation, several acceleration levels, component ablations, and an external set of prompts for FLUX.

Critical weaknesses

The Endpoint-Based Scheduling Cost Needs More Direct Validation. Forecasting from historical features and using dynamic programming to schedule full evaluations have already been explored in TaylorSeer and DPCache. The more interesting contribution here is using final-output damage to assess the importance of each timestep. It is still difficult to tell how much this particular choice contributes. Comparisons with existing methods change both the predictor and the scheduling method. The uniform-versus-DP ablation mainly shows that choosing computation positions matters. The theoretical analysis explains error propagation, but offers limited support for the specific cost function used in the method. During calibration, the method reuses an old output at a single step without extrapolation. During inference, it skips multiple steps and applies two-anchor extrapolation. The size and distribution of errors may differ between these settings. The results show that the selected schedules work well, but do not tell us how reliably the surrogate identifies better schedules. The cost exponents, temporal weighting, and maximum cached span may also influence the outcome.

Practical Gains Need to Be Weighed Against Calibration Cost and Generation Quality. The appendix reports 6,000 calibration probes for DiT and 640 each for FLUX and Wan, with 49 full network evaluations per probe. This is a substantial setup cost. The paper does not report calibration time or how many generations are needed to recover that cost. Measured latency is also missing for the reduced-step sampling baselines, leaving an important part of the efficiency comparison incomplete. The gains from Stage 2 are uneven across quality metrics. For DiT, its FID is slightly worse than Stage 1 at all four budgets. At B15, for example, FID changes from 2.5957 to 2.6882. These results raise a practical question about when the additional calibration is worth adopting.

Questions for authors

Q1. Could the authors keep the output predictor, computation budget, and schedule constraints fixed, and compare different scheduling costs? Useful alternatives would include local output error and a cost based only on the number of steps since the last full evaluation. This would make the value of endpoint-based calibration much clearer.

Q2. Could the authors evaluate a set of candidate schedules at the same budget and compare their surrogate costs with their actual LPIPS scores? I would also like to see whether the schedule rankings and selected schedules' performance remain stable when the cost parameters and maximum cached span change.

Q3. Once calibration is included, at what workload size does EDCache become worthwhile compared with Stage 1 and simply reducing the number of sampling steps? A comparison using measured runtime and generation quality would help answer this.

1 comment

Dear Authors,

Thank you for the detailed response and the additional experiments. The comparisons that hold the predictor, compute budget, and scheduling constraints fixed address my main concern. They show that the endpoint-based cost improves fidelity over the tested alternatives. The schedule-ranking analysis also helps explain how well the proxy works across models.

I appreciate the calibration cost and latency breakdown. Please retain the distinction in the revised paper: calibration can pay for itself through inference savings relative to full computation, while its benefit over Stage 1 is improved fidelity at the same inference latency.

I am raising my rating from 2 to 3 (Weak acceptance). My confidence remains 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.