acceptodds
Under review as a conference paper at ICLR 2027

Terminal-Sensitivity-Aware Feature Caching for Efficient Diffusion Inference

Abstract

Feature caching accelerates diffusion inference by reusing intermediate features across denoising steps, but existing adaptive methods typically make reuse decisions from estimated local denoiser-output differences. The same local difference can produce different terminal errors because its effect depends on the scheduler step size and the sensitivity of the remaining denoising trajectory. To account for these factors, we introduce Terminal Sensitivity Caching (TSenCache), a training-free method that estimates the terminal impact of each reuse decision. TSenCache guides reuse decisions by combining a sample-adaptive local-error estimate with the target scheduler step size and a calibrated sensitivity of the remaining trajectory. Keeping the scheduler step size separate also allows the calibrated sensitivity profiles to transfer across sampling settings through noise-level interpolation. Experiments on FLUX.1-dev and Wan2.1 demonstrate that TSenCache preserves higher reconstruction fidelity than the evaluated caching methods at comparable inference speed.

Primary area:
generative models
Keywords:
Diffusion Model, Model Acceleration, Caching Mechanism, Training-Free

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

This paper proposes a feature caching method that identifies steps with a large impact on the final output error, unlike existing methods that consider only local output differences.

Critical strengths

The method achieves a favorable trade-off between speedup and output quality, as shown in Tables 1 and 2. Table 5 also demonstrates its applicability to various methods.

Critical weaknesses
  1. The use of scheduler step size and a local output-difference surrogate offers limited novelty. Furthermore, the remaining-trajectory sensitivity relies on a calibration set, limiting the technical contribution.

  2. In Table 3, when evaluating the effect of αF(k+1)\alpha_F(k+1)—which can be considered a key component of this paper—comparing e^k\hat{e}_k alone with αF(k+1)e^k\alpha_F(k+1)\hat{e}_k actually shows degraded performance in terms of SSIM and LPIPS. However, sufficient analysis regarding this phenomenon is lacking.

Questions for authors
  1. How exactly was the calibration set constructed, and what assumptions guided its selection?

  2. How sensitive is the method’s performance to the choice of calibration set?

5 comments

Thank you for recognizing the speed–quality trade-off and applicability across local estimators. We clarify the methodological contribution, explain the component ablation, and provide additional evidence on calibration-set sensitivity below.


W1. Novelty of the step-size factor, local surrogate and calibrated remaining-trajectory sensitivity.

Our contribution is a new terminal-impact-aware criterion for input-dependent online feature caching. Similar local errors at different steps can have substantially different effects on the final output; improving the local-error estimate alone therefore does not establish which reuse decisions are tolerable (Figure 1; Appendix I).

To our knowledge, TSenCache is the first online, sample-adaptive feature-caching method to base reuse decisions on estimated terminal impact. Existing online methods [1–3] estimate local output changes; LeMiCa and BudCache target terminal reconstruction through a fixed schedule shared across inputs. TSenCache instead accounts for the remaining trajectory while adapting each reuse decision to the current sample. Reviewers nTFS and 9cwQ also recognized the importance of considering final-output error.

Calibration does not weaken the methodological contribution. TeaCache [1] (CVPR 2025), MagCache [2] (NeurIPS 2025), and SenCache [3] (CVPR 2026) calibrate, respectively, a mapping from input to output differences, residual-magnitude ratios, and denoiser sensitivities. These precedents demonstrate that designing how to calibrate task-relevant information can itself be a methodological contribution.

The methodological contribution lies in how to calibrate terminal impact and the online decisions this enables. TSenCache measures how a reuse perturbation propagates through the remaining trajectory to the final output. This information enables the online policy to distinguish tolerable local errors from those requiring full computation, without evaluating future steps during inference.

Figure 6 and Table 8 demonstrate that terminal weighting improves the alignment of the same local estimator with measured terminal impact, including across reuse gaps. The fixed/online comparison in nTFS Q1 (Table 11) establishes the benefit of using this information for sample-specific decisions. The criterion improves fidelity with all four local estimators in Table 5, and Table 4 demonstrates profile reuse across sampling configurations.

References

[1] Feng Liu et al. "Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model." CVPR, 2025.
[2] Zehong Ma et al. "MagCache: Fast Video Generation with Magnitude-Aware Cache." NeurIPS, 2025.
[3] Yasaman Haghighi and Alexandre Alahi. "SenCache: Accelerating Diffusion Model Inference via Sensitivity-Aware Caching." CVPR, 2026.

W2. Why does αF⋅e^\alpha_F \cdot \hat{e} worsen SSIM and LPIPS compared with e^\hat{e} alone in Table 3?

αF⋅e^\alpha_F \cdot \hat{e} is a partial ablation, not the complete TSenCache criterion. As discussed in §4.4, the default FLUX schedule has larger step sizes during the later sampling steps. Omitting ∣Δt∣\lvert\Delta t\rvert disregards the larger latent perturbation produced by a given output error at these steps, shifting computation toward earlier steps. Protecting earlier steps improves PSNR, but insufficient later refreshes reduce the structural and perceptual fidelity captured by SSIM and LPIPS. The recorded computation allocation illustrates this behavior (Table 12):

Score, FLUX.1-dev at 10 NFEFull computations in steps 1–9Full computations in steps 19–28PSNR ↑SSIM ↑LPIPS ↓
e^\hat{e}4.793.2921.9210.80150.2153
∣Δt∣⋅e^\lvert\Delta t\rvert \cdot \hat{e}2.295.8115.8930.63720.4572
αF⋅e^\alpha_F \cdot \hat{e}7.771.0023.0440.73480.3338
αF⋅∣Δt∣⋅e^\alpha_F \cdot \lvert\Delta t\rvert \cdot \hat{e} (TSenCache, main configuration)6.022.3624.6710.83630.1695

Weighting the local score by ∣Δt∣\lvert\Delta t\rvert alone shifts computation in the opposite direction and reduces early-step protection. The complete score combines both effects, preserving early computation while restoring later refreshes and improving all three fidelity metrics.


Q1. How was the calibration set constructed, and what assumption motivates it?

For the main experiments in the submission (§4.1), we follow SenCache's small-set approach: 8 COCO captions for FLUX.1-dev and 8 MixKit descriptions for Wan2.1, disjoint from evaluation prompts. Calibration requires only prompts, without training or ground-truth outputs. The underlying assumption is that the dominant timestep-dependent downstream sensitivity is shared across prompts, while the online local-error estimate accounts for sample-specific differences. Eq. 9 and Appendix A specify the measurement procedure.


Q2. How sensitive is performance to the choice of calibration set?

The fidelity gains persist across four pairwise disjoint calibration sets of 8 prompts each. Figure 7 in the submission shows similar sensitivity profiles as the number of calibration prompts increases. We conducted an additional experiment to assess how replacing the calibration prompts affects final generation quality. Each set was used separately to calibrate a profile and evaluate FLUX / DrawBench-200 at 10 NFE (Table 7). No prompt is shared between any calibration sets.

Calibration setNFEPSNR ↑SSIM ↑LPIPS ↓
Set A (8 prompts)1023.6750.82330.1836
Set B (8 prompts)1024.4070.83540.1699
Set C (8 prompts)1024.2440.83120.1764
Set D (8 prompts)1023.6700.82190.1883
SenCache (Table 1)1022.1080.80280.2129

All four sets outperform the online baselines on all three fidelity metrics. The additional experiment therefore complements the profile-similarity analysis by demonstrating that the final-quality benefit persists across distinct calibration sets.


We hope these clarifications establish the contribution of terminal-impact-aware online caching and address your questions about calibration. We would appreciate your reassessment in light of the clarified contribution and additional experimental evidence.

Reviewer_yvwROpenReview ↗

Official Comment by Reviewer yvwR

  1. In response to my first weakness, the authors identify the terminal-aware criterion as their main contribution over the local output-difference criterion. However, Table 3 shows that both SSIM and LPIPS decrease when the proposed terminal sensitivity αF(k+1)\alpha_F(k+1) is incorporated into the local output-difference surrogate e^k\hat{e}_k (previous method). This result seriously undermines the claimed contribution of terminal sensitivity. Describing αF(k+1)e^k\alpha_F(k+1)\hat{e}_k as a partial ablation does not address my concern. Moreover, it is unclear why Table 3 from the original submission is reproduced in the rebuttal, as it provides no new information relevant to this concern.

  2. My concern regarding the calibration set has been resolved.

Overall, my assessment of the main contribution remains unchanged, and I therefore maintain my score of 1.

Reviewer_9cwQOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the limitations of existing adaptive diffusion caching methods, which typically make reuse decisions based on local output differences. The authors observe that similar local errors at different denoising steps can lead to substantially different errors in the final output, due to varying scheduler step sizes and the sensitivity of the remaining denoising trajectory. Motivated by this observation, they propose TSenCache, a training-free caching method that estimates the terminal impact of each reuse decision. Specifically, it combines an online estimate of local error with the scheduler step size and an offline-calibrated terminal sensitivity to guide cache reuse.

Critical strengths
  1. The motivation is well justified, and the paper identifies an important problem. Diffusion caching should focus on the error in the final output rather than local errors alone.

  2. The paper is clearly written and logically structured.

  3. The method achieves good reconstruction fidelity across the evaluated benchmarks.

Critical weaknesses
  1. Although the paper identifies a good problem, I am not fully convinced by its analysis. I understand that analyzing isolated cache decisions is a common practice in diffusion caching papers. However, Figure 1(c) shows an almost monotonically decreasing terminal error, which may simply reflect that errors introduced earlier propagate through more denoising steps in a Markovian process. It is difficult to distinguish whether this pattern comes from greater sensitivity at specific timesteps or simply from their earlier positions. Moreover, the analysis considers only a single cache reuse with full computation at all other steps, whereas practical caching involves multiple reuse decisions. It remains unclear whether the same error pattern holds under realistic caching trajectories.

  2. The quantitative and qualitative evaluations are not sufficiently comprehensive. Tables 1 and 2 compare all baselines at only one speedup, while Figure 3 evaluates different NFE budgets only against SenCache, making it difficult to assess the method's advantages across different speedups against different baselines. Some qualitative examples are also unconvincing, such as the apparent failure in the pizza example in Figure 10 and the blurry subjects in Figure 11. Although the blurriness is also present in the full-computation outputs, it makes the qualitative comparison less informative.

  3. The main paper lacks comparisons with caching methods that explicitly target terminal reconstruction error. For example, offline search methods such as BudCache directly optimize the reconstruction loss at the final output and are therefore particularly relevant to the paper's motivation. Although these methods are compared in the appendix, I believe this comparison should be included in the main experimental results.

Questions for authors

This paper is borderline for me (around 2.5), mainly due to the weaknesses mentioned above. However, I understand that evaluating and analyzing terminal errors from intermediate denoising steps is inherently challenging. I also find this research topic interesting, so I am willing to give the paper a score of 3 to encourage the authors to address my concerns during the rebuttal. I will then decide whether to maintain or lower my score based on their responses.

  1. Regarding Weakness 1, I have a simple alternative: replace the calibrated remaining-trajectory sensitivity with a monotonically decreasing function of the denoising step, without offline calibration. This would only require setting the initial and final weights and choosing a decay rate. Could the authors compare this simple design against TSenCache? This could help assess how much of the improvement comes simply from prioritizing earlier steps.

  2. Could the authors provide a complete comparison with all baselines across different speedup levels (or NFE budgets)?

  3. Could the authors provide additional metrics beyond reconstruction fidelity, such as image quality metrics and VBench scores? Even if TSenCache does not perform as well on these metrics, this would not affect my assessment of the method's effectiveness. The goal is simply to make the experimental evaluation more complete and comprehensive.

4 comments

Thank you for recognizing the importance of final-output error and for identifying specific ways to strengthen the analysis and evaluation. Reviewer nTFS also noted the motivation and improved reconstruction fidelity. We address your concerns below with additional experiments and revised qualitative comparisons.


W1 / Q1. Could a monotonically decreasing weight explain the improvement simply by prioritizing earlier steps?

A generic early-step preference captures part of the terminal-sensitivity effect, while the measured profile provides a further improvement. Earlier perturbations can have a larger final effect because they propagate through more denoising steps. This propagation contributes to terminal sensitivity, but a generic decay captures only part of its timestep dependence. The requested comparison directly tests whether prioritizing earlier steps is sufficient.

We replaced αF\alpha_F with linear and exponential decays using the same maximum and minimum weights as αF\alpha_F, while retaining the local estimator, scheduler factor and online decision rule. All three conditions use FLUX.1-dev / DrawBench-200 at 10 NFE, without a consecutive-reuse cap or forced final refresh.

Terminal weightNFEPSNR ↑SSIM ↑LPIPS ↓
Linear decay1023.0530.79630.2091
Exponential decay1023.7540.77240.2821
Calibrated αF\alpha_F1025.1360.81200.2049

The measured profile remains high during the early steps, decreases sharply, and subsequently plateaus; matching its endpoints alone does not reproduce this shape (Figure 10). Both simple decays exceed the online baselines in Table 1 in PSNR, whereas the measured profile improves further on all three fidelity metrics. Calibration identifies this profile without requiring a search over decay functions (Table 6).


W1, continued. Does the error pattern persist when multiple steps use cached features?

The local–terminal mismatch persists on actual cached trajectories. We used the trajectories and full-computation/reuse decisions recorded during TSenCache generation at 10 NFE for the same 16 prompts as Figure 1. At each step, we created full-computation and reuse branches from the same cached prefix. Both branches then replayed the subsequent recorded decisions while independently evolving their latent and cache states. We measured the same local-output difference Δv\Delta v, next-latent difference Δx\Delta x, and final-image difference ΔI\Delta I as in Figure 1.

Terminal impact remains largest during the early steps and decreases markedly toward the middle of sampling. Local error increases again during the later steps without a comparable increase in final-image error. Figure 11 compares these curves with the isolated-reuse analysis, confirming that the mismatch persists under repeated reuse.


W2 / Q2. Comparisons across different NFE budgets.

We expanded the comparison to eight methods at five NFE budgets on FLUX.1-dev / DrawBench-200, including LeMiCa and BudCache. TSenCache achieves the best or near-best PSNR throughout and outperforms every runtime-adaptive baseline in PSNR:

NFETSenCacheLeMiCaBudCacheSeaCacheDiCacheMagCacheSenCacheTeaCache
720.8420.9420.3119.3119.7019.7320.4316.09
1024.6724.1724.0622.5522.8022.0022.1116.89
1328.3528.0826.3826.7326.5324.9024.6819.27
1630.9630.9927.4229.9228.7028.2928.6220.28
2035.2435.2430.6134.9730.6334.7531.8122.71

Values are PSNR (dB); Table 15 provides the full comparison and measured NFE.

The LTX-Video comparison likewise demonstrates a consistent advantage over SenCache across the tested budgets (nTFS Q2; Table 16).


W2, continued. Uninformative qualitative examples.

We replaced the pizza example with a scissors example in which SenCache alters the handle structure, while TSenCache preserves the full-computation structure (Figure 12). The new Wan2.1 examples have clearly resolved subjects in the full-computation references: SenCache changes the astronaut's helmet and suit (Figure 13) and the child's hairstyle and expression (Figure 14), whereas TSenCache preserves these features.

W3. BudCache and LeMiCa should appear in the main experimental results.

Both methods are now included in the main comparison (Table 1) and the five-budget comparison (Table 15). TSenCache outperforms both methods on all three fidelity metrics at the main operating point. LeMiCa and BudCache optimize a shared schedule offline, whereas TSenCache makes sample-specific online decisions. The fixed/online comparison using the same calibration in nTFS Q1 directly examines the contribution of these decisions.


Q3. Additional image-quality and alignment metrics, including VBench.

We added four image-quality and alignment metrics on FLUX / DrawBench-200 and evaluated all 16 VBench dimensions on Wan2.1 (Tables 13–14):

MethodCLIP Score ↑ImageReward ↑HPSv2 ↑PickScore ↑
Full computation31.280.9310.30522.80
TeaCache31.480.8530.29422.68
DiCache31.280.8530.29622.66
MagCache31.380.8720.30322.69
SeaCache31.360.8740.30022.70
SenCache31.360.8570.30122.71
TaylorSeer31.370.9190.30622.78
LeMiCa31.310.8940.29822.61
BudCache31.470.9080.30622.75
TSenCache31.370.8920.30222.70
Wan2.1, VBench (%)Total ↑Quality ↑Semantic ↑
Full computation81.6985.2967.31
TeaCache81.4885.1166.95
DiCache81.7985.3767.47
MagCache81.7685.3367.51
SeaCache81.6685.1867.60
SenCache81.5385.0867.35
TaylorSeer81.0485.0165.17
TSenCache81.5085.1067.08

These scores have limited reliability for distinguishing caching-induced deviations in this comparison. The outputs retain similar semantic content and layout, yielding closely clustered scores despite differences in reconstruction fidelity. Several caching methods receive higher scores than full computation; a higher score therefore does not necessarily indicate better preservation of the original output.


The monotone-weight comparison and cached-trajectory analysis directly address your concerns about the terminal-error analysis. The expanded baseline comparisons and revised qualitative examples provide broader evidence for the reconstruction-fidelity benefit. We hope these additions address your concerns and would appreciate your consideration of this evidence in the final assessment.

Thank you for the response and additional experiments. I have three following questions:

  1. The original BudCache paper reports PSNR/SSIM/LPIPS of 24.903/0.8374/0.1759 at 10 NFE on FLUX.1-dev, whereas your Table 1 reports 24.065/0.8253/0.1759. Furthermore, BudCache's performance improvement across NFE budgets in Table 15 seems unusually slow, with an increasingly large gap compared to LeMiCa. Could you explain these discrepancies and clarify whether BudCache was independently re-searched at each budget?

  2. Similarly, the original SenCache paper reports PSNR/SSIM/LPIPS of 30.7233/0.9387/0.0390 at 25 NFE on Wan2.1, compared to 26.955/0.8977/0.0620 in your Table 2. Could you clarify the reason for these substantial differences?

  3. For the monotonic-weight ablation, could you clarify the exact implementations of the linear and exponential decays, including how their decay rates were determined, and provide the per-step full-computation/cache frequencies across all 28 steps for all three variants? Also, why does the calibrated variant in Table 6 (25.136/0.8120/0.2049) differ substantially from the main TSenCache result in Table 1 (24.671/0.8363/0.1695) at the same 10 NFE?

Thank you for examining the additional results. We clarify the baseline evaluations and the monotone-weight comparison below.


Q1. BudCache results and independent search at each NFE budget.

Reconstruction scores can differ from those reported in the original paper because different random seeds produce different generated samples and full-computation references, even when the DrawBench prompts are identical. Our evaluation uses the same model, prompts, random seeds and full-computation references for all methods in the experiments, ensuring a fair comparison across methods.

Although the reason for the different improvement rates between BudCache and LeMiCa is unclear, we want to emphasize that we used BudCache's official GitHub code for caching and schedule search and independently optimized a schedule for each budget of 7, 10, 13, 16, and 20 NFE. No schedule was obtained by reusing or truncating the 10-NFE schedule. The saved schedules confirm that each meets its target budget. The final search losses are:

NFE710131620
Final search loss (latent MSE)0.078050.027120.013830.006430.00249

The search loss decreases at every budget, and evaluation PSNR also improves.


Q2. SenCache results on Wan2.1.

Table 2 uses different hyperparameters from the slow setting reported in the SenCache paper. Applying the paper's reported slow-setting hyperparameters to the released implementation did not reproduce its reported 25 NFE: our four-prompt check yielded 31 NFE on average. The paper and the official GitHub README also specify different hyperparameters.

For this reason, our 25-NFE comparison is based on the official GitHub code and README settings, with only the thresholds adjusted to match the target budget. Because of the different operating setting and independently generated full-computation references, our absolute scores can differ from those reported in the paper. All methods in Table 2 are evaluated under identical conditions, ensuring a fair comparison across methods.


Q3. Decay definitions, per-step frequencies, and the difference between Tables 1 and 6.

The decay rates are determined by the calibrated profile's endpoints, without a separate rate search. The first step always uses full computation. For one-indexed steps k=2,…,28k=2,\ldots,28, let uk=(k−2)/26u_k=(k-2)/26, M=max⁡αFM=\max\alpha_F, and m=min⁡αFm=\min\alpha_F. We use

wlinear(k)=(1−uk)M+ukm,w_{\mathrm{linear}}(k)=(1-u_k)M+u_km,

wexp(k)=M(m/M)uk=Mexp⁡[−λuk],λ=log⁡(M/m).w_{\mathrm{exp}}(k)=M(m/M)^{u_k}=M\exp[-\lambda u_k],\qquad \lambda=\log(M/m).

The measured endpoints M=142.1093M=142.1093 and m=1.6049m=1.6049 give λ=4.4835\lambda=4.4835. All three variants share the same local-error estimator, scheduler step sizes, and score-based online decision rule; only the terminal weight changes. Each reuse threshold is tuned to the same target of 10 NFE.

Tables 1 and 6 use different operating configurations. TSenCache follows SenCache in limiting consecutive reuse, using a default limit of 4. As described in Appendix H, the monotone-weight experiment in Table 6 disables auxiliary rules for all three variants to isolate how the terminal weight affects computation allocation and reconstruction fidelity. The calibrated profile achieves the best PSNR, SSIM, and LPIPS among the three weights in this comparison. We have revised the relevant text in Section 4.4 and Appendix H to make this distinction clearer.

To check whether this conclusion also holds under the main configuration, we additionally compared the three weights with auxiliary rules enabled. The calibrated profile again achieves the best fidelity on all three metrics:

Terminal weightNFEPSNR ↑SSIM ↑LPIPS ↓
Linear1022.2760.79990.2120
Exponential1023.7160.81020.1975
Calibrated1024.6710.83630.1695

Figure 10(b) now shows full-computation frequencies at all 28 steps for all three variants, averaged over the 200 DrawBench prompts. Calibrated and exponential weights allocate almost the same amount of full computation to the first nine steps—7.01 and 7.00 evaluations on average—yet the calibrated profile preserves higher fidelity on all three metrics. This indicates that fidelity depends on which steps receive computation, not only on how much computation is concentrated near the start.

Reviewer_nTFSOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper builds on top of the recent works accelerating diffusion models through feature caching. The authors nicely present previous work and their limitations. They propose to use the estimated terminal error, rather than local model-output differences, to guide caching decisions.

The main observation is that similar local errors at different denoising steps can have very different effects on the final output. The proposed method addresses this by combining an estimate of the local error with the scheduler step size and a calibrated sensitivity of the remaining denoising trajectory.

Critical strengths

The paper is well motivated, and the proposed approach is interesting.

I appreciate the evaluation on both image and video generation models. The quantitative and qualitative results are encouraging, showing improved reconstruction fidelity at comparable inference speeds.

Critical weaknesses

The paper is relevant and interesting. Yet, i have few comments that will be framed as question in the next section.

  • offline vs online adaptation. The results are relatively close to those of offline scheduling methods such as LeMiCa and BudCache (Table 8). It is unclear how much of the improvement comes from the terminal-sensitivity profile itself rather than sample-specific online decisions.
  • evaluation across architectures. The experiments cover FLUX.1-dev for image generation and Wan2.1 for video generation which is nice. It is quite easy as a review to ask for more experiments. But in this case, previous work noticed that sensitivity patterns can differ substantially across video architectures. It will be nice to test the methods across other video gen architectures.
  • on consecutive cache reuse. Appendix D nicely studies cache gaps of up to four steps, but only reports correlations with terminal impact. The effect of longer consecutive reuse intervals on reconstruction quality will be interesting.
Questions for authors

To sum up the above comments, here are the questions:

  1. What is the main benefit of online caching compared with offline scheduling? Could the authors compare TSenCache with a fixed schedule derived from the same terminal-sensitivity profile, using the same average NFE? This would help isolate the contribution of online adaptation.
  2. Does TSenCache generalize to other video architectures, such as LTX-Video?
  3. How does TSenCache perform with longer consecutive cache-reuse intervals? Does terminal-sensitivity weighting help maintain reconstruction fidelity as the cache gap increases?
2 comments

Thank you for your positive assessment of the motivation and reconstruction results, and for your questions on online adaptation, architectural generalization, and consecutive reuse. We address each question below with experimental evidence. References to tables and figures refer to the revised manuscript.


Q1. What does online adaptation add over a fixed schedule derived from the same calibration?

Online adaptation improves reconstruction fidelity at the same NFE. We construct a fixed schedule from the average per-step terminal-error measurements obtained during TSenCache's calibration and apply that schedule to every prompt. Both methods use the same FLUX.1-dev / DrawBench-200 inputs and seeds.

MethodPSNR ↑SSIM ↑LPIPS ↓
Fixed schedule (mean measured terminal error)22.9420.81750.1961
TSenCache (online)24.6710.83630.1695

A fixed schedule follows the average error pattern across calibration prompts. TSenCache combines the shared terminal sensitivity with the current sample's local error, allowing the same step to use full computation for one input and cached features for another. The comparison demonstrates the additional fidelity benefit of sample-specific online decisions (Table 11).


Q2. Does TSenCache generalize to another video architecture, such as LTX-Video?

The same online criterion achieves the highest reconstruction fidelity among the six compared methods on LTX-Video. We evaluated LTX-Video 0.9.1 on all 944 VBench prompts at 25 NFE out of 50 sampling steps (Table 2).

LTX-Video, 944 prompts, 25 NFEPSNR ↑SSIM ↑LPIPS ↓
TeaCache19.7340.75010.2991
DiCache23.1130.83160.1943
MagCache23.3070.84040.1871
SeaCache22.3020.81420.2118
SenCache21.2370.82260.2177
TSenCache23.7230.85370.1692

We additionally compared TSenCache with SenCache at four target NFE budgets using the same 944 prompts (Table 16):

MethodNFEPSNR ↑SSIM ↑LPIPS ↓
SenCache2017.560.74140.4172
TSenCache2021.130.79110.2534
SenCache2521.240.82260.2177
TSenCache2523.720.85370.1692
SenCache3122.480.85180.1717
TSenCache3025.970.88730.1204
SenCache3627.330.91110.0880
TSenCache3629.080.92320.0733

TSenCache outperforms SenCache on all three fidelity metrics across the tested budgets, demonstrating that the benefit on LTX-Video extends beyond a single operating point.


Q3. Does terminal-sensitivity weighting help with longer consecutive reuse intervals?

Terminal-sensitivity weighting preserves higher reconstruction fidelity under longer consecutive reuse. We compared TSenCache with the local-error criterion using the same estimator, examining final reconstruction quality and consecutive reuse lengths at 10 NFE (Table 9):

Score (FLUX.1-dev, DrawBench-200)NFEMean longest runMax runPSNR ↑SSIM ↑LPIPS ↓
Local error e^\hat{e}104.92621.9210.80150.2153
TSenCache107.981125.1940.81840.2029

TSenCache achieves the highest fidelity on all three metrics despite allowing longer consecutive reuse intervals. Final quality depends on both the length and position of these intervals. The results support using terminal sensitivity to allocate longer reuse where its downstream effect is smaller, rather than relying on local error alone.

Thanks for the comments. I ll update my rating accordingly.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their constructive feedback. We are encouraged that the reviewers found our approach well motivated (nTFS, 9cwQ), with Reviewer 9cwQ highlighting the importance of guiding caching by the error in the final output rather than local errors. The reviewers also recognized the favorable trade-off between inference speed and reconstruction fidelity on image and video generation (nTFS, yvwR).

Following the reviews, we have revised the paper as follows:

  • Generality across architectures: added LTX-Video, where TSenCache achieves the best fidelity among six caching methods at 25 NFE and outperforms SenCache across budgets (nTFS).
  • More comprehensive comparison: moved LeMiCa and BudCache into the main FLUX comparison and reported results for all seven baselines in Table 15 across five target budgets from 7 to 20 NFE (9cwQ).
  • Controlled analyses: compared online decisions with a fixed schedule derived from the same calibration (nTFS) and the calibrated sensitivity with monotone decays (9cwQ), and added results on calibration-set choice, longer consecutive reuse, cached-trajectory error patterns, and additional quality metrics.

We sincerely appreciate the reviewers' time and effort, which helped us substantially improve the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.