acceptodds
Under review as a conference paper at ICLR 2027

RACE: Refresh-Anchored Cache for Diffusion Acceleration

16 views

Abstract

Flow matching methods have made advanced progress in image and video generation, yet their iterative sampling remains computationally expensive. Cache-based acceleration explores computational redundancy and approximates velocity in certain skipped steps. However, velocity approximation errors accumulate through sampler updates, resulting in the perturbed state for velocity prediction in cache refresh. To improve velocity prediction in cache refresh, the state displacement in the preceding updates should be adjusted. Moreover, we propose that the distribution of state and velocity updates is related along local generation trajectory. This distribution constraint can be easily applied on the estimation of velocity at reuse steps. Building on these insights, we propose **RACE** (Refresh-Anchored Cache), a training-free framework that combines state-guided velocity extrapolation with refresh-anchored integral correction. RACE uses the evolving state to guide velocity prediction and refresh-endpoint errors to compensate for state deviations accumulated during reuse steps. Each refresh step thus serves as an anchor both for subsequent velocity approximation and preceding state correction. RACE is compatible with existing cache schedulers and requires no additional network computation or hyper-parameters tuning. Extensive experiments on FLUX.1-dev, Wan 2.1, HunyuanVideo, and MiniMax-H3 demonstrate that RACE achieves state-of-the-art latency-fidelity trade-offs. We will release our code soon.

Primary area:
generative models
Keywords:
Diffusion Model Acceleration, Cache-Based Inference, Flow Matching

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_FzMROpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes RACE, a training-free method using state-guided velocity extrapolation with a refresh-anchored correction mechanism. The authors point out that cache-based velocity reuse will lead to error accumulation as the number of skipped steps increases. They observe an approximately linear relationship between state and velocity changes along local denoising trajectories. Based on the observation, RACE utilizes the state-velocity relationship of historical steps to estimate the velocity of skipped steps and compares the actual velocity and predictions to estimate accumulated errors and make corrections. The paper provides abundant experiments to demonstrate these claims and the improvement.

Critical strengths
  1. Easy and Useful: The RACE could be utilized in other training-free, cache-based diffusion acceleration methods orthogonally and shows promising improvements (30+dB PSNR results are really impressive), which provides good scalability.
  2. Clear theory and numerical analysis: The paper discusses the near-linear relationship between state and velocity changes and error accumulation with clear numerical analysis that is easy to understand.
  3. Innovative: Providing a novel adjustment method to handle the error accumulation problems in cache-based diffusion acceleration, which will inspire more attempts based on these discoveries and techniques.
Critical weaknesses
  1. Video Quality Evaluation: For 30 dB PSNR configurations, current indicators are powerful enough to demonstrate the closure between the accelerated and original trajectories. However, a VBench result may strengthen the argument for the 20 dB results.
  2. Generalization of the Near-linear Assumption: It shows that the proposed method still works quite well on backbones that don’t exactly obey the assumption. Additional discussion of the method's generalization ability in non-ideal situations would be helpful.
  3. Without Indicating the Improvement Directly: Another table showing the improvement of the refresh mechanism compared with previous caching methods such as TeaCache, EasyCache, SenCache, and SeaCache will help strengthen the point.
Questions for authors
  1. Is there an optimal configuration of this caching and refreshing process for each backbone, or can the optimal configuration be determined by numerical analysis?
  2. Will it be helpful when utilizing other approximate functions other than linear? If not, what might be the reason?
3 comments
AuthorsOpenReview ↗

Part1: W1, W2

Thank you for your valuable assessment and suggestions.

W1: Video Quality Evaluation.

We additionally report VBench Quality, Semantic, and Total scores for Original, reduced-step sampling, and the caching methods on Wan2.1, HunyuanVideo, and MiniMax-H3. These complement the reference-fidelity evaluation using PSNR, LPIPS, and SSIM.

RACE achieves the highest Total scores among the listed caching methods at both computational budgets on all three models. Compared with SeaCache, RACE improves Total by 0.511 and 0.674 on Wan2.1, 0.239 and 0.098 on HunyuanVideo, and 0.156 and 0.128 on MiniMax-H3, at the high and low computational budgets, respectively. These results show that RACE maintains overall video quality while improving reference fidelity.

MethodWan2.1 ParameterWan2.1 PSNRWan2.1 QualityWan2.1 SemanticWan2.1 Total
Original-—84.08066.62880.589
25 steps-14.7884.08765.74680.419
15 steps-13.2882.90265.74779.471
High computational budget
TeaCacheδ = 0.0923.5384.21466.69480.710
TaylorSeerS = 218.3884.43966.18580.788
SenCacheε = 0.0129.6284.50065.40380.681
DiCacheδ = 0.1027.2384.56365.40880.732
SeaCacheδ = 0.2029.3883.93166.45180.435
RACEδ = 0.2031.1584.46666.86680.946
Low computational budget
TeaCacheδ = 0.2019.1683.61465.48979.989
TaylorSeerS = 414.2083.43560.79278.907
SenCacheε = 0.1021.6684.01566.49480.511
DiCacheδ = 0.2023.6384.25165.33380.468
SeaCacheδ = 0.3523.5383.47265.95479.969
RACEδ = 0.3524.4084.11066.77480.643
MethodHunyuanVideo ParameterHunyuanVideo PSNRHunyuanVideo QualityHunyuanVideo SemanticHunyuanVideo Total
Original-—83.17872.15680.973
25 steps-20.1282.99171.95280.783
15 steps-17.2582.07070.98979.854
High computational budget
TeaCacheδ = 0.1223.5583.03970.94980.621
TaylorSeerS = 224.9383.31571.53380.959
SenCacheε = 0.2830.1983.23571.48880.886
DiCacheδ = 0.06532.4083.01571.23480.659
SeaCacheδ = 0.2632.4483.21271.92680.955
RACEδ = 0.2635.7183.41172.32481.194
Low computational budget
TeaCacheδ = 0.1620.7482.92371.52480.643
TaylorSeerS = 321.0183.32870.58180.779
SenCacheε = 0.4527.5283.12670.26580.554
DiCacheδ = 0.1327.8483.10171.95680.872
SeaCacheδ = 0.3528.1683.09572.25480.927
RACEδ = 0.3531.1883.20072.32581.025
MethodMiniMax-H3 ParameterMiniMax-H3 PSNRMiniMax-H3 QualityMiniMax-H3 SemanticMiniMax-H3 Total
Original-—84.47964.68580.520
25 steps-17.7884.03865.68380.367
15 steps-15.7083.82464.83480.026
High computational budget
TeaCacheδ = 0.2318.6684.70565.28280.821
TaylorSeerS = 218.9084.37864.75980.454
SenCacheε = 0.3723.4784.73964.14380.620
DiCacheδ = 0.0722.2884.60465.22680.728
SeaCacheδ = 0.2823.3484.74564.42280.681
RACEδ = 0.2825.1684.74565.20580.837
Low computational budget
TeaCacheδ = 0.4217.3483.74465.62980.121
TaylorSeerS = 413.9577.74448.98971.993
SenCacheε = 2.2017.7484.10965.25480.338
DiCacheδ = 0.3117.8784.48864.36080.462
SeaCacheδ = 0.5517.6984.10665.92180.469
RACEδ = 0.5519.7284.61764.51580.597

W2: Generalization of the Near-linear Assumption.

The method does not require the entire sampling trajectory to be globally linear; it relies primarily on local approximations within short reuse intervals. SGVE models the local state–velocity relationship using the two most recent full-computation refresh points and determines the extrapolation coefficient by projecting the current state onto the latest displacement direction. Each refresh updates the local secant, allowing predictions to adapt as the sampling trajectory evolves. In Appendix E, the second and third principal components exhibit more pronounced curvature, yet SGVE still tracks the local trajectories well. This indicates that the mechanism also applies when the linear relationship is imperfect.

AuthorsOpenReview ↗

Part2: W3, Q1, Q2

W3: Indicating the Improvement for different cache scheduler Directly.

Compared with TeaCache, EasyCache, SenCache, and SeaCache, RACE focuses on reducing generation errors and modifying the generation trajectory. It is compatible with different cache schedulers. Because RACE changes the sampling trajectory, the actual refresh locations may also change. On FLUX.1-dev, we compared these four schedulers with and without RACE. Adding RACE improves PSNR by 0.479, 2.911, 3.293, and 3.282 dB, respectively, with improvements in the other metrics as well. These results demonstrate RACE’s effectiveness across cache schedulers. We observed that TeaCache still selects reuse in some intervals that deviate from local linearity, partly limiting RACE’s correction gains. Its improvements are therefore consistent but relatively modest.

MethodTFLOPsPSNR ↑LPIPS ↓SSIM ↑PSNR gain
TeaCache921.15920.6310.20610.8085—
TeaCache + RACE930.44021.1100.19380.8218+0.479
EasyCache931.01128.3940.07000.9256—
EasyCache + RACE927.32931.3050.03640.9541+2.911
SenCache930.45329.0690.06060.9309—
SenCache + RACE933.53132.3620.03490.9552+3.293
SeaCache930.53629.5780.05460.9358—
SeaCache + RACE927.41232.8600.02880.9620+3.282

Q1: Is there an optimal configuration of this caching and refreshing process for each backbone?

We primarily use the hyperparameter δ to control the cache scheduler’s reuse aggressiveness. We varied δ in SeaCache + RACE on FLUX.1-dev and evaluated computational cost and image quality using 200 DrawBench prompts:

δTFLOPs ↓PSNR ↑LPIPS ↓SSIM ↑
0.18751007.27034.690.02130.9705
0.2250930.51332.860.02880.9620
0.2625851.43031.160.04040.9491
0.3000776.99929.870.04980.9394
0.3750670.00426.900.08460.9046
0.5250556.03124.380.11820.8745
0.6000483.92622.340.15640.8445

Increasing δ makes reuse more aggressive, reducing both computational cost and output fidelity. We further fitted the relationships among δ, computational cost, and PSNR to quantify the quality cost of adjusting δ and guide the choice of computational budget and threshold.

Across the seven operating points, the fitted relationship is PSNR = −28.5406δ + 38.9757 (R² = 0.9683). In terms of computational cost F, measured in TFLOPs, the fit is PSNR = 0.023348F + 11.2862 (R² = 0.9975), showing a more pronounced linear relationship. R² is the coefficient of determination which measures how much variation in the observations is explained by the fitted model; values closer to 1 indicate a better fit. Thus, δ can control computational cost and fidelity, while the fitted slopes quantify the magnitude of this trade-off.

Overall, choosing δ involves a clear compute–fidelity trade-off, and no single δ is optimal for all budgets.

Q2: Will it be helpful when utilizing other approximate functions other than linear?

We examined nonlinear approximations in two ways: (1) incorporating older refresh points into polynomial fitting for SGVE; and (2) using higher powers of β to weight the per-step error estimates in RAIC.

For SGVE, we compared history windows containing 2, 3, and 4 refresh points, fitting a nonlinear polynomial in each window for per-step velocity prediction. We evaluated these settings on FLUX.1-dev using DrawBench prompts:

Number of historical pointsTFLOPs ↓PSNR ↑LPIPS ↓SSIM ↑
2938.14232.93890.02850.9623
3937.58431.67720.03060.9606
4937.58430.80360.03150.9610

Increasing the history from 2 to 3 or 4 points reduces PSNR by approximately 1.262 and 2.135 dB, without improving LPIPS or SSIM. This indicates that extending the history does not yield more accurate predictions of the current velocity. SGVE relies on the local state–velocity relationship, while older refresh points may come from trajectory segments whose dynamics have already changed. The two most recent refresh points therefore better reflect current local changes.

For RAIC, we replaced the linear weight β with β^1.5、β² and β³ to evaluate nonlinear weighting. These experiments also used FLUX.1-dev and 200 DrawBench prompts, with each variant’s threshold set according to the computational budget:

RAIC weightTFLOPs ↓PSNR ↑LPIPS ↓SSIM ↑
β938.14232.93890.02850.9623
β1.5β^{1.5}937.95632.66540.02920.9613
β²937.95632.32710.03050.9599
β³937.7731.98710.03200.9581

Increasing the exponent of β from 1 to 1.5, 2, and 3 progressively reduces PSNR and worsens both LPIPS and SSIM.

Overall, neither quadratic fitting nor correction weights with higher powers improved fidelity. For SGVE, older history may shift the fit away from current local changes. For RAIC, higher powers of β reduce compensation at intermediate steps, whereas the error diagnostics in Fig. 4(a) indicates growth closer to linear. We therefore retain two-point SGVE and linear RAIC.

Thanks for your rebuttal. My concerns are well addressed.

Reviewer_6A87OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces RACE, a training-free acceleration framework for flow matching and diffusion models. Traditional cache methods suffer from velocity approximation errors that propagate across skipped steps and distort subsequent state updates. RACE solves this by combining two components: (1) SGVE, which uses the linear relationship between state and velocity updates across local generation trajectories to guide velocity prediction at reuse steps, and (2) RAIC, which leverages refresh-endpoint velocity errors to estimate and compensate for accumulated state deviations. Experiments demonstrate that RACE achieves good performance across several models.

Critical strengths
  1. The observation and formulation of the local linear relation between state-velocity updates provide a mathematically sound foundation for velocity extrapolation.
  2. The method is evaluated across a broad spectrum of image and video generation models under multiple computational budgets.
  3. The proposed method is training-free, and compatible with existing cache schedulers.
Critical weaknesses
  1. The “no hyperparameters” and “same compute” claims are overstated. The method depends on the cache scheduler and its threshold, which is tuned per model and budget.
  2. Although RACE avoids additional network parameter computations, computing inner products and least-squares projections across high-dimensional latent states introduces tensor operation overhead. This could impact latency gains in ultra-low latency regimes.
  3. The effectiveness of RAIC depends on the accuracy of the interval coordinate approximation. If the interval between refresh steps is too wide, the linear growth assumption of interior errors may degrade.
Questions for authors
  1. The authors are suggested to provide the measured memory and time overhead introduced by computing high-dimensional inner products and projections for SGVE during iterative sampling.
  2. How robust is the integral correction when facing highly irregular or adaptive step-size schedules generated by advanced ODE solvers or dynamic cache schedulers?
  3. Please report the sensitivity to δ and window size.
3 comments
AuthorsOpenReview ↗

Part1: W1, W2, W3, Q1.

Thank you for your valuable assessment and suggestions.

W1: The “no hyperparameters” and “same compute” claims.

We agree that both claims require qualification. RACE retains the cache scheduler and its threshold δ, which must be set for each model and target budget. “No additional tunable hyperparameters” means that, given this configuration, SGVE and RAIC add no tunable extrapolation or correction coefficients. SGVE’s projection coefficient is computed analytically from the current state and the two latest refresh states, while RAIC’s weights are determined by actual sampling times.

“Same compute” means that both modules reuse existing refresh results for prediction and correction without additional full-network evaluations. The few added inner products and vector operations incur only minor overhead. RACE can therefore improve output fidelity at the same target budget. Computational, runtime, and GPU memory overheads are reported below.

W2 / Q1: Memory and time overhead introduced by SGVE during iterative sampling.

Using FLUX.1-dev on DrawBench and Wan2.1 1.3B on VBench, we compared full 50-step sampling with SeaCache, SeaCache + SGVE, and RACE at two budget levels. At comparable budgets, the three accelerated variants use identical refresh decisions to isolate prediction and correction overhead. The table reports FLOPs for the complete generation pipeline and runtime, peak GPU memory, and SGVE projection time during sampling.

ModelMethodTFLOPsPeak Memory, mean / max GiBGiBLatency, mean ± std ssAverage total runtime of SGVE msmsSGVE runtime per call ms/callms/call
FLUX.1-devFull 50 steps186132.09 / 32.337.8 ± 0.05-—
FLUX.1-devSeaCache948.832.18 / 32.404.09 ± 0.12-—
FLUX.1-devSeaCache + SGVE948.432.16 / 32.384.03 ± 0.1515.140.41
FLUX.1-devRACE948.432.10 / 32.404.03 ± 0.1415.160.41
FLUX.1-devSeaCache565.132.16 / 32.402.59 ± 0.05-—
FLUX.1-devSeaCache + SGVE564.732.13 / 32.382.52 ± 0.0617.070.39
FLUX.1-devRACE564.732.14 / 32.402.53 ± 0.0617.080.40
Wan2.1 1.3BFull 50 steps993418.25 / 19.1254.45 ± 0.16-—
Wan2.1 1.3BSeaCache4879.419.39 / 20.2626.64 ± 0.33-—
Wan2.1 1.3BSeaCache + SGVE4879.119.47 / 20.3326.59 ± 0.348.020.16
Wan2.1 1.3BRACE4879.119.56 / 20.6626.58 ± 0.348.110.15
Wan2.1 1.3BSeaCache3063.119.92 / 20.6116.67 ± 0.72-—
Wan2.1 1.3BSeaCache + SGVE3062.819.99 / 20.6916.63 ± 0.7210.880.14
Wan2.1 1.3BRACE3062.819.61 / 21.0416.64 ± 0.729.760.12

SGVE averages 0.39–0.41 ms per projection on FLUX and 0.12–0.16 ms on Wan, totaling less than 20 ms and accounting for less than 0.7% of total latency. GPU memory changes vary by model and budget but remain small; both means and maxima are reported above.

W3: The effectiveness of RAIC over reuse intervals with different widths.

Longer intervals may exhibit greater curvature and directional changes, making endpoint errors less representative. Appendix G shows that, when the error’s second derivative is bounded, the integral estimation error bound grows cubically with interval length. Interval length is therefore a key factor in approximation reliability.

On FLUX.1 with DrawBench, we measured state L2 errors before and after RAIC at 50% and 30% budgets. For every SGVE reuse interval bounded by consecutive full-computation points, we started from the same interval-initial state and performed full computation at every step to obtain an endpoint reference. We then compared the endpoint L2 errors of uncorrected SeaCache + SGVE and RAIC-corrected states against this reference. At both budgets, RAIC provides relatively consistent correction for intervals with at least two reuse steps, reducing state error in over 70% of these intervals.

≈50% computational budget
Reuse countPercentage in all intervalsSeaCache + SGVE errorSeaCache + SGVE + RAIC errorError ReductionPercentage of intervals with reduced endpoint error
144%0.003070.002944.4%62.70%
253%0.005030.0038723.10%91.10%
33%0.009760.00638.50%100.00%
≈30% computational budget
Reuse countPercentage in all intervalsSeaCache + SGVE errorSeaCache + SGVE + RAIC errorError ReductionPercentage of intervals with reduced endpoint error
122.5%0.004320.004251.70%61.00%
29.7%0.00450.0036518.70%87.50%
324.1%0.012630.011548.60%71.10%
45.6%0.023370.0187619.70%73.00%
526.3%0.022740.0131542.20%100.00%
611.8%0.027720.0148746.30%100.00%
AuthorsOpenReview ↗

Part2: Q2, Q3.

Q2: The robustness in higher-order solvers and irregular refresh schedules

RACE retains fidelity gains over the corresponding caching baselines under the tested higher-order solvers, irregular refresh intervals, and adaptive step-size schedules.

We added Heun experiments on FLUX.1-dev using 200 DrawBench prompts and full 50-step Heun outputs as references:

MethodTFLOPs ↓PSNR ↑LPIPS ↓SSIM ↑
Full 50-step Heun3684.42———
25-step Heun1823.6016.87200.31700.7314
SeaCache1824.1829.66220.05420.9386
RACE1824.1333.44190.02800.9651
15-step Heun1079.2714.59110.42160.6528
SeaCache1080.0121.89540.17920.8288
RACE1079.9523.67680.12700.8695

RACE improves PSNR over SeaCache by 3.78 and 1.78 dB at the high and low budgets, respectively, supporting its effectiveness with Heun.

We use SeaCache as the primary adaptive step-size scheduler; its nonuniform refresh schedule adapts to denoising. Experiments with SeaCache, SenCache, and EasyCache show substantial gains across all three, supporting RACE’s robustness across adaptive step-size schedules.

MethodTFLOPsPSNR ↑LPIPS ↓SSIM ↑PSNR gain
EasyCache931.01128.3940.07000.9256—
EasyCache + RACE927.32931.3050.03640.9541+2.911
SenCache930.45329.0690.06060.9309—
SenCache + RACE933.53132.3620.03490.9552+3.293
SeaCache930.53629.5780.05460.9358—
SeaCache + RACE927.41232.8600.02880.9620+3.282

Q3: Sensitivity to δ and window size

δ controls reuse aggressiveness and the scheduler’s refresh ratio; Fig. 7aa reports PSNR across refresh ratios. We varied δ in SeaCache + RACE on FLUX.1-dev and evaluated computational cost and image quality using 200 DrawBench prompts:

δTFLOPs ↓PSNR ↑LPIPS ↓SSIM ↑
0.18751007.27034.690.02130.9705
0.2250930.51332.860.02880.9620
0.2625851.43031.160.04040.9491
0.3000776.99929.870.04980.9394
0.3750670.00426.900.08460.9046
0.5250556.03124.380.11820.8745
0.6000483.92622.340.15640.8445

Increasing δ makes reuse more aggressive, reducing both computational cost and output fidelity. To quantify this quality cost and guide budget and threshold selection, we fitted the relationships among δ, compute, and PSNR.

Across the seven operating points, the fits are PSNR = −28.5406δ + 38.9757 R2=09˙683R² = 0\.9683 and PSNR = 0.023348F + 11.2862 R2=09˙975R² = 0\.9975, where F is computational cost in TFLOPs. R², the coefficient of determination, measures the variation in the observations explained by the fit; values closer to 1 indicate better fits. Thus, δ controls a clear compute–fidelity trade-off, with the fitted slopes quantifying the resulting changes. No single δ is optimal for all budgets.

We also ablated SGVE’s history length using 2, 3, and 4 refresh points. Window size denotes the span covered by historical refresh points; multiple historical observations are used to construct a nonlinear prediction function:

Historical refresh pointsTFLOPs ↓PSNR ↑LPIPS ↓SSIM ↑
2938.14232.93890.02850.9623
3937.58431.67720.03060.9606
4937.58430.80360.03150.9610

The secant through two consecutive refresh points performs best in this experiment; a longer history does not improve results. See W3 for diagnostics by reuse-interval length.

Reviewer_6A87OpenReview ↗

addressed some of my concerns

Thanks for the detailed response and additional results. The feedback has addressed some of my concerns. I thus would like the maintain my rating of weak acceptance.

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

This paper introduces RACE (Refresh-Anchored Cache), a training-free framework for cache-based acceleration, evaluated on flow-matching image and video models. The idea is to improve velocity predictions at reuse steps while also correcting errors that have already accumulated in the latent state. State-Guided Velocity Extrapolation (SGVE) uses the current state and the two most recent refresh observations to predict velocity through a local state–velocity secant. Refresh-Anchored Integral Correction (RAIC) uses the prediction error measured at the next refresh to estimate the time-weighted velocity error over the preceding reuse interval, assuming roughly linear error growth with sampling time. This correction is added to the state update without another full-network evaluation. Results on FLUX.1-dev, Wan 2.1, HunyuanVideo, and MiniMax-H3 show better agreement with full-computation outputs than the compared caching baselines at broadly comparable budgets.

Critical strengths
  1. A simple approach with consistent practical gains. I like that RACE treats velocity prediction and accumulated state error as two related but distinct problems. Both components use information already available during sampling, need no training or extra full-network evaluations, and work with several cache schedulers. The improvements in reference fidelity across four image and video models make this a useful contribution to cache-based acceleration.

  2. The components are motivated by observations and backed by ablations. The state–velocity plots give some intuition for SGVE, and Figure 4 provides full-dimensional measurements supporting the endpoint error model on FLUX.1-dev. Table 6 is particularly helpful: PSNR increases from 29.63 dB with neither component to 31.38 dB with SGVE alone, 31.85 dB with RAIC alone, and 32.96 dB with both.

Critical weaknesses

None

Questions for authors
  1. Can you isolate RAIC's effect on the trajectory? Building on Figures 3 and 8, could you compare full computation, SGVE only, and SGVE + RAIC using identical initial noise, conditioning, time grids, and a fixed shared refresh schedule for the two accelerated variants? A shared PCA plot would be useful, but I would put more weight on full-dimensional relative state errors across prompts and sampling steps. At refreshes, please compare the next-step state with and without the added correction against the reference at the same time, so that normal sampling progress is not confused with the correction's effect. This would help clarify the gap between estimating accumulated residuals and reducing actual trajectory deviation.

  2. How reliable is the endpoint error model beyond the setting in Figure 4? RAIC approximates the reuse-step error as ej≈βjebe_j \approx \beta_j e_b. Figure 4 supports this on FLUX.1-dev, with SGVE enabled and RAIC disabled, while Appendix G gives a conditional interpolation-error bound. Could you show how the fit varies by sampling stage and reuse-interval length, and whether similar behavior holds on a video model with RAIC active? I would also like to know how often the correction increases state deviation when intermediate errors change direction or do not grow roughly linearly.

  3. Do the reference-fidelity gains also preserve video quality? PSNR, LPIPS, and SSIM are reasonable measures of agreement with the original model, but they do not fully capture motion quality or temporal consistency. Although the video prompts come from VBench, I could not find reported VBench quality scores. Could you provide relevant VBench scores for the original model, RACE, and a strong caching baseline at comparable budgets, especially under aggressive reuse?

3 comments
AuthorsOpenReview ↗

Part1: Q1, Q2

Thank you for your valuable assessment and suggestions.

Q1: Can you isolate RAIC's effect on the trajectory?

Using identical initial noise, conditioning, and time grids, we performed PCA analyses of full computation, SGVE only, and SGVE + RAIC, with the two accelerated variants sharing a fixed refresh schedule (visualizations in Appendix L.3 of the updated PDF). States and velocities were projected separately using PCA. We observed that, with SGVE alone, the state–velocity evolution during reuse deviates from the full-computation reference. Adding RAIC brings the trajectory closer to this reference through corrections at refresh points, illustrating its role in correcting deviations accumulated during reuse.

To quantitatively evaluate the errors, we further measured state L2 errors before and after RAIC on FLUX.1 using DrawBench at 50% and 30% computational budgets. We extracted all reuse intervals using SGVE approximations, each bounded by two consecutive full-computation points. Starting from the same state at the beginning of each interval, we performed full computation at every step to obtain a reference endpoint state. We then measured the endpoint L2 errors of SeaCache + SGVE before and after RAIC correction relative to this reference. As shown below, RAIC provides relatively consistent correction at both budgets when reuse occurs at least twice, reducing state error in over 70% of these intervals.

≈50% computational budget
Reuse countPercentage in all intervalsSeaCache + SGVE errorSeaCache + SGVE + RAIC errorError ReductionPercentage of intervals with reduced endpoint error
144%0.003070.002944.4%62.7%
253%0.005030.0038723.1%91.1%
33%0.009760.006038.5%100.0%
≈30% computational budget
Reuse countPercentage in all intervalsSeaCache + SGVE errorSeaCache + SGVE + RAIC errorError ReductionPercentage of intervals with reduced endpoint error
122.5%0.004320.004251.7%61.0%
29.7%0.004500.0036518.7%87.5%
324.1%0.012630.011548.6%71.1%
45.6%0.023370.0187619.7%73.0%
526.3%0.022740.0131542.2%100.0%
611.8%0.027720.0148746.3%100.0%

Q2: How reliable is the endpoint error model beyond the setting in Figure 4?

We added diagnostics analogous to Fig. 4 on HunyuanVideo, aggregated over the full VBench prompt set, and compared projection coefficients and directional alignment with RAIC enabled and disabled on FLUX.1-dev and HunyuanVideo (visualizations in Appendix L.1 of the updated PDF). This extends the diagnostics to video models and sampling trajectories with correction enabled.

As shown in Fig.10(a)(b), HunyuanVideo exhibits error patterns similar to those in Fig. 4: the projection coefficient qj=⟨ej,eb⟩/∣eb∣22q_j=\langle e_j,e_b\rangle/|e_b|_2^2 increases with the interval coordinate βj\beta_j and closely follows qj=βjq_j=\beta_j, while the mean cosine similarity between interior and endpoint errors is approximately 0.75–0.90. These observations support ej≈βjebe_j\approx\beta_j e_b in terms of projection magnitude and directional alignment, providing a basis for estimating errors within video-model reuse intervals from endpoint residuals.

As shown in Fig.10(c) Since βj<1\beta_j<1 inside the interval, directly copying the endpoint error overestimates the intermediate error component along that direction, especially near the interval start. RAIC therefore scales the endpoint residual by βj\beta_j and accumulates the correction using the actual step sizes. With the no-correction baseline normalized to 1, directly copying the endpoint error yields pointwise and interval error MSEs of 1.48 and 1.69, whereas RACE reduces them to 0.28 and 0.29. This shows that modeling error growth within the interval is crucial for accurately estimating accumulated errors.

As shown in Fig.11, keeping SGVE enabled and toggling only RAIC yields broadly similar projection-coefficient and cosine-similarity curves for both models. HunyuanVideo’s projection curves largely overlap and remain close to the linear reference, while the local fluctuations in the middle and later portions of the FLUX.1-dev curves appear in both settings. Directional alignment increases as the interval progresses in both models. Thus, after RAIC changes the sampling trajectory, the endpoint residual retains useful directional information. This supports using ej≈βjebe_j\approx\beta_j e_b when SGVE and RAIC operate jointly.

The table in Q1 also quantifies cases where RAIC increases actual state deviation. With dynamic schedulers such as SeaCache, longer reuse intervals often exhibit consistent velocity directions, allowing RAIC to substantially reduce errors in most intervals. Shorter intervals tend to involve greater curvature or abrupt changes, where RAIC still improves over half the cases. Although some corrections increase state deviation, the absolute errors are typically small, limiting their impact on subsequent sampling.

Q3:Do the reference-fidelity gains also preserve video quality?

We additionally report VBench scores for Original, reduced-step sampling, and the caching methods on Wan2.1, HunyuanVideo, and MiniMax-H3 to complement the reference-fidelity evaluation using PSNR, LPIPS, and SSIM.

RACE achieves the highest Total scores among the listed caching methods at both computational budgets on all three video models. Compared with SeaCache, RACE improves Total by 0.511 and 0.674 on Wan2.1, 0.239 and 0.098 on HunyuanVideo, and 0.156 and 0.128 on MiniMax-H3, at the high and low computational budgets, respectively. These results show that RACE maintains overall video quality while improving reference fidelity.

MethodWan2.1 ParameterWan2.1 QualityWan2.1 SemanticWan2.1 Total
Original-84.08066.62880.589
25 steps-84.08765.74680.419
15 steps-82.90265.74779.471
High computational budget
TeaCacheδ = 0.0984.21466.69480.71
TaylorSeerS = 284.43966.18580.788
SenCacheε = 0.0184.50065.40380.681
DiCacheδ = 0.1084.56365.40880.732
SeaCacheδ = 0.2083.93166.45180.435
RACEδ = 0.2084.46666.86680.946
Low computational budget
TeaCacheδ = 0.2083.61465.48979.989
TaylorSeerS = 483.43560.79278.907
SenCacheε = 0.1084.01566.49480.511
DiCacheδ = 0.2084.25165.33380.468
SeaCacheδ = 0.3583.47265.95479.969
RACEδ = 0.3584.11066.77480.643
MethodHunyuanVideo ParameterHunyuanVideo QualityHunyuanVideo SemanticHunyuanVideo Total
Original-83.17872.15680.973
25 steps-82.99171.95280.783
15 steps-82.07070.98979.854
High computational budget
TeaCacheδ = 0.1283.03970.94980.621
TaylorSeerS = 283.31571.53380.959
SenCacheε = 0.2883.23571.48880.886
DiCacheδ = 0.06583.01571.23480.659
SeaCacheδ = 0.2683.21271.92680.955
RACEδ = 0.2683.41172.32481.194
Low computational budget
TeaCacheδ = 0.1682.92371.52480.643
TaylorSeerS = 383.32870.58180.779
SenCacheε = 0.4583.12670.26580.554
DiCacheδ = 0.1383.10171.95680.872
SeaCacheδ = 0.3583.09572.25480.927
RACEδ = 0.3583.20072.32581.025
MethodMiniMax-H3 ParameterMiniMax-H3 PSNRMiniMax-H3 QualityMiniMax-H3 SemanticMiniMax-H3 Total
Original-—84.47964.68580.520
25 steps-17.7884.03865.68380.367
15 steps-15.7083.82464.83480.026
High computational budget
TeaCacheδ = 0.2318.6684.70565.28280.821
TaylorSeerS = 218.9084.37864.75980.454
SenCacheε = 0.3723.4784.73964.14380.62
DiCacheδ = 0.0722.2884.60465.22680.728
SeaCacheδ = 0.2823.3484.74564.42280.681
RACEδ = 0.2825.1684.74565.20580.837
Low computational budget
TeaCacheδ = 0.4217.3483.74465.62980.121
TaylorSeerS = 413.9577.74448.98971.993
SenCacheε = 2.2017.7484.10965.25480.338
DiCacheδ = 0.3117.8784.48864.36080.462
SeaCacheδ = 0.5517.6984.10665.92180.469
RACEδ = 0.5519.7284.61764.51580.597

Thank you for the detailed rebuttal and additional experiments, which have largely addressed my concerns.

The additional trajectory analyses and VBench results provide stronger evidence supporting the effectiveness of RACE.

Overall, I will maintain my rating of 3.

Reviewer_1cA5OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RACE, a training-free caching method for accelerating diffusion models, aiming to reduce image and video generation costs while preserving full-computation outputs. It combines two components: SGVE predicts velocities at reuse steps from the states and velocities observed at the two latest full-computation steps, while RAIC uses the prediction error observed at the next refresh to estimate and compensate for accumulated state deviations. The central motivation is that improving the current prediction does not automatically remove errors accumulated earlier. Experiments on four models show improved trade-offs between computational cost and reference fidelity under the reported configurations, with ablations supporting both components.

Critical strengths
  1. RACE requires neither model retraining nor additional full-network evaluations. The paper also describes an implementation that avoids storing intermediate errors for the entire interval, demonstrating consideration of practical memory overhead in addition to computational savings.

  2. Cross-model experiments and ablations provide empirical support. Improvements are consistent across four image/video models. The ablations show benefits from both components individually and further improvement when combined.

Critical weaknesses

1.RAIC compares the full-network and predicted velocities at the endpoint reached after reuse steps, then uses this difference to estimate earlier accumulated errors. However, this endpoint may already differ from the corresponding state reached through full computation at every step. The resulting correction therefore does not automatically equal the actual state deviation.

2.Figure 4 mainly covers one image model and configuration with RAIC disabled. The main experiments include several video models and different budgets. Further evidence is therefore needed to establish whether the same error behavior holds during sampling with both SGVE and RAIC enabled.

Questions for authors
  1. Regarding Weakness 1, could the authors explain how the velocity difference measured at a potentially shifted endpoint can estimate the accumulated state deviation relative to full computation? Under what conditions is this approximation reliable?

  2. Do the error patterns in Figure 4 hold under other settings? Could similar diagnostics be extended to at least one video model? It would be helpful to examine sampling with both SGVE and RAIC enabled.

  3. Regarding the grid search in Section 5.1, please clarify how the target budgets and baseline parameters are determined, including the grid-search ranges and configuration selection criteria. And explain why the low budget is described as ~30% while the HunyuanVideo RACE ratio in Table 2 is ~40% (5,884 vs. 14,703 TFLOPs).

3 comments
AuthorsOpenReview ↗

Part1: W1,Q1.

Thank you for your valuable assessment and suggestions.

W1 / Q1: Reliability of Estimating Accumulated State Deviation from the Velocity Difference at a Potentially Shifted Endpoint.

The endpoint velocity residual does not equal the state deviation relative to the full-computation trajectory. To clarify when RAIC is reliable, we distinguish local velocity errors from trajectory state deviations.

Let xj∗x_j^* and xjx_j denote the full-computation and accelerated trajectories, respectively, with state deviation dj=xj∗−xjd_j=x_j^*-x_j. Let$u_j$ be the velocity actually used for sampling and ϵj=vθ(xj,tj)−uj\epsilon_j=v_\theta(x_j,t_j)-u_j the local velocity error. At reuse steps, ϵj=ej\epsilon_j=e_j; at refresh steps, ϵj=0\epsilon_j=0. The diagnostic endpoint residual$e_b=v_\theta(x_b,t_b)-\hat v_b$ is computed using the pre-refresh predictor, with both velocities evaluated at the same accelerated state. Let cjc_j denote the additional state correction, which is zero at reuse steps. Taking Euler updates as an example, $d_{j+1}=d_j+\Delta t_j\epsilon_j+\Delta t_j[v_\theta(x_j^*,t_j)-v_\theta(x_j,t_j)]-c_j. $

After incorporating the correction at the interval start into the initial deviation, the state deviation before the next endpoint correction consists of three components: the initial deviation, the integral of local velocity errors, and the contribution from vector-field changes caused by trajectory displacement. RAIC approximates the second component using C^b=∑j=ab−1Δtjβjeb\hat C_b=\sum_{j=a}^{b-1}\Delta t_j\beta_j e_b . Thus, endpoint displacement does not prevent measuring the velocity prediction error at that state; this residual estimates error accumulation along the current accelerated trajectory. The estimate is more reliable when errors within the interval are directionally consistent, vary approximately linearly with actual sampling time, and the vector field is stable under state perturbations. Appendix G further bounds the residual integral estimation error, with the bound growing cubically with interval length.

For quantitative evaluation, we measured state L2 errors before and after RAIC on FLUX.1 using DrawBench at 50% and 30% computational budgets. We extracted all reuse intervals using SGVE approximations, each bounded by two consecutive full-computation points. Starting from the same state at the beginning of each interval, we performed full computation at every step to obtain a reference endpoint state. We then measured the endpoint L2 errors of SeaCache + SGVE before and after RAIC correction relative to this reference. As shown below, RAIC provides relatively consistent correction at both budgets when reuse occurs at least twice, reducing state error in over 70% of these intervals.

≈50% computational budget
Reuse countPercentage in all intervalsSeaCache + SGVE errorSeaCache + SGVE + RAIC errorError ReductionPercentage of intervals with reduced endpoint error
144%0.003070.002944.4%62.70%
253%0.005030.0038723.10%91.10%
33%0.009760.00638.50%100.00%
≈30% computational budget
Reuse countPercentage in all intervalsSeaCache + SGVE errorSeaCache + SGVE + RAIC errorError ReductionPercentage of intervals with reduced endpoint error
122.5%0.004320.004251.70%61.00%
29.7%0.00450.0036518.70%87.50%
324.1%0.012630.011548.60%71.10%
45.6%0.023370.0187619.70%73.00%
526.3%0.022740.0131542.20%100.00%
611.8%0.027720.0148746.30%100.00%
AuthorsOpenReview ↗

Part2: W2, Q2, Q3.

W2 / Q2: Error patterns under video models and with both SGVE and RAIC enabled.

We added diagnostics analogous to Fig. 4 on HunyuanVideo, aggregated over the full VBench prompt set, and compared projection coefficients and directional alignment with RAIC enabled and disabled on FLUX.1-dev and HunyuanVideo visualizationsinAppendixL1˙,L2˙,oftheupdatedPDFvisualizations in Appendix L\.1, L\.2, of the updated PDF. This extends the diagnostics to video models and sampling trajectories with correction enabled.

As shown in Fig. 10(a)(b), HunyuanVideo exhibits error patterns similar to those in Fig. 4: the projection coefficient qj=⟨ej,eb⟩/∣eb∣22q_j=\langle e_j,e_b\rangle/|e_b|_2^2 increases with the interval coordinate βj\beta_j and closely follows qj=βjq_j=\beta_j, while the mean cosine similarity between interior and endpoint errors is approximately 0.75–0.90. These observations support ej≈βjebe_j\approx\beta_j e_b in terms of projection magnitude and directional alignment, respectively, providing a basis for estimating errors within video-model reuse intervals from endpoint residuals.

In Fig. 10(c), since βj<1\beta_j<1 inside the interval, directly copying the endpoint error overestimates the intermediate error component along that direction, especially near the interval start. RAIC therefore scales the endpoint residual by βj\beta_j and accumulates the correction using the actual step sizes. With the no-correction baseline normalized to 1, directly copying the endpoint error yields pointwise and interval error MSEs of 1.48 and 1.69, whereas SGVE reduces them to 0.28 and 0.29—approximately 72% and 71% below the no-correction baseline, respectively. This shows that modeling error growth within the interval is crucial for accurately estimating accumulated errors.

As shown in Fig. 11, Keeping SGVE enabled and toggling only RAIC yields broadly similar projection-coefficient and cosine-similarity curves for both models. HunyuanVideo’s projection curves largely overlap and remain close to the linear reference, while the local fluctuations in the middle and later portions of the FLUX.1-dev curves appear in both settings. Directional alignment increases as the interval progresses in both models. Thus, after RAIC changes the sampling trajectory, the endpoint residual retains useful directional information, and its corresponding error component can still be approximately characterized by βj\beta_j. This supports using ej≈βjebe_j\approx\beta_j e_b when SGVE and RAIC operate jointly.

Q3: Parameter search, computational budget definition, and low-budget HunyuanVideo results.

The target computational budget is defined by the total TFLOPs of the complete generation pipeline, including DiT blocks, cache-refresh probes, and SGVE/RAIC operations. We primarily determine baseline parameters through binary search. We first select 80 prompts at regular intervals from VBench or DrawBench, then select 64 of these in the same way to run the full-computation baseline and measure FLOPs. We use binary search to find parameters matching the target budget until the budget mismatch is below 0.5%, then validate the selected parameters on the remaining 16 prompts. In our experiments, the evaluated FLOPs were within 1% of the expected values.

For HunyuanVideo, the 5884 TFLOPs in Table 2 indeed correspond to 40.0% of the 14703 TFLOPs required for full computation. Describing the entire low-budget group as “approximately 30%” was imprecise; this operating point should be labeled as an approximately 40% computational budget. We added experiments at approximately 4400–4500 TFLOPs, as shown below. At 4454 TFLOPs, RACE achieves 26.98 dB PSNR, 0.108 LPIPS, and 0.864 SSIM, outperforming the other caching methods in the table on all three metrics. Compared with SeaCache, it improves PSNR by 2.19 dB, reduces LPIPS by approximately 28.5%, and increases SSIM by 0.037. These additional results cover lower computational costs than the original 5884 TFLOPs operating point.

MethodParameterTFLOPs ↓PSNR ↑LPIPS ↓SSIM ↑
Original—14703———
TeaCacheδ=0.24442220.170.2670.722
TaylorSeerS=4441320.750.2320.737
SenCacheε=0.60453024.300.1620.816
DiCacheδ=0.31449525.710.1440.842
SeaCacheδ=0.50445224.790.1510.827
RACEδ=0.50445426.980.1080.864

Thank you for your careful response. My comments have been addressed, and I have no further comments.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overview of RACE and Additional Experiments During the Rebuttal Period

We thank all reviewers for their thoughtful feedback and constructive suggestions.

Our paper introduces RACE, a training-free caching framework for accelerating diffusion models. We focus on two interconnected issues: velocity prediction errors during reuse steps and their accumulation in the sampling state. To address these issues, we propose State-Guided Velocity Extrapolation (SGVE), which improves prediction using the local state–velocity relationship, and Refresh-Anchored Integral Correction (RAIC), which uses endpoint residuals obtained at refresh steps to estimate and compensate for local velocity errors accumulated over the preceding interval. Both components reuse information already available during sampling, require no additional full-network evaluations, and introduce no tunable coefficients beyond those of the cache scheduler.

To further strengthen the paper, we conducted the following additional experiments during the rebuttal period:

  1. Direct validation of RAIC’s effect on state error correction. We added trajectory visualizations under a shared refresh schedule and comparisons of state errors before and after correctionvisualizationsinAppendixL3˙oftherevisedPDFvisualizations in Appendix L\.3 of the revised PDF. On FLUX.1-dev, at approximately 50% and 30% compute budgets, RAIC reduces endpoint state errors in over 70% of intervals containing at least two reuse steps. We also report cases where correction increases the error to clarify the method’s reliability and limitations.

  2. Additional video-model error diagnostics and VBench evaluation. We extended the endpoint error analysis to HunyuanVideo and compared error patterns with and without RAIC on both image and video modelsvisualizationsinAppendixL1˙,L2˙oftherevisedPDFvisualizations in Appendix L\.1, L\.2 of the revised PDF. We also added VBench Quality, Semantic, and Total scores for Wan2.1, HunyuanVideo, and MiniMax-H3. At both reported budget levels across all three video models, RACE achieves the highest VBench Total score among the compared caching methods, providing complementary evidence of overall video quality beyond reference fidelity.

  3. Validation across different solvers and cache schedulers. In the additional Heun experiments, RACE improves PSNR over SeaCache by 3.78 dB and 1.78 dB at the high and low budgets, respectively. On FLUX.1-dev, combining RACE with TeaCache, EasyCache, SenCache, or SeaCache yields improvements in every case.

  4. Additional practical overhead measurements and design ablations. We report runtime and peak GPU memory usage on FLUX.1-dev and Wan2.1. SGVE’s projection computations take less than 20 ms in total, accounting for less than 0.7% of total latency. We also added experiments on scheduling-threshold sensitivity, longer history windows, and nonlinear correction weights. Under the tested settings, these alternatives do not outperform the current two-point SGVE and linear RAIC designs.

In addition, we clarified the compute budgets and parameter selection procedure and added HunyuanVideo experiments at an approximately 30% compute budget.

These additions strengthen the experimental evidence for state error correction, video quality, compatibility, and practical overhead, while allowing us to describe RACE’s benefits and conditions of applicability more precisely.

Thank you once again for your time and valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.