acceptodds
Under review as a conference paper at ICLR 2027

ARCS: APPROXIMATION-RISK-AWARE COMPUTATION SCHEDULING FOR FEATURE FORECASTING IN DIFFUSION TRANSFORMERS

Abstract

Diffusion Transformers achieve strong generation quality but require expensive network evaluations across many denoising timesteps. Feature caching reduces this cost by reusing intermediate activations at subsequent timesteps. Recent methods mathematically approximate the temporal evolution of cached features to forecast their values, reducing errors associated with direct reuse. However, these forecasting methods often rely on predefined computation schedules, implicitly assuming that forecasts remain reliable for a predetermined duration. In practice, forecasting reliability varies substantially throughout denoising, making when to compute as important as how much to compute. We propose Approximation-Risk-Aware Computation Scheduling (ARCS), a training-free method that adaptively allocates full computations according to estimated forecasting risk. We instantiate ARCS for Taylor feature forecasting, where adaptive computation produces irregularly spaced computation anchors. To handle these irregular anchors, we formulate Taylor forecasting with Newton divided differences and derive non-uniform approximation-error proxies that provide inexpensive online risk signals for determining when full computation is needed. Across image and video generation models, ARCS improves generation quality over fixed-interval Taylor forecasting, with the largest gains observed under aggressive DiT acceleration. In particular, ARCS reduces FID and sFID by up to 18.2% and 19.8%, respectively, compared with fixed-interval Taylor forecasting at matched full-computation budgets.

Primary area:
generative models
Keywords:
diffusion transformers, diffusion acceleration, feature caching, feature forecasting, adaptive computation, efficient inference

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ew7kOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This work achieves SOTA gains in image and video generation with some conceptual novelty but suffers from insufficient core innovation, limited practical value ,and weak experimental evidence.

Critical strengths

None

Critical weaknesses
  • Weak innovation : The work only optimizes scheduling on top of existing Taylor feature prediction, with no new mechanism. The Taylor-error-based risk proxy is simple and non-generalizable, only working for Taylor prediction .

  • Low practical deployment value: ARCS adds extra inference overhead using cached anchors. It has higher latency than SpeCa/CG-Taylor under the same compute budget. Tedious threshold and order tuning are required

  • Insufficient experiments: Some baselines lack order-matched comparisons. Ablations only validate the proposed proxy without systematic module/parameter studies. Tests are limited to DiT, FLUX and Hunyuan; stability at extreme acceleration rates is unproven.

  • Incomplete paper: Limitations like absent error propagation modeling are only briefly mentioned with no in-depth analysis. Future work is vague and lacks concrete directions, resulting in poor paper closure.

Questions for authors

see above

4 comments
AuthorsOpenReview ↗

Response to Reviewer ew7k (W1-W2)

We thank the reviewer for their feedback and for the opportunity to clarify our contributions and address the concerns raised.

W1: Novelty and Generalizability

As discussed in Section 2, existing feature forecasting methods, including TaylorSeer, primarily focus on how accurately intermediate features can be predicted under limited computational budgets. In contrast, our work addresses a complementary but distinct question: When should these predictions be trusted, and when should full computation be performed?

The key contribution of ARCS is deriving a scheduling signal from the approximation-error structure of the underlying forecaster. Unlike existing adaptive scheduling methods that rely on feature-change estimates or separate validation signals, ARCS determines when to perform full computation using a risk proxy derived from approximation error.

Specifically, we formulate Taylor forecasting using Newton divided differences under irregularly spaced computation anchors and derive a risk proxy from cached feature history. We further introduce progressive forecasting order to accommodate limited anchor history and cross-order risk normalization to maintain comparability of risk scores across different prediction orders. These components enable adaptive Full/Forecast decisions without additional network evaluations for risk estimation.

Therefore, our methodological contribution is not the introduction of a new predictor, but rather establishing a connection between forecasting reliability and computation scheduling through approximation-error analysis.

We acknowledge that the current implementation of ARCS focuses on Taylor forecasting. However, the underlying principle of deriving scheduling signals from a predictor's approximation-error structure is not inherently restricted to Taylor-based methods.

Extending this approach to other predictors would require deriving an appropriate error proxy for each forecasting formulation, as explicitly discussed in Section 5 (Limitations and Future Work). We identify this extension as a promising direction for future research.

W2: Practical Utility and Inference Overhead

Latency and Generation Quality

We acknowledge that ARCS exhibits higher latency than SpeCa and CG-Taylor on DiT-XL/2. However, as shown in Table 1, the latency overhead relative to SpeCa is approximately 11.9% at 11 NFC, whereas it decreases to approximately 4.4% and 6.2% at the more aggressive 8-NFC and 6-NFC operating points, respectively.

Notably, at 6 NFC, although ARCS incurs approximately 6.2% higher latency than SpeCa, it reduces FID from 8.55 to 5.17, representing a 39.5% improvement, while increasing IS from 153.34 to 200.09.

CG-Taylor achieves the lowest latency of 0.0531 seconds under the same 6-NFC budget, but its FID deteriorates substantially to 73.45.

These results demonstrate that, despite incurring additional inference latency, ARCS maintains high generation quality even under aggressive acceleration.

Importantly, the latency trend observed on DiT-XL/2 should not be generalized to other models.

On FLUX.1-dev (Table 2), ARCS achieves lower latency and higher generation quality than SpeCa and CG-Taylor in both acceleration regimes. At approximately 7 NFC, ARCS achieves 8.6% and 38.1% lower latency, respectively, while attaining an ImageReward of 0.8450, compared with 0.6729 and -1.0849.

On HunyuanVideo, ARCS likewise achieves lower latency and higher generation quality than SpeCa in both regimes. At approximately 6 NFC, ARCS reduces latency by 14.0% (22.46 to 19.32 seconds) while improving VBench from 66.24 to 77.44.

Even when ARCS incurs additional latency, it substantially improves generation quality, while achieving simultaneous gains in inference speed and generation quality in several settings.

Threshold and Order Configuration

We acknowledge that ARCS requires threshold calibration to meet a target full-computation budget.

This involves selecting the base threshold (γ0\gamma_0) on a calibration set to match the target NFC, after which it remains fixed during evaluation. The threshold schedule parameter (rγr_\gamma) is fixed at 1.15 across all models.

For forecasting order, only the model-specific maximum order is specified. The effective order increases automatically with available anchor history, requiring no separate configuration for each order.

Thus, although ARCS requires initial calibration, threshold and order configurations follow defined procedures without extensive manual tuning.

AuthorsOpenReview ↗

Response to Reviewer ew7k (W3-W4)

W3: Experimental Comparisons, Ablations, and Evaluation Scope

Order-Matched Comparison

Tables 1 and 2 report results using the configurations recommended by the respective baseline methods, while additional order-matched comparisons are provided in Appendix B.4.

However, we acknowledge that these comparison settings were not sufficiently clarified in the main text and that an order-matched comparison with SpeCa on DiT-XL/2 was missing. The reviewer's comment prompted us to re-examine our comparison settings and identify this missing experiment, and we appreciate the opportunity to address this omission.

To address this concern, we conducted additional order-matched experiments with SpeCa on DiT-XL/2. The results are summarized below.

ConfigurationMax. orderBase thresholdNFCFID ↓sFID ↓IS ↑
SpeCa default40.211.02.695.21231.34
SpeCa order-matched20.211.02.705.27231.72
SpeCa default40.38.63.366.35218.41
SpeCa order-matched20.38.43.526.75216.92
SpeCa default41.05.98.558.46153.34
SpeCa order-matched21.05.99.1910.45152.14

We will incorporate these additional results into the revised manuscript and clarify the comparison settings.

Ablation Studies

Beyond the error-proxy comparison presented in Section 4.3, Appendix B provides additional analyses of the proposed components and parameter configurations.

Specifically, Appendix B.1 investigates sensitivity to the threshold schedule parameter, Appendix B.2 evaluates the contributions of progressive forecasting order and cross-order risk normalization, and Appendix B.3 compares module-level risk aggregation strategies, including Mean, Max, and RMS.

These analyses are presented in Tables 4–6 of Appendix B.

Therefore, our ablation studies are not limited to validating the proposed risk proxy; they also examine the major components of ARCS and its parameter choices.

Evaluation Scope and Aggressive Acceleration Regimes

We selected DiT-XL/2, FLUX.1-dev, and HunyuanVideo to evaluate class-conditional image generation, text-to-image generation, and text-to-video generation, respectively.

These models have also been used in prior diffusion acceleration studies, including TaylorSeer, SpeCa, and RFC. Building on these established evaluation settings, we evaluate ARCS across three distinct generation tasks.

Furthermore, Tables 1 and 2 include operating points where full computation is performed only 6.0 times (12.0%) on DiT-XL/2, 7.1 times (14.2%) on FLUX.1-dev, and 6.1 times (12.2%) on HunyuanVideo, out of 50 sampling steps.

These budgets are even lower than the most aggressive settings evaluated in RFC, namely 7 NFC on DiT, 8 NFC on FLUX, and 7 NFC on HunyuanVideo.

Therefore, our evaluation covers not only moderate acceleration settings but also aggressive computational budgets comparable to, or more constrained than, those considered in prior diffusion acceleration studies.

W4: Error Propagation and Future Research Directions

We agree that ARCS does not explicitly model error propagation throughout the denoising process.

However, error propagation is an inherent challenge shared by feature forecasting-based diffusion acceleration methods. Our work focuses on determining when full computation should be performed under a limited computational budget, despite this underlying challenge.

Regarding future research directions, Section 5 explicitly identifies the following three directions:

  1. Extending error-derived scheduling to other forecasting formulations.
  2. Incorporating error propagation into risk estimation.
  3. Developing online threshold adaptation for sample-specific computational budget control.

In particular, extending ARCS to other forecasters would require analyzing the approximation-error structure of each forecasting formulation and deriving a corresponding risk proxy that can be estimated from cached feature histories. While the current implementation focuses on Taylor forecasting, the underlying principle of using approximation-error-derived risk to guide Full/Forecast scheduling is not inherently restricted to Taylor-based methods. This would allow the scheduling framework to be extended to other forecasting formulations, provided that suitable risk proxies can be derived.


We hope these clarifications and additional experimental results help address the reviewer's concerns and better communicate the contributions and scope of our work.

We would be glad to provide further clarification and address any additional questions the reviewer may have.

Reviewer_ew7kOpenReview ↗

Feedback

Thanks for your rebuttal. Some of my questions have been addressed. But I am still concerned about Novelty. I tend to reject this paper. I decide raise my score from 1 to 2.

AuthorsOpenReview ↗

Thank You for Your Feedback and Reconsideration

Thank you very much for taking the time to carefully reconsider our rebuttal and for acknowledging that some of your concerns have been addressed.

We sincerely appreciate your thoughtful feedback and your willingness to reconsider the rating. We also hope that our responses have helped better communicate the motivation and contributions of our work.

Thank you again for your valuable time and consideration, especially given your busy schedule!

Reviewer_3LkvOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ARCS, an adaptive computation scheduler for feature forecasting in diffusion transformers. It constructs a risk proxy from higher-order divided differences of historical full-computation features to decide when to forecast and when to recompute, while accommodating irregular computation anchors. Experiments cover DiT-XL/2, FLUX.1-dev, and HunyuanVideo, with the clearest benefits appearing in image generation under aggressive acceleration. The approach has practical value, although the explanation of proxy effectiveness and some experimental accounting require further clarification.

Critical strengths

1.Clear gains under aggressive acceleration, supported by measured latency.

2.An inexpensive scheduling signal aligned with the objective.

3.Substantial ablations and baseline controls.

Critical weaknesses

1.The final proxy can miss derivative-estimation error, and its practical impact is unclear.

2.The correlation analysis does not isolate the shared effect of forecast distance.

3.Computational-budget accounting needs clarification. In FLUX Table 2, TaylorSeer reports 557.33 T FLOPs at 7.0 NFC, whereas ARCS reports 487.81 T at 7.1 NFC; near 10 NFC, their FLOPs are identical. Different profiling or initialization conventions could explain this, so it is not established as an error, but it complicates the interpretation of matched budgets. The statistical scope of NFC, FLOPs, and latency should be specified, with consistent-budget comparisons at the principal operating points. (Table 2 and Appendix C)

Questions for authors

see weakness

4 comments
AuthorsOpenReview ↗

Response to Reviewer 3Lkv (W1)

W1: Practical Impact of Derivative-Estimation Errors

We thank the reviewer for highlighting that our final risk proxy does not explicitly account for derivative-estimation errors and questioning their practical impact.

As discussed in Eq. (7), our theoretical forecasting-error bound includes both the Taylor truncation error and the derivative-estimation error, whereas the final ARCS scheduling proxy uses only the truncation-error term. We therefore acknowledge that actual forecasting errors may be underestimated.

To investigate the practical impact of this limitation, we extended the oracle forecasting-error analysis presented in Section 4.4 by further comparing proxy values against actual forecasting errors.

Under the 11-NFC setting, ARCS selected Forecast computation in all 32 batches at the final denoising timestep, (t=0). The mean proxy value was 223.48, corresponding to only 16.3% of the mean decision threshold of 1,368.5. In contrast, the mean actual forecasting error was 0.3888, approximately 3.0 times the overall mean error of 0.1294 across timesteps where Forecast computation was performed.

These findings demonstrate that the proxy can remain low even when the actual forecasting error is high, indicating that it may not adequately capture forecasting errors at certain timesteps.

To further assess how this limitation affects actual generation quality, we compared ARCS against an online oracle-guided scheduler that determines Full/Forecast computation based on observed forecasting errors. The decision threshold of the online oracle-guided scheduler was calibrated to achieve an average number of Full computations comparable to ARCS.

MethodNFC ↓FID ↓sFID ↓IS ↑
ARCS (Ours)11.012.704.88233.14
Online Oracle-guided11.162.735.12235.13

Under approximately matched Full-computation budgets, ARCS achieved generation quality comparable to that of the online oracle-guided scheduler. Specifically, ARCS achieved slightly better FID (2.70 vs. 2.73) and sFID (4.88 vs. 5.12), while the online oracle-guided scheduler achieved a higher IS (235.13 vs. 233.14).

These results support that, despite not explicitly accounting for derivative-estimation errors, the proposed proxy remains a useful practical signal for effectively allocating Full computations under the evaluated setting.

We will clarify this limitation and incorporate the additional analysis and generation-quality comparison into the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer 3Lkv (W2-W3)

W2: Correlation Analysis — The Effect of Forecast Distance

We agree that forecast distance may jointly influence the correlation between the proposed risk proxy and actual forecasting error. In Taylor-based forecasting, approximation errors can increase with forecast distance, which our proxy also reflects. Thus, their shared dependence on forecast distance may partially contribute to the observed correlation.

We interpret the correlation analysis as supporting empirical evidence for the proposed proxy. To further examine its effectiveness, we extended the oracle forecasting-error analysis from our response to Weakness 1 by comparing ARCS's Full-computation positions with timesteps exhibiting high actual forecasting errors.

We measured actual forecasting errors at each timestep along the same ARCS generation trajectory. For each batch, the 9 timesteps with the highest oracle forecasting errors were designated as oracle-selected Full positions. Together with 2 mandatory initial Full computations, this yields 11 Full computations, matching ARCS's budget. We compared these positions with those selected by ARCS.

Comparison MetricResult
Overall Full/Forecast decision agreement95.83% (1,472/1,536)
Exact agreement with oracle-selected Full positions88.89% (256/288)
Oracle-selected Full positions within 1 timestep of an ARCS Full position98.26%
Mean nearest distance0.142 timesteps

Beyond the observed proxy-error correlation, these results show substantial agreement between high-error timesteps and ARCS's Full-computation positions. Notably, 98.26% of oracle-selected Full positions were within 1 timestep of an ARCS Full position, with a mean nearest distance of 0.142 timesteps.

Because this analysis uses the same ARCS trajectory, it cannot fully isolate forecast distance or capture trajectory differences under alternative schedules. Nevertheless, it provides complementary evidence of alignment between high-error timesteps and ARCS's Full-computation allocation.

As reported in Weakness 1, ARCS also achieved comparable generation quality to the Online Oracle-guided scheduler at approximately matched Full-computation budgets, further supporting its practical effectiveness.

In the revised manuscript, we will clarify forecast distance's potential influence on correlation and its interpretive limitations in Section 4.4 and include the oracle-based scheduling analysis.


W3: Correction of FLUX Computational-Budget Accounting

We thank the reviewer for pointing out this inconsistency. Upon re-examination, we found that TaylorSeer's three initial warm-up Full computations had not been correctly accounted for, resulting in 8 NFC instead of the intended 7 NFC. Consequently, the original comparison inadvertently favored TaylorSeer by allowing one additional Full computation. We corrected the interval from (N=9) to (N=10) and reran the experiments. The corrected results show that ARCS achieves a larger performance advantage over TaylorSeer at the properly matched budget. We have updated Table 2 in the revised manuscript accordingly and appreciate the reviewer's careful observation, which helped us identify and resolve this issue.

Following the reviewer's suggestion, we will also clarify the statistical scope of NFC, FLOPs, and latency in Appendix C of the revised manuscript.

The corrected FLUX results from Table 2 are presented below.

MethodNFCLatency (s) ↓FLOPs (T) ↓ImageReward ↑CLIP ↑PSNR ↑SSIM ↑LPIPS ↓
Full computation50.020.663477.450.96620.2776———
TaylorSeer (N=6)10.04.62696.390.91480.277718.44180.66360.4147
SpeCa10.24.83789.090.91940.274917.62340.62380.4597
CG-Taylor10.07.49744.930.76010.276019.16900.63980.4452
ARCS (Ours)10.14.76702.650.94060.280220.33400.70830.3168
TaylorSeer (N=10)7.03.42487.810.62430.270116.52060.55310.5741
SpeCa7.23.83524.400.67290.271816.09760.48050.6441
CG-Taylor7.05.65539.99-1.08490.241814.31630.28970.8477
ARCS (Ours)7.13.50493.370.84500.279718.00100.61620.4424

We hope that our responses and additional experiments have adequately addressed the reviewer's concerns. We sincerely appreciate the reviewer's constructive feedback and welcome any further questions or comments.

good response and i will maintain my positive rating

AuthorsOpenReview ↗

Response to Reviewer 3Lkv

Thank you for your positive feedback and for taking the time to review our responses. We sincerely appreciate your constructive comments, which have helped us improve the clarity and quality of our work.

We remain open to any further questions or suggestions you may have.

Reviewer_xzNvOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates when to re-execute the full computation in diffusion Transformer feature prediction. ARCS uses Newton’s difference-quotient method to adapt to irregular computation anchors and employs a risk proxy—constructed from truncated Taylor terms—to schedule Full/Forecast decisions, eliminating the need for additional network forward passes for risk estimation. The method also incorporates progressive prediction orders, cross-order normalization, and module-level risk aggregation. Results on DiT-XL/2, FLUX.1-dev, and HunyuanVideo demonstrate that this method improves the trade-off between generation quality and latency under various computational budgets.

Critical strengths
  1. Problem-focused with a simple implementation. The reliability of feature predictions varies with the denoising process, so there is a reasonable justification for recalculating them based on risk. The agent relies solely on cached history and does not require additional network reliability checks.
  2. The benefits of some key design choices are clear. In Table 1, DiT reduces the FID from 6.32 to 5.17 at 6 NFC, with latency remaining largely comparable. In Table 2, FLUX’s results at approximately 7 NFC simultaneously improve multiple quality metrics and latency, indicating that the benefits are not solely due to changes in the number of computations.
  3. Control experiments support specific design choices. The appendix includes experiments comparing threshold schedules, the number of progressive stages, cross-stage normalization, aggregation methods, and baseline stage counts. Correlation analysis between agent and actual errors provides supplementary evidence, and the agent’s additional full computations do not affect actual scheduling.
Critical weaknesses
  1. There are specific cases where the final proxy fails to reflect the approximation error of the derivative. The quotient derivative estimate from Eq. (5) is substituted into the Taylor series expansion of Eq. (2); this is not equivalent to directly using Newton’s interpolation polynomials. For the scalar trajectory F(t)=(t−τ0)2F(t)=(t-\tau_0)^2, let the interval between the two nearest anchor points be h>0h>0, and assume there are sufficient anchor points to achieve an effective order of 2. Then, based on the estimates given in the paper, D1=hD_1=h and D2=2D_2=2, so that:

    \qquad |F(\tau_0+k)-\widehat F_2(\tau_0+k)|=|hk|.$$ The third-order difference quotient, however, is zero, so the final truncated proxy is zero. This example does not refute the complete error decomposition that includes derivative approximation terms, nor does it constitute a rejection of methods explicitly stated to be heuristic proxies; it illustrates that the correlation results alone are insufficient to explain under what conditions the proxy may fail to detect errors. This boundary is directly related to the paper’s core scheduling mechanism and warrants explicit discussion.
  2. The benefits need to be explained in terms of operating points and evaluation objectives. When DiT is set to 11 NFC, ARCS and TaylorSeer achieve FID values of 2.70 and 2.71, respectively, but with latencies of 0.1216 seconds and 0.1095 seconds. When HunyuanVideo is set to approximately 6 NFC, ARCS shows an improvement in PSNR, while VBench, SSIM, and LPIPS do not show corresponding improvements. These results do not negate the value of the method, but they suggest that similar NFC values are insufficient to generalize actual benefits. The paper already includes quality-latency curves; it is recommended to use these to clearly identify the range where the method excels and to distinguish PSNR improvements from overall generation quality.

  3. Existing calibration protocols provide limited evidence for unseen prompts. Appendix C explains that thresholds are based on NFC calibration rather than generation quality; which alleviates concerns about selecting operating points based on test quality. However, FLUX’s final set of 200 prompts includes those used for calibration and budget verification, with only the random seed altered. Therefore, the overall results cannot, on their own, demonstrate the stability of a fixed threshold for unseen prompts. A separate aggregation of prompts not included in calibration or budget verification in the existing evaluation would provide a direct, albeit limited, examination.

Questions for authors
  1. For the quadratic trajectory examples mentioned above, are there any protection mechanisms not described in the paper? Can examples be reported from existing oracle error diagnostics where low proxy values correspond to high errors, to illustrate the actual impact of this blind spot?
  2. Is it possible to separately summarize the prompt results in FLUX that were not used for calibration or budget verification, and report the mean and dispersion of the actual NFC values?
  3. Based on the existing delay curves and tables, what are the computational budget and quality objectives for which the authors believe this method offers the greatest advantages? Which settings are more suitable for continuing to use fixed-interval predictions?
3 comments
AuthorsOpenReview ↗

Response to Reviewer xzNv (W1, Q1)

W1: Proxy Blind Spots and Their Practical Impact (Question 1)

Safeguards and Empirical Examples of Blind Spots

We sincerely thank the reviewer for carefully examining our method and constructing a concrete quadratic counterexample that clearly highlights a potential limitation of our proposed risk proxy.

The reviewer's counterexample demonstrates that forecasting errors can arise from inaccuracies in derivative estimation even when the Taylor truncation error is zero. As discussed in Eq. (7), our forecasting error bound accounts for both the Taylor truncation error and the derivative-estimation error. However, the final ARCS scheduler uses a proxy based only on the truncation-error term. We therefore acknowledge that our proxy does not explicitly account for derivative-estimation errors and may exhibit blind spots where actual forecasting errors are not adequately captured.

Regarding Question 1, our current implementation does not include a separate mechanism specifically designed to detect such proxy blind spots.

To investigate whether these limitations arise in practice, we extended the oracle forecasting-error diagnostics presented in Section 4.4.

At the final denoising timestep (t=0t=0), ARCS selected Forecast computation in all 32 batches. The mean proxy value was 223.48, corresponding to only 16.3% of the mean decision threshold (1,368.5). In contrast, the mean oracle relative forecasting error was 0.3888, approximately 3.0 times the overall mean error of 0.1294 across Forecast observations and the highest mean error among Forecast timesteps.

Furthermore, 27 out of 1,248 Forecast observations (2.16%) had proxy values below 25% of the decision threshold while their actual oracle errors fell within the top 10% of the Forecast-error distribution. All 27 cases occurred at t=0t=0.

These findings empirically confirm that the proposed proxy can underestimate actual forecasting errors in certain cases, consistent with the blind spots highlighted by the reviewer.

Oracle-Based Scheduling Comparison Under a Matched Full-Computation Budget

To further assess how effectively the proposed proxy allocates Full computations despite these blind spots, we conducted an additional oracle-based scheduling analysis along the same ARCS trajectory.

Specifically, we measured actual forecasting errors at each timestep and retrospectively selected the nine timesteps with the highest oracle errors in each batch as oracle Full positions. Together with the two mandatory initial Full computations, this corresponds to the same budget of 11 Full computations as ARCS.

The comparison yielded the following results:

  • Overall Full/Forecast decision agreement: 95.83% (1,472/1,536)
  • Exact match rate for oracle-selected Full positions: 88.89% (256/288), excluding the two mandatory initial Full computations
  • Oracle-selected Full positions within one timestep of an ARCS Full position: 98.26%, with a mean nearest distance of 0.142 timesteps

These results indicate that, despite certain proxy blind spots, ARCS exhibits substantial agreement with Full-computation positions associated with high actual forecasting errors along the same trajectory.

Comparison of Actual Generation Quality

Finally, to examine how these scheduling differences affect actual generation quality, we conducted an additional comparison between ARCS and an online oracle-guided scheduler.

The online oracle-guided scheduler computes both Full and Forecast features along its own evolving generation trajectory and determines whether to perform Full computation based on the observed relative forecasting error. Its decision threshold was calibrated to achieve an average number of selected Full computations comparable to ARCS.

The generation-quality comparison under approximately matched Full-computation budgets is summarized below.

MethodNFCFID ↓sFID ↓IS ↑
ARCS (Ours)11.012.704.88233.14
Online Oracle-guided11.162.735.12235.13

Under approximately matched Full-computation budgets, ARCS achieved slightly better FID (2.70 vs. 2.73) and sFID (4.88 vs. 5.12), while the online oracle-guided scheduler achieved a higher IS (235.13 vs. 233.14). These results demonstrate that ARCS achieves competitive generation quality relative to an error-informed scheduling reference, despite the identified proxy blind spots.

Although our proxy does not fully capture all sources of forecasting error, these additional analyses suggest that it remains a useful practical signal for effectively allocating Full computations. We will clarify this limitation and incorporate the additional findings into the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer xzNv (W2, Q3-W3, Q2)

W2: Operating Points and Evaluation Objectives (Question 3)

We agree with the reviewer that matched NFC alone is insufficient to establish practical benefits, and that the preferred scheduling strategy depends on both the acceleration regime and the evaluation objectives.

Overall, our results suggest that ARCS offers its greatest advantages under aggressive acceleration, particularly for DiT-XL/2 and FLUX.1-dev, where its quality benefits over fixed-interval TaylorSeer become more pronounced as the Full-computation budget decreases.

DiT-XL/2

For DiT-XL/2, at 11 NFC, ARCS and TaylorSeer achieve nearly identical FID, while ARCS incurs higher latency. However, under the more aggressive 6-NFC setting, ARCS reduces FID from 6.32 to 5.17 while maintaining nearly identical latency (0.0794 s vs. 0.0791 s). This demonstrates a more favorable quality–latency trade-off under stronger acceleration.

FLUX.1-dev

For FLUX.1-dev, we additionally evaluated a less aggressive 14-NFC setting to better characterize the relationship between acceleration and generation quality. The following table summarizes the results across all three operating points.

NFCMethodLatency (s) ↓ImageReward ↑CLIP ↑PSNR ↑SSIM ↑LPIPS ↓
14.0TaylorSeer (N=4)6.130.93960.279320.20730.73430.3037
14.0ARCS (Ours)6.290.95280.278623.44900.80280.2087
10.0TaylorSeer (N=6)4.620.91480.277718.44180.66360.4147
10.1ARCS (Ours)4.760.94060.280220.33400.70830.3168
7.0TaylorSeer (N=10)3.420.62430.270116.52060.55310.5741
7.1ARCS (Ours)3.500.84500.279718.00100.61620.4424

Across all three FLUX operating points, ARCS improves ImageReward, PSNR, SSIM, and LPIPS, with comparable or better CLIP scores. Under stronger acceleration, ImageReward and CLIP gains become more pronounced, alongside larger LPIPS reductions, while PSNR and SSIM improvements remain positive but non-monotonic.

Latency overhead also decreases from 0.16 s at 14 NFC to 0.08 s at approximately 7 NFC.

HunyuanVideo

For HunyuanVideo, the results are more mixed. Although ARCS improves PSNR at both operating points, the more aggressive setting does not yield corresponding improvements in VBench, SSIM, or LPIPS.

Recommended Operating Regimes

ARCS is most beneficial when generation quality under constrained Full-computation budgets is prioritized, particularly for DiT and FLUX under aggressive acceleration. TaylorSeer remains reasonable when latency is prioritized or quality gains are limited or mixed, as in 11-NFC DiT and HunyuanVideo.

We will clarify these quality–latency trade-offs and operating regimes in the revised manuscript.


W3: Generalization to Unseen Prompts (Question 2)

We thank the reviewer for the constructive feedback and specific guidance on evaluating ARCS's stability on unseen prompts.

Following this suggestion, we separately analyzed 136 FLUX prompts (indices 64–199) that were excluded from both threshold calibration and budget verification.

Actual NFC Distribution on Unseen Prompts (Mean ± Std).

ThresholdTarget NFCAll 200Unseen 136Unseen Min–Max
γ0=5250\gamma_0=52501010.090 ± 0.75010.066 ± 0.7409–12
γ0=10500\gamma_0=1050077.080 ± 0.6957.000 ± 0.6756–9

On unseen prompts, mean NFC closely matches the target budgets, with standard deviations comparable to the full 200-prompt set, indicating stable computation budgets without recalibration.

Generation Quality on Unseen Prompts (indices 64–199, 136 prompts).

MethodNFCCLIP ↑ImageReward ↑PSNR ↑SSIM ↑LPIPS ↓
TaylorSeer (N=6,m=1N=6, m=1)10.00.27740.870618.21030.64150.4201
ARCS10.10.28090.917819.81030.68280.3335
TaylorSeer (N=10,m=1N=10, m=1)7.00.27210.648916.38470.54440.5622
ARCS7.00.27980.828917.54530.58330.4660

ARCS outperforms TaylorSeer across all five quality metrics at both budgets, maintaining its advantage on unseen prompts.

We will include these results in the revised manuscript.


We hope that our responses and additional experiments have adequately addressed the reviewer's questions and concerns. We sincerely appreciate the reviewer's constructive feedback, which has been valuable in helping us identify areas for improvement and further strengthen our work.

We would be glad to provide further clarification and welcome any additional questions or comments the reviewer may have.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer xzNv,

We would like to kindly follow up on our rebuttal and ask if you have any remaining questions or concerns regarding our responses.

We would be more than happy to provide any additional clarification if needed!

We sincerely appreciate your time and thoughtful consideration, especially given your busy schedule. Thank you again for your valuable feedback!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Manuscript Revision Update

Dear Area Chair and Reviewers,

We have uploaded a revised version of our manuscript incorporating the feedback received during the rebuttal period.

The main revisions include:

  1. Additional oracle-based analyses and evaluation.
  2. Corrected FLUX computational-budget accounting and updated results.
  3. Additional order-matched baseline comparisons.
  4. Clarification of quality–latency trade-offs and proxy limitations.
  5. Improved experimental details, reproducibility, and discussion of limitations.

We sincerely thank the Area Chair and reviewers for their valuable time, thoughtful feedback, and constructive suggestions. Your comments have been extremely helpful in improving the clarity and quality of our work.

We greatly appreciate your careful consideration.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.