acceptodds
Under review as a conference paper at ICLR 2027

BoundaryDelta: Spline-Integrated Average-Velocity Prediction for Efficient Diffusion Sampling

Abstract

Cache-based acceleration for diffusion and flow-matching models typically reuses or extrapolates an instantaneous prediction, even though each solver step depends on velocity accumulated over an interval. We introduce BoundaryDelta, a training-free framework that instead estimates the interval-average velocity. At each cached step, BoundaryDelta constructs a causal, boundary-constrained cubic Hermite velocity surrogate from completed model evaluations, then integrates it analytically to advance without an additional transformer call. Because approximation quality depends on where genuine evaluations occur, we couple this controller with budget-specific anchor schedules found offline. Across Qwen-Image and Stable Diffusion 3.5 Medium, BoundaryDelta delivers the strongest overall FULL-reference fidelity among equal-NFE baselines across aggressive and moderate call budgets. In a separate image-conditioned 3D study, BoundaryDelta retains FULL-50-level held-out ground-truth geometry for Hunyuan3D-Shape-v2.1 Small using 17 of 50 transformer evaluations. On Qwen-Image, BoundaryDelta reduces transformer evaluations by 50–80% at the primary 10-, 17-, and 25-call settings and achieves up to paired steady-state speedup.

Primary area:
generative models
Keywords:
diffusion acceleration, flow matching, training-free caching, average velocity, interpolation, schedule optimization

Reviews

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

This paper describes a training-free faster diffusion and flow-matching models sampling method, which uses Hermite spline to avoid velocity evaluation at every timestep, but instead predicts the estimated intermediate velocities. By conducting experiments on various base models, the novel sampling method achieves the state-of-the-art efficiency, while preserving the generated image quality under different metrics.

Critical strengths

Here are some major strengths of the paper:

  1. The theoretical foundation of the method is clear and thorough, with explicit motivation and rigorous Hermite spline derivations.
  2. Experiments inside of the paper are comprehensive, across different models, different baselines, different tasks, and different metrics.
Critical weaknesses

Here are some weaknesses:

  1. The Bn experiments (for example B10), although their configurations are able to be inferred from the paper context, should be defined more explicitly.
  2. The anchor optimization part is a bit suspicious; see the questions below.
Questions for authors

The paper looks sound in general, but I have one major concern, which can be split into the following detailed questions

  1. The exact-budget anchor optimization is described inside of Section 5; however they seem to be costly, since whenever the NFE budget and the size of the generated image changes, they need to be re-optimized by minimizing the loss (13). Can you give an analysis of the budget/wallclock time needed for the exact-budget anchor optimization?
  2. Should the anchor optimization NFE budget be also included in the comparison, since they're highly configuration-dependent? For example inside of Table 1, would it be a more fair comparison if the B10 and B17 are counted as using 10/17 NFEs+optimization NFEs, instead of purely 10/17 NFEs?
  3. Does the actual optimization development prompt–seed pairs used in the anchor optimization procedure matter? More precisely, if we happen to choose some different/bad samples used during the optimization, will the final anchor positions change drastically and therefore affect the final results?

I will raise my score to 3 or 4 if the anchor optimization related questions are addressed. Edit: I raised my score to 3.

5 comments

Response to Reviewer sUux — Part 1/3

Thank you for recognizing the rigorous derivation and the breadth of our evaluation. Your questions concern the practical cost and reliability of anchor scheduling. We address them with measured optimization costs, a frozen-schedule resolution-transfer test, and an additional search on different development prompts.

Budget definition. B10, B17 and B25 mean exactly 10, 17 and 25 guided transformer evaluations per generated sample, including the startup anchors, on the unchanged 50-transition solver grid. The remaining transitions use cached updates; FULL-50 evaluates the model at every transition.

Q1. What does anchor optimization cost, and must a configuration change require another search?

Measured offline cost. Our fresh Qwen-Image CFG4 searches use ten development prompts and two optimizer seeds. Q200 means that each search run evaluates 200 distinct candidate anchor schedules: eight starting schedules followed by 192 additional candidates. Each candidate is scored by its average terminal-latent MSE over the ten development prompts. Each search runs on four AMD MI210 GPUs and took approximately 8.3 hours at B17 and 12.0 hours at B25. Table S1 reports the combined cost of both optimizer seeds, together with the subsequent candidate-selection and validation stages.

Table S1. Measured BoundaryDelta offline preparation, both optimizer seeds combined. Calibration includes search, candidate collection, selection and validation.

BudgetSchedule search (GPU-hours)Full measured calibration (GPU-hours)
B1039.8653.94
B1766.0282.62
B2596.05104.02

Each reported call budget uses its own separately prepared schedule, with the one-time per-budget cost listed above. A new resolution did not require another search in our test: the B17 masks transferred unchanged from 1664×928 to 1328×1328 (Table S2).

Schedule reuse across resolutions. We transferred both methods' original B17 masks unchanged from 1664×928 to 1328×1328, keeping guidance at CFG4. Table S2 uses the same 50 DrawBench prompts and generation seed 42, with a resolution-matched FULL-50 reference.

Table S2. Frozen-mask resolution transfer, B17. Same 50 prompts, seed 42, resolution-matched FULL-50 and fixed CFG4; no re-search. Lower MSE/L1/LPIPS and higher SSIM/PSNR are better. Bold/underline mark the better/second mean within each setting.

SettingMethodMSE ↓L1 ↓LPIPS ↓SSIM ↑PSNR ↑
1664×928, CFG4BoundaryDelta0.00680.04430.06590.941228.73
Released MeanCache0.0091‾\underline{0.0091}0.0566‾\underline{0.0566}0.0876‾\underline{0.0876}0.9250‾\underline{0.9250}26.84‾\underline{26.84}
1328×1328, CFG4BoundaryDelta0.00650.03880.05910.938129.62
Released MeanCache0.0092‾\underline{0.0092}0.0543‾\underline{0.0543}0.0854‾\underline{0.0854}0.9155‾\underline{0.9155}27.33‾\underline{27.33}

At the new resolution, BoundaryDelta reduces MSE/L1/LPIPS by 29.93%/28.54%/30.76% versus released MeanCache, with higher SSIM and PSNR. The searched masks therefore transfer between these resolution settings while preserving BoundaryDelta's fidelity advantage over released MeanCache.

Response to Reviewer sUux — Part 2/3

Q2. Should optimization evaluations be included in the inference comparison?

We propose reporting offline preparation and per-image inference as separate, complementary costs. Table S1 makes the optimization cost explicit, while B17 describes the 17 guided evaluations performed for each generated sample using the frozen schedule. If preparing a schedule costs CofflineC_{\mathrm{offline}} guided evaluations and it is reused for MM generations, the total is Coffline+17MC_{\mathrm{offline}}+17M, and the average per generation is 17+Coffline/M17+C_{\mathrm{offline}}/M. This accounts for preparation once and makes its contribution explicit for any deployment volume.

Once obtained, the same mask is reused across prompts and can also transfer across the tested resolution settings, as Table S2 demonstrates. The relevant parallel with a pretrained or distilled checkpoint is that both produce reusable artifacts: a user can download the prepared checkpoint or anchor schedule and apply it directly to subsequent generations. BoundaryDelta keeps the pretrained weights fixed and optimizes only the evaluation locations. We therefore report schedule preparation alongside the inference comparison, while retaining the exact per-sample B10/B17/B25 counts. The preparation cost is amortized as the number of generated samples grows.

Response to Reviewer sUux — Part 3/3

Q3. Does changing the development prompt–seed panel alter the schedule and final performance?

Using a different ten-prompt search panel, the complete development pipeline selected a B17 schedule differing from our deployed schedule at only one evaluation location. Their archived terminal losses on the same original search panel differ by just 0.48%, with the new schedule slightly lower. On all 200 DrawBench prompts, it also improves all five fidelity means over the deployed schedule, including 2.52% lower MSE and 2.69% lower LPIPS.

A different development split. We re-split the development pool so that the ten search prompts were disjoint from the original ones. At B17 we ran the complete pipeline: Q200 search with optimizer seeds 0/1, selection on 15 prompts × two generation seeds, and confirmation on 25 prompts × five seeds, all separate from DrawBench. Both searches used CFG4 and generation seed 42.

We then froze the alternate-split selected B17 mask and evaluated it on all 200 DrawBench prompts at generation seed 42.

Table S3. Held-out test of the alternate-split selected mask, B17. All 200 DrawBench prompts, seed 42. Lower MSE/L1/LPIPS and higher SSIM/PSNR are better; bold/underline mark the best/second-best displayed means.

Method / scheduleMSE ↓L1 ↓LPIPS ↓SSIM ↑PSNR ↑
BoundaryDelta, alternate-split selected mask0.005980.04090.05830.945629.40
BoundaryDelta, original deployed mask0.00613‾\underline{0.00613}0.0419‾\underline{0.0419}0.0599‾\underline{0.0599}0.9445‾\underline{0.9445}29.24‾\underline{29.24}
Released MeanCache0.008350.05420.08090.928727.36

The alternate-split mask preserves BoundaryDelta's advantage (MSE/L1/LPIPS −28.4%/−24.5%/−28.0% versus released MeanCache) and slightly improves on the deployed mask (MSE −2.5%, LPIPS −2.7%). The two BoundaryDelta masks have nearly identical automated quality means: HPSv2.1 0.29967/0.29966, VQAScore 0.8574/0.8575, CLIP-G 33.47/33.46 and ImageReward 1.211/1.213, respectively. Selection and confirmation average over two and five generation seeds per prompt, respectively, providing development evidence across noise draws.

Separately, the original frozen schedules retain all five fidelity point-estimate improvements over released MeanCache on the same 50-prompt panel at generation seeds 42, 43 and 44, across B10/B17/B25. This complementary deployment-noise test is also reported in our response to Reviewer 8DyQ (Table R4); we repeat the three-seed averages below to keep this answer self-contained.

Table S4. Three-generation-seed averages, original frozen systems. Each seed uses the same 50 prompts. Metric directions and within-budget highlighting follow Table S2; CLIP-G is 100×cosine.

BudgetMethodMSE ↓L1 ↓LPIPS ↓SSIM ↑PSNR ↑CLIP-G ↑ImageReward ↑HPSv2.1 ↑
B10BoundaryDelta0.02760.10740.19210.834721.7531.801.0920.2885
Released MeanCache0.0493‾\underline{0.0493}0.1396‾\underline{0.1396}0.2606‾\underline{0.2606}0.7859‾\underline{0.7859}18.81‾\underline{18.81}31.71‾\underline{31.71}1.084‾\underline{1.084}0.2882‾\underline{0.2882}
B17BoundaryDelta0.00620.04250.06220.944828.9932.001.1010.2927
Released MeanCache0.0087‾\underline{0.0087}0.0547‾\underline{0.0547}0.0835‾\underline{0.0835}0.9279‾\underline{0.9279}26.92‾\underline{26.92}31.75‾\underline{31.75}1.087‾\underline{1.087}0.2917‾\underline{0.2917}
B25BoundaryDelta0.00240.02380.02820.974133.9031.841.0950.2931
Released MeanCache0.0042‾\underline{0.0042}0.0410‾\underline{0.0410}0.0551‾\underline{0.0551}0.9558‾\underline{0.9558}30.40‾\underline{30.40}31.69‾\underline{31.69}1.092‾\underline{1.092}0.2928‾\underline{0.2928}

In short, schedule search is a one-time, measured cost per call budget. Once found, a schedule is reused across prompts and transferred to the tested new resolution. The alternate development pipeline selected a closely related B17 schedule with nearly identical archived search loss and improved held-out fidelity.

Reviewer_ASNhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work proposes a training-free method for accelerating diffusion sampling by estimating the interval-average velocity and offline anchor scheduling. The proposed method was evaluated with Qwen-Image and Stable Diffusion 3.5 Medium on a number of generation applications.

Critical strengths

The proposed method is training-free, yielding reasonably good results over MeanCache. The proposed formula is elegant and looks interesting.

Critical weaknesses

While the proposed method seems to yield promising results over MeanCache and some other methods, the evaluation and the comparison are not comprehensive, so it is hard to justify that the proposed method is indeed SOTA. I am not sure if using MSE in the loss (Eq. (13)) and in the evaluation is fair enough and it is ambiguous how efficient this offline anchor optimization is. What is the justification of employing cubic Hermite interpolant? It is unclear if this proposed method is truly optimal - more explanation is required to provide insights to the readers.

Questions for authors

Please see the above and address the concerns: I am particularly concerned in insufficient evaluation, unclear motivation for cubic Hermite interpolant and unclear explanation for offline anchor scheduling. Moreover, it is also important to demonstrate the usefulness of this work not only in terms of training-free temporal acceleration, but also in terms of other acceleration methods. Lastly, what is the meaning of this work over practical generative models distilled from the original models? Will this method outperform them? Will this method work on them? Please see below for some recent related works that were not discussed and compared.

  1. Lihui Gu et al., ScalingCache: Extreme Acceleration of DiTs through Difference Scaling and Dynamic Interval Caching, ICLR 2026
  2. Divya Jyoti Bajpai et al., FlowCast: Trajectory Forecasting for Scalable Zero-Cost Speculative Flow Matching, ICLR 2026
  3. Wenliang Zhao et al., UniPC: A Unified Predictor-Corrector Framework for Fast Sampling of Diffusion Models, NeurIPS 2023
  4. Qinsheng Zhang et al., Fast Sampling of Diffusion Models with Exponential Integrator, ICLR 2023
  5. Shikang Zheng et al., Forecast then Calibrate: Feature Caching as ODE for Efficient Diffusion Transformers, AAAI 2026
  6. Junwen Tan et al., VDE: Training-Free Accelerating Rectified Flow Model via Velocity Decomposition and Estimation, CVPR 2026
  7. Zhengyang Geng et al., Mean Flows for One-step Generative Modeling, NeurIPS 2025
  8. Zhengyang Geng et al., Improved Mean Flows: On the Challenges of Fastforward Generative Models, CVPR 2026
  9. Wongi Jeong et al., Training-free Mixed-Resolution Latent Upsampling for Spatially Accelerated Diffusion Transformers, CVPR 2026
9 comments

Response to Reviewer ASNh — Part 1/5

Thank you for recognizing the elegance of the formula and the improvements over MeanCache. We clarify the choice of cubic Hermite interpolation, the search objective, and the cost of anchor optimization below. We also provide a controlled comparison with simpler polynomial predictors and additional VDE, UniPC and DEIS comparisons.

Our claim is scoped to the evaluated methods and operating points. At B17/B25 on Qwen-Image, BoundaryDelta gives the strongest FULL-reference fidelity among the released caching methods, official VDE presets and shorter-grid samplers compared below at comparable call counts. On SD3.5 Medium, it leads seven of eight metrics at both tested budgets.

Q1. Why cubic Hermite, and in what sense is the construction justified?

The cubic follows from the information we choose to preserve across a cached interval. Two causally estimated endpoint values and two slope estimates give four scalar constraints per velocity component. A cubic is the lowest polynomial degree that can satisfy four generic constraints, and its interval average has a closed form. Integrating this constructed surrogate analytically gives a lightweight estimate of the quantity that advances the solver: the velocity accumulated across the interval.

BoundaryDelta keeps instantaneous velocity observations separate from predicted interval averages. MeanCache's released Qwen implementation adds predicted averages to its velocity history and reuses them as pointwise velocities in later cached predictions (released implementation). BoundaryDelta uses genuine full evaluations for its instantaneous-velocity anchors; predicted averages advance the latent without replacing those observations. This preserves the distinction between the velocity at an anchor and its average over an interval.

On 50 saved FULL trajectories, we tested cached predictions whose latest full evaluation was one to four solver steps earlier. Every rule used the same three earlier anchors and current saved latent, with the current model output withheld. The target was the saved FULL step's displacement divided by its step length. Each trial reset to the saved state, excluding accumulated prediction error. Steps 6–49 gave 8,800 local tests per method.

Table A1. Local one-step error of BoundaryDelta and integrated extrapolation rules. Same 50 saved FULL trajectories, steps 6–49, latest-anchor gaps of one to four steps; 8,800 tests per method. Mean relative-L2 error is lower for a more accurate prediction; bold/underline mark the best/second-best displayed means.

Causal predictorMean relative-L2 error ↓
BoundaryDelta0.1628
Integrated linear extrapolation0.3455‾\underline{0.3455}
Integrated quadratic extrapolation1.3510

The integrated linear and quadratic extrapolation rules have 2.1× and 8.3× BoundaryDelta's local error. BoundaryDelta uses anchor-difference information while avoiding the error growth observed for these two extrapolation rules. This supports its construction relative to those alternatives; the full-rollout comparison with identical schedule search separately shows its fidelity contribution over MeanCache at B17/B25 (our response to Reviewer 8DyQ, Table R1).

Response to Reviewer ASNh — Part 2/5

Q2. Is using terminal MSE for search and evaluation fair?

We use terminal latent MSE to search efficiently for schedules that preserve the original sampler's output. Each candidate already produces a terminal latent, so computing this loss requires neither image decoding nor an additional perceptual model during search. This makes it practical to compare many candidate schedules within the offline search budget.

Our own historical development study supports this choice. Across 11, 9 and 10 distinct schedules at B10/B17/B25, respectively, latent MSE closely tracked LPIPS, with Spearman correlations of 0.982, 0.983 and 1.000. Table A2 also reports agreement with SSIM and PSNR: a value near 1 means that schedules preferred by MSE are also preferred by the decoded-image fidelity metric.

Table A2. Agreement between schedule rankings from latent MSE and image-fidelity metrics. Entries are Spearman correlations, with signs aligned so that positive values mean lower MSE agrees with lower LPIPS or higher SSIM/PSNR. Within each row, the three correlations use the same candidate schedules.

BudgetDistinct schedulesLPIPSSSIMPSNR
B10110.9820.9730.973
B1790.9830.9831.000
B25101.0001.0001.000

The close ranking agreement in these small development sets supports MSE as an inexpensive way to search for fidelity-preserving schedules. After freezing the schedules, we independently evaluate them on held-out prompts using MSE, L1, LPIPS, SSIM and PSNR, and assess prompt alignment and preference through a separate reference-free metric family. This combines an efficient preservation objective during search with broader evaluation of the generated outputs.

Response to Reviewer ASNh — Part 3/5

Q3. Why this schedule optimizer, how much does it cost, and is it optimal?

BoundaryDelta couples the predictor with the full evaluations that support it. The paper's controller–schedule ablations show why placement matters: the same number of calls can leave very different gaps between anchors. We therefore compared several ways to search for a good placement before choosing the self-adjusting swap method.

Why this optimizer. Our historical five-prompt screen at B10/B17/B25 compared six exact-budget methods: self-adjusting (1+Λ)(1+\Lambda)-swap, randomized hill climbing, conditional-Bernoulli cross-entropy search, cross-entropy with variable-neighborhood refinement, schedule-aware variable-neighborhood search (PAT-VNS), and gap-density search. Swap reached the lowest final search loss at each method's documented allowance and the best mean rank after a matched 56 candidates: 1.67, versus 2.33 for PAT-VNS. It exchanges anchors while preserving the call count and adjusts one mutation radius, returning an empirically effective finite-budget schedule without a global-optimality claim.

Measured cost and reuse. In our fresh Qwen-Image CFG4 replication, each optimizer seed evaluates 200 distinct schedules on ten development prompts, including eight starting schedules. Each BoundaryDelta lane uses four AMD MI210 GPUs and took approximately 8.3 hours at B17 and 12.0 hours at B25. The following allocations combine both optimizer seeds:

Table A4. Fresh offline preparation costs in GPU-hours. Full calibration includes downstream candidate selection and validation/recovery stages under each budget's protocol.

BudgetBD searchMC searchBD full calibrationMC full calibration
B1039.8640.0553.9454.05
B1766.0267.0482.6284.57
B2596.0596.43104.02104.54
Reviewer_8DyQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies caching-based acceleration for diffusion and flow-matching models. BoundaryDelta estimates the average velocity over a solver interval by analytically integrating a cubic Hermite surrogate constructed from completed model evaluations and the current latent state. The resulting update skips transformer calls while retaining the original 50-step grid. The paper additionally searches for budget-specific anchor schedules offline by minimizing terminal-latent MSE against FULL-50 on development data, then freezes them for evaluation. Experiments on Qwen-Image and Stable Diffusion 3.5 Medium report improved FULL-reference fidelity over equal-call baselines. A separate Hunyuan3D study evaluates held-out geometry, and Qwen-Image achieves up to 4.72× paired steady-state speedup.

Critical strengths
  1. The paper presents a lightweight, causal update that requires no model training or additional transformer evaluations at cached steps. Coupling the prediction rule with anchor placement is a reasonable way to address errors caused by long cache intervals.
  2. The evaluation uses paired initial noise, all 200 DrawBench prompts, and development data separated from final evaluation. The measured Qwen speedups and fidelity improvements provide useful evidence for acceleration under the tested configurations, beyond selected qualitative examples.
Critical weaknesses
  1. The predictor's independent contribution is unclear. MeanCache already considers average velocity and performs better at B10 under the same schedule (Table 11). It is difficult to judge how much of the improvement comes from the new predictor rather than schedule optimization.
  2. The evaluation mainly measures similarity to FULL-50, which does not directly establish better generation quality. Preference metrics do not consistently improve, and robustness across random seeds and changes to the model configuration remains insufficiently assessed.
Questions for authors
  1. How much does the new predictor improve over MeanCache when both receive the same schedule-search procedure and budget?
  2. Is there evidence of better user-perceived quality at comparable latency, and robustness across random seeds and model configurations?
5 comments

Response to Reviewer 8DyQ — Part 1/4

Thank you for recognizing the value of our lightweight causal update and paired evaluation. Under the same schedule-search procedure and allowance, BoundaryDelta improves all five fidelity means at B25 and four of five at B17. We complement this attribution result with quality scores at comparable speed and tests across generation seeds and resolutions.

For method-score columns, bold marks the best point estimate and underlining the second best within each budget or setting. Table values are rounded consistently by metric; percentage changes use unrounded values.

Q1. How much does the predictor improve over MeanCache under the same schedule-search procedure and budget?

BoundaryDelta combines a causal spline-integrated predictor with the anchor schedule supporting it. Its contribution is not average velocity alone. To compare the predictors under the same optimization procedure, we searched schedules separately for BoundaryDelta and MeanCache using the same swap optimizer, eight matched starting schedules, ten development prompts, and Q200 allowance per optimizer seed, at each call budget. Both optimizer seeds, 0 and 1, contribute candidates. Inference uses Qwen-Image, CFG4, 1664×928 resolution, BF16 and the unchanged 50-step grid. B10/B17/B25 mean exactly 10/17/25 whole guided transformer evaluations.

Candidates are selected on another 15 development prompts with two generation seeds. DrawBench is not used for search or selection. The final comparison below uses all 200 DrawBench prompts, generation seed 42, paired initial noise and matching FULL-50 references.

Table R1. B17/B25 final test after identical search allowances. All 200 DrawBench prompts, generation seed 42. Lower MSE/L1/LPIPS and higher SSIM/PSNR are better; CLIP-G (100×cosine), ImageReward and HPSv2.1 are reference-free scores, with higher values better.

BudgetMethodMSE ↓L1 ↓LPIPS ↓SSIM ↑PSNR ↑ (dB)CLIP-G ↑ImageReward ↑HPSv2.1 ↑
B17BoundaryDelta0.0061‾\underline{0.0061}0.04120.05920.945729.4233.431.2140.2998
MeanCache, same search0.00590.0443‾\underline{0.0443}0.0608‾\underline{0.0608}0.9448‾\underline{0.9448}29.23‾\underline{29.23}33.34‾\underline{33.34}1.209‾\underline{1.209}0.2986‾\underline{0.2986}
B25BoundaryDelta0.00210.02250.02500.975434.8833.43‾\underline{33.43}1.2190.3004
MeanCache, same search0.0024‾\underline{0.0024}0.0249‾\underline{0.0249}0.0285‾\underline{0.0285}0.9735‾\underline{0.9735}34.10‾\underline{34.10}33.441.212‾\underline{1.212}0.2994‾\underline{0.2994}

At B25, BoundaryDelta improves all five fidelity metrics under identical search: MSE −10.13%, L1 −9.49%, LPIPS −12.36%, with higher SSIM and PSNR, all with paired 95% intervals favoring BoundaryDelta. At B17 it improves four of five fidelity means (L1 −7.08%, LPIPS −2.66%, higher SSIM and +0.19 dB PSNR; the L1 interval favors BoundaryDelta), while MSE is about 2% higher, with an interval overlapping zero. HPSv2.1 also favors BoundaryDelta at both budgets in paired intervals.

At B10, the most aggressive budget, MeanCache's predictor is better under equal search, consistent with Table 11. Anchor placement remains important: BoundaryDelta's complete system reduces MSE by 46.8% versus released MeanCache. At B17/B25, the predictor adds fidelity under the same search procedure. Together with the original-system MSE reductions of 46.8%/26.6%/48.9% at B10/B17/B25, this supports the coupled design: anchor placement is particularly important when calls are scarce, and the integrated predictor adds fidelity at the denser tested budgets.

Response to Reviewer 8DyQ — Part 2/4

Q2. Quality at comparable latency, random seeds, and model configurations

Quality at comparable latency. Table R3 compares automated preference (ImageReward and HPSv2.1) and prompt alignment (CLIP-G and VQAScore) at closely matched speed. These complement the FULL-reference fidelity results; we have not conducted a human study.

Table R3. Automated preference and alignment at matched speed. All 200 DrawBench prompts, seed 42; higher is better. Speed-ups are the paper's paired steady-state values for the original systems; CLIP-G is 100×cosine. Bold/underline rank quality means among the three equal-call systems within each budget; FULL-50 is the unranked reference. FULL-50's budget dash means the full-reference setting.

BudgetMethodSpeed-up ↑CLIP-G ↑ImageReward ↑HPSv2.1 ↑VQAScore ↑
—FULL-50 (reference)1.00×33.351.21930.30120.8542
B10BoundaryDelta4.72×33.44‾\underline{33.44}1.20750.29620.8532‾\underline{0.8532}
Released MeanCache4.75×33.451.2049‾\underline{1.2049}0.2960‾\underline{0.2960}0.8607
Native 10-step Qwen4.74×32.950.76450.24680.8249
B17BoundaryDelta2.84×33.46‾\underline{33.46}1.2130‾\underline{1.2130}0.29970.8575
Released MeanCache2.86×33.291.22450.2990‾\underline{0.2990}0.8573‾\underline{0.8573}
Native 17-step Qwen2.85×33.721.11350.28740.8512
B25BoundaryDelta1.98×33.431.2187‾\underline{1.2187}0.30040.8557‾\underline{0.8557}
Released MeanCache1.96×33.231.22030.3000‾\underline{0.3000}0.8551
Native 25-step Qwen1.96×33.38‾\underline{33.38}1.20030.29910.8591

BoundaryDelta ranks first or second on every displayed quality score at every budget, while remaining close to FULL-50 at 2.0–4.7× paired speed-ups: within 1.7% on ImageReward and HPSv2.1, and within 0.4% on CLIP-G and VQAScore. Simply running fewer native steps at the same speed loses substantially at aggressive budgets: at B10, ImageReward falls from 1.21 to 0.76 and HPSv2.1 from 0.296 to 0.247. Against equally searched MeanCache (Table R1), HPSv2.1 also favors BoundaryDelta at B17 and B25.

Response to Reviewer 8DyQ — Part 3/4

Generation seeds. We evaluated the original frozen BoundaryDelta and MeanCache systems at seeds 42, 43 and 44, using the same fixed 50 DrawBench prompts at all three budgets, with seed-matched initial noise and FULL references. No masks are searched again. Table R4 reports equal-weight averages of the three seed means, retaining all eight original metrics.

Table R4. Three-generation-seed averages, original frozen systems. Each seed uses the same 50 prompts.

BudgetMethodMSE ↓L1 ↓LPIPS ↓SSIM ↑PSNR ↑CLIP-G ↑ImageReward ↑HPSv2.1 ↑
B10BoundaryDelta0.02760.10740.19210.834721.7531.801.0920.2885
Released MeanCache0.0493‾\underline{0.0493}0.1396‾\underline{0.1396}0.2606‾\underline{0.2606}0.7859‾\underline{0.7859}18.81‾\underline{18.81}31.71‾\underline{31.71}1.084‾\underline{1.084}0.2882‾\underline{0.2882}
B17BoundaryDelta0.00620.04250.06220.944828.9932.001.1010.2927
Released MeanCache0.0087‾\underline{0.0087}0.0547‾\underline{0.0547}0.0835‾\underline{0.0835}0.9279‾\underline{0.9279}26.92‾\underline{26.92}31.75‾\underline{31.75}1.087‾\underline{1.087}0.2917‾\underline{0.2917}
B25BoundaryDelta0.00240.02380.02820.974133.9031.841.0950.2931
Released MeanCache0.0042‾\underline{0.0042}0.0410‾\underline{0.0410}0.0551‾\underline{0.0551}0.9558‾\underline{0.9558}30.40‾\underline{30.40}31.69‾\underline{31.69}1.092‾\underline{1.092}0.2928‾\underline{0.2928}

BoundaryDelta improves all five fidelity point estimates in every budget–seed panel (seeds 42/43/44, the same 50 prompts, original frozen masks, no re-search). Averaged over seeds, MSE falls by 44.1%, 28.6% and 43.8% and LPIPS by 26.3%, 25.5% and 48.9% at B10/B17/B25; all three automated quality means are also higher.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of contribution, rebuttal evidence and revised manuscript

We thank the reviewers and Area Chair for their consideration. We have responded to each concern with explanations or additional experiments and incorporated the principal evidence into the revised manuscript. We summarize the contribution and the main rebuttal findings below.

BoundaryDelta couples a causal interval-average velocity predictor with a reusable, fixed-budget evaluation schedule. It keeps the pretrained weights unchanged and targets faster generation while preserving the chosen sampler's outputs. On Qwen-Image, the original B10/B17 systems use 80%/66% fewer evaluations and achieve 4.723×/2.841× paired steady-state speedup, with 27.6%/26.0% lower LPIPS than released MeanCache (Table 1).

1. Predictor contribution under the same search. Both predictors receive the same terminal-MSE objective, development data, starting schedules, optimizer and search allowance. Both optimizer seeds contribute to development-based selection; the frozen schedules are then evaluated on all 200 DrawBench prompts. At B25, BoundaryDelta improves all five fidelity means over searched MeanCache, including 10.13% lower MSE and 12.36% lower LPIPS. At B17, four of five fidelity means improve, including 7.08% lower L1. The searched MeanCache variant also improves all five fidelity means over released MeanCache at both budgets. These results separate the benefit of our schedule search from the additional predictor contribution (Table 3; Appendix D.2).

2. Motivation and fairness of terminal-MSE optimization. Section 5 explains why the final latent measures the accumulated endpoint deviation caused by caching. It is inexpensive to compare because each candidate already produces this tensor, avoiding image decoding and perceptual-model scoring during search. Across the historical B10/B17/B25 candidate sets, MSE–LPIPS ranking correlations are 0.982/0.983/1.000; Table 5 also reports SSIM, PSNR and CLIP correlations. Giving MeanCache the same MSE-based search directly addresses the tuning-fairness concern, while the additional metrics assess preservation beyond the optimized loss (Section 5; Appendix B.3).

3. Why the Hermite construction. The revised explanation connects the available anchor information to a boundary-constrained velocity curve whose interval average is obtained analytically. It clarifies which quantities are causal estimates and which coefficients follow exactly from integration. The trajectory and causal-prediction diagnostics provide numerical support for this modeling choice (Section 4; Appendix A.2–A.3).

4. Preparation cost, development sensitivity and reuse. Table 1 and Appendix B.4 report offline preparation separately from per-image inference. Repeating B17 preparation on a different development split gives similar held-out fidelity: the two freshly selected schedules differ by 1.23% in MSE and 1.53% in LPIPS on all 200 DrawBench prompts (Appendix D.3). Reusing the original B17 masks at a new resolution requires no additional search and retains BoundaryDelta's five-metric fidelity advantage over released MeanCache, including 30.76% lower LPIPS (Appendix D.4.3).

5. Broader evaluation and practical scope. The revision includes frozen-mask tests at generation seeds 42, 43 and 44 on the same 50 DrawBench prompts, where BoundaryDelta improves all five fidelity means over released MeanCache at B10/B17/B25. It also reports CLIP-G, ImageReward, HPSv2.1 and VQAScore on the complete original 200-prompt set, complementing FULL-reference fidelity with reference-free assessment. Additional VDE, UniPC and DEIS comparisons broaden the acceleration baselines (Appendices D.4–D.5). Section 2 clarifies the relationship to other acceleration approaches and learned few-step generators.

Together with the original two image backbones and image-conditioned 3D study, these results support BoundaryDelta as a practical, fixed-weight acceleration method with an analytically motivated predictor and reusable schedule preparation.

Reviewer sUux has confirmed that their concerns were addressed. We would appreciate the reviewers' and Area Chair's assessment of the revised manuscript and this additional evidence, particularly whether any specific substantive concerns remain unresolved.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.