acceptodds
Under review as a conference paper at ICLR 2027

SpecFlow: Fast and Lossless Speculative Sampling for Flow Matching

Abstract

Flow Matching (FM) has emerged as a leading framework for high-quality image and video generation, yet inference remains expensive because ODE integration requires sequential network evaluations. Existing acceleration methods either require additional training or rely on approximations, generally without exact distributional guarantees. We ask whether FM can instead admit the lossless acceleration of speculative decoding, where draft quality affects efficiency but not correctness. A direct extension is challenging because deterministic FM transitions are degenerate (Dirac) and lack the densities required for exact probabilistic verification. We address this by performing speculation on a marginal-preserving stochastic realization of the FM probability flow, whose Gaussian transitions enable exact verification and correction of draft proposals. We obtain the required score directly from the FM velocity, requiring no auxiliary model or additional network evaluations. Building on this, we introduce SpecFlow, a training-free speculative sampler that exactly reproduces its chosen discretized stochastic target sampler for any valid draft; making SpecFlow lossless. We further introduce a curvature-adaptive drafting strategy that jointly adapts prediction order and speculative horizon to improve efficiency. Across image generation, image editing, and video generation, SpecFlow achieves inference speedup while maintaining the target sampler's quality.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Flow Matching, Speculative Decoding

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_TFqqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper incorporates the speculative decoding in autoregressive models to the sampling of flow-matching based generative models. By re-introducing the diffusion score and schedule, the stochastic term is added back to the flow-based models, so that the speculative decoding can be applied. Experiments across different setups such as text-to-image, image editing, and video generation are conducted, with wall-clock computational efficiency provided to demonstrate the effectiveness of the SpecFlow.

Critical strengths

The idea of incorporating the autoregressive speculative decoding is novel in the flow-matching based model regime.

Critical weaknesses

See the questions down below.

Questions for authors

Here are my questions:

  1. Based on the setup of the paper, it is natural to apply the speculative decoding to noise-based or score-based diffusion models, instead of the flow-based models; is there any reasons the flow-based models were chosen to be the primary backbone models?
  2. Several classical baseline models are missing from the paper, for example the DPM-Solver++ [1] and the UniPC [2] solvers; it is beneficial to compare to at least one of them, in at least one of the setups (eg. text-to-image, image editing, etc.).
  3. I'm a bit surprised that the Euler–Maruyama solver attains the same image quality as the Euler's method across NFE=5, 10, 25, 50, as shown in Table 1. Although the marginal probability distributions are the same, the stochastic differential equation solver needs more timesteps (therefore more NFEs) than the ODE solver to sample stably due to the Wiener process term; one not identical but similar case example is DDIM [3] (Euler) v.s. DDPM [4] (Euler–Maruyama). Can you provide an intuitive explanation of why would the SpecFlow be able to avoid this issue? If I understand correctly, the speculative decoding should only accelerate the sampling but not improve the quality under the same budget.

[1] Lu, Cheng, et al. "DPM-Solver++: Fast solver for guided sampling of diffusion probabilistic models." Machine Intelligence Research 22.4 (2025): 730-751.
[2] Zhao, Wenliang, et al. "UniPC: A unified predictor-corrector framework for fast sampling of diffusion models." Advances in Neural Information Processing Systems 36 (2023): 49842-49869.
[3] Song, Jiaming, Chenlin Meng, and Stefano Ermon. "Denoising diffusion implicit models." arXiv preprint arXiv:2010.02502 (2020).
[4] Ho, Jonathan, Ajay Jain, and Pieter Abbeel. "Denoising diffusion probabilistic models." Advances in neural information processing systems 33 (2020): 6840-6851.

5 comments

We thank the reviewer for recognizing the novelty of applying speculative decoding to flow-matching models and for the constructive questions. We address each point below.

Q1. Why focus on flow-matching models rather than score-based diffusion models?

A1: We focus on Flow Matching (FM) because extending exact speculative decoding (SD) to its standard deterministic sampling procedure presents a distinct challenge that is not present in stochastic diffusion samplers. There are existing methods that have extended SD to diffusion and score based models but the major concern why it was not extendable to FM is due to deterministic paths from source distribution to target distribution.

As explained in Section 3.2 (Eq. 6), deterministic FM transitions are Dirac measures. Consequently, conventional likelihood-ratio verification cannot meaningfully accept a proposed next state unless it exactly matches the deterministic transition. In contrast, score-based diffusion samplers already have non-degenerate stochastic transitions, for which exact speculative sampling methods have been developed (Section 2).

Our contribution is to bridge this gap for pretrained FM models. We construct a marginal-preserving stochastic realization of the FM probability flow (Section 3.3, Eq. 7), derive the required Gaussian target transitions (Eq. 10), and recover the score estimate directly from the pretrained FM velocity (Eqs. 8–9), without an auxiliary score network or retraining.

Also, This is particularly relevant because modern image, editing, and video generators, including BAGEL, FLUX, and HunyuanVideo, use flow-based architectures. Our framework addresses these existing pretrained models rather than requiring a separately trained stochastic generator.

The underlying speculative verification principle can also be applied to score-based diffusion models with evaluable stochastic transitions. However, the deterministic nature of standard FM sampling is precisely the additional challenge addressed by SpecFlow.


Q2. Can you compare against classical higher-order solvers ...

A2: Higher-order solvers are relevant baselines, but they address a different computational objective. Methods such as DPM-Solver++ and UniPC improve numerical integration efficiency, whereas SpecFlow targets sequential sampling dependencies through speculative drafting and verification.

To address the requested numerical comparison, we additionally evaluated DPM-Solver++(2M) on BAGEL-7B-MoT at 1024×10241024\times1024, using all 553 GenEval prompts and identical initial latents. We compare fidelity relative to deterministic Full-50.

MethodSpeedup ↑GenEval ↑CLIP-IQA ↑LPIPS ↓PSNR ↑
Full-501.00×0.7800.846——
Full-202.46×0.7560.7970.27618.8
DPM-Solver++(2M)-202.46×0.7470.8110.27918.6
SpecFlow2.52×0.7800.8510.04831.7

At comparable speedup, SpecFlow achieves substantially lower LPIPS and higher PSNR than DPM-Solver++(2M). This indicates closer per-image agreement with Full-50.

We will include this baseline and clearly distinguish the two verification settings in the revision.


Q3. Why does Euler–Maruyama appear to maintain ...

A3: We would like to clarify that similar generation quality does not imply that deterministic Euler and Euler--Maruyama (EM) produce identical trajectories or distributions at finite NFEs. Our stochastic formulation is constructed to preserve the marginal distributions of the original FM probability flow in continuous time, under the ideal velocity and score assumptions (Appendix A.1.1). At finite sampling steps, discretization and learned-velocity errors can introduce differences between the two samplers.

Importantly, Table 1 compares Full-N sampling with SpecFlow-N, rather than directly comparing deterministic Euler and serial EM. The observed similarity in generation quality therefore should not be interpreted as numerical equivalence between the two solvers.

We already investigate this distinction in Appendix B.2 (Table 6), where deterministic and stochastic sampling exhibit close empirical agreement at finer discretizations, and Appendix B.5 (Table 9), which shows that larger diffusion scales can degrade stochastic-target quality.

Thus, the observed quality preservation is consistent with our stochastic construction, but is not assumed or guaranteed at arbitrary finite NFEs. SpecFlow's exactness guarantee applies specifically to its chosen discretized stochastic target, while its agreement with deterministic FM is evaluated separately.


Final Remarks

We sincerely thank the reviewer for their constructive feedback and thoughtful evaluation. Based on these clarifications and the additional evidence, we respectfully ask the reviewer to reconsider their assessment and increase their score.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer TFqq,

It is a gentle reminder to acknowledge our rebuttal and, if there are no further concerns, we request to increase the scores.

Best Regards,

Authors

Thanks the authors for the reply. However, for the DPM-Solver++ baseline comparisons in the Q2 response, I'm still not convinced.

  1. It seems like the DPM-Solver++ at 20 NFE is suspiciously worse than the vanilla Euler at 20 NFE for most of the metrics. Could the authors kindly provide the configs passed to the DPM-Solver++ experiment (I suppose the HuggingFace Diffuser version of DPM-Solver++ is used)? I'm a bit afraid that some mis-configurations caused unfair comparisons.
  2. Do you mind providing the LPIPS and PSNR values for the Full-50 experiment as well?
Reviewer_aMedOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper brings speculative decoding to flow matching samplers without any training. The obstacle is that a deterministic ODE step has no transition density, so there is nothing to verify a draft against. The paper gets around this by running speculation on the SDE that shares the same marginals as the probability flow. Its discretized transitions are Gaussian, the score it needs comes in closed form from the velocity, drafts of future states are produced by extrapolating the recent velocity history with a low order polynomial, and the drafted states are then verified in one batched forward pass and corrected on the first rejection. A curvature statistic adapts the polynomial order and how many states are drafted at a time. Theorem 1 says the committed chain has exactly the law of the serial stochastic sampler for any draft, and on that basis the paper reports roughly 2.5 times wall clock speedup at batch size one on image, editing and video models.

Critical strengths

(1) The lossless claim is actually proven and honestly scoped. Theorem 1 follows the standard speculative decoding argument one transition at a time, the residual sampler used after a rejection has the right density, and the acceptance probability has a closed form that in turn motivates the adaptive drafter. The authors do not claim equivalence with the deterministic sampler people actually use, and they test the two sources of discrepancy separately.

(2) The method is training free, needs no auxiliary draft model, and drafting costs no network evaluations, so the whole mechanism is one batched verification per block plus cheap algebra. Latency is also measured end to end rather than inferred from evaluation counts.

Critical weaknesses

(1) The paper takes the speculative decoding argument from language models without asking why it works there. In a language model a single token step is memory bound, so verifying several draft tokens in one forward pass costs little more than one token, and the parallelism is close to free. A high resolution forward pass of a diffusion transformer, by contrast, is compute bound even at batch one, so verifying several draft states in one pass costs close to several times a single step. The speedup is then a trade between fewer sequential passes and more work per pass, and whether it comes out positive depends on the acceptance rate and on how saturated the GPU already is. The paper never names this issue, never reports compute, and measures everything at batch one, which happens to be the single setting most favorable to filling an idle GPU with extra work. The one evaluation count in the latency table is ambiguous between the number of verification passes and the number of states evaluated, and if it is passes, the total number of states could exceed what the full sampler evaluates. In that case the comparison against caching methods, which reduce compute rather than redistribute it, is not like for like at any serving batch size. Video is the most compute bound case of all, and yet the speedup there comes with no detail on draft length or batching.

(2) Acceptance rate is what the whole method turns on, and it is never measured on a real model. The appendix promises average acceptance, accepted steps per pass, number of passes and an acceptance over time curve, but none of it appears for any real backbone. The only acceptance numbers in the paper are on a 2D Gaussian mixture, and the drafting order, horizon, diffusion scale and correction ablations all live there as well. Since losslessness guarantees quality by construction, acceptance and pass counts are the only evidence that could explain where the real model speedup comes from, and that evidence is missing.

(3) The novelty is modest and the baselines are incomplete. Exact speculative sampling for samplers with Gaussian transitions, including the reflection coupling used here after a rejection, already exists, and converting the flow ODE to its SDE with the score velocity identity is a standard step, as the authors acknowledge. What is new is the polynomial drafter and the adaptive rule, but those are evaluated only on the toy. Meanwhile the most natural training free competitor is missing. The drafter integrates a second order extrapolation of the velocity, which is what a multistep ODE solver does, so a higher order solver at a matching number of evaluations would spend roughly the same compute with no verification machinery, and although it is cited it is not compared. Parallel samplers based on Picard iteration are neither cited nor compared. Finally, the editing and video results have no numbers, only plots, and the speedups printed on the sample figures disagree with the main table.

Questions for authors

(1) Is the evaluation count in the latency table the number of batched verification passes or the number of states evaluated? Please report both, with total compute per image relative to the full sampler.

(2) Please report acceptance rate, accepted steps per pass, number of passes and the acceptance over time curve for each real model, and the maximum draft length used.

(3) Please report latency and throughput at batch 4 and 8 for your method and the caching baselines. If the speedup shrinks with batch size, say so and characterize the regime in which the method helps.

(4) Are the serial forward passes needed after each rejection included in the counts and latency?

(5) For the distribution test on the real model, what feature representation, sample size and bandwidth were used? The appendix shows the same test missing a sizable likelihood drop on the toy.

6 comments

We thank the reviewer for their detailed feedback and address each concern below.

W1 / Q1. Why should speculative verification accelerate a compute-bound diffusion transformer? What does the model-evaluation count represent?

A1: The fundamental benefit of speculative decoding is not limited to memory-bound autoregressive models. Its broader principle is to replace sequentially dependent model evaluations with speculative proposals that can be verified in parallel. This applies to any iterative computation where future states depend on previous outputs, provided that future states can be predicted inexpensively and verified efficiently.

In autoregressive language models, speculative decoding additionally benefits from shared-prefix computation and favorable hardware utilization. In flow matching, the source of acceleration is different: each sampling step depends on the previous state, making the original trajectory inherently sequential. SpecFlow uses previously computed velocities to predict multiple future states without additional network evaluations, then verifies these states together in a batched forward pass. Thus, instead of executing every target-model evaluation sequentially, multiple speculative positions can be evaluated concurrently, reducing the number of sequential verification stages.

Importantly, this does not imply that evaluating multiple states costs the same as evaluating one state. For compute-bound diffusion transformers, the achievable speedup depends on the cost of batched verification, the number of accepted speculative steps, and rejection overhead. However, being compute-bound does not itself rule out acceleration: the relevant question is whether the reduction in sequential execution time outweighs the additional parallel computation.

In the paper, in Section 3.3.1 and Algorithm 1 describe how SpecFlow drafts future states using stored velocities, verifies them in parallel, and corrects the first rejected transition. Section 4 (Evaluation Protocol) explicitly states that speculative states are batched on a single GPU and that speedup is measured using end-to-end wall-clock latency rather than inferred from model-evaluation counts.

Appendix B.6 (Table 10) reports the following BAGEL results at 1024×1024 on a single NVIDIA H100 GPU: These reported measurements show that SpecFlow's acceleration is evaluated in actual wall-clock time, including drafting, batched verification, acceptance testing, rejection handling, and residual correction. Therefore, the reported speedup is not obtained by assuming that a batched forward pass has the cost of a single-state evaluation.

Clarification of model-evaluation counts: The reported model evaluations represent the average number of sequential target-model forward passes required to complete sampling, including batched verification and additional evaluations after rejection. Each batched verification counts as one forward pass, even when multiple speculative states are evaluated together. End-to-end speedup is measured separately using actual wall-clock latency, which includes all computational overhead.


W2 / Q2. Please report acceptance rate, accepted steps per pass, number of passes, acceptance over time, and maximum draft length for each real model.

A2: We thank the reviewer for the suggestion, We have provided the number of forward passes as NFE in table 1, and other details are present in the table given below.

More experiments: We evaluated the SpecFlow on BAGEL text-to-image generation, BAGEL image editing, and HunyuanVideo generation. We recorded acceptance decisions throughout the complete sampling trajectory.

Model / taskSpeedupAcceptanceAccepted steps/passTotal model passesMax draft
BAGEL T2I (553 prompts)2.52×0.704.9019.88
BAGEL editing (110 instructions)2.40×0.664.7120.58
HunyuanVideo (16 prompts)2.32×0.654.6221.68

All configurations use a 50-step reference sampler. The total model-pass column includes the additional model evaluations required by the measured execution.

Acceptance over sampling time: The acceptance profiles show a similar pattern across the three evaluated tasks. Acceptance is highest in the middle of the sampling trajectory, where the velocity field is relatively predictable, and lower near the noise and data endpoints. In particular, most full-length speculative blocks are accepted in the middle region, whereas acceptance decreases substantially near the final data endpoint. This also explains why adaptive drafting can be more effective than applying equally aggressive speculation throughout the trajectory.


AuthorsOpenReview ↗

Rebuttal Extended

W2 / Q3. Please report batch-4 and batch-8 latency and throughput against caching baselines. Does the speedup decrease as batch size increases?

A3: Speculative decoding accelerates sequential generation by evaluating future states in parallel. As batch size increases, GPU utilization also increases, potentially reducing the additional benefit of speculative parallelism. However, this does not necessarily eliminate the speedup.

New batch-scaling experiments: To evaluate this, we tested SpecFlow on BAGEL using a single NVIDIA H100 GPU at batch sizes B∈{1,4,8,16}B\in\{1,4,8,16\}, with 100 prompts per setting and the same 50-step reference sampler.

MethodB=1B=4B=8B=16
SpecFlow2.52×2.38×2.37×2.29×

The results show that SpecFlow maintains substantial acceleration even at larger batch sizes. Speedup decreases only moderately from 2.52× at batch size 1 to 2.29× at batch size 16, retaining approximately 91% of its initial relative speedup. This demonstrates that the benefit of speculative execution remains significant across the evaluated batch sizes.

At still larger batch sizes, we expect the advantage from parallel verification to diminish as GPU resources become increasingly utilized. This is a general hardware-efficiency trade-off in speculative execution rather than a limitation unique to SpecFlow.

In summary, SpecFlow maintains over 2.2× acceleration across all evaluated batch sizes, demonstrating that its benefits extend beyond single-request inference.


W3. What is novel compared with existing exact speculative sampling, Gaussian reflection coupling, and FlowCast?

Response. The key novelty of SpecFlow is enabling exact, training-free speculative sampling for pretrained Flow Matching (FM) models, whose standard deterministic transitions do not directly support probabilistic verification. Our contribution has three main components:

1. Enabling exact speculation for FM: Deterministic FM transitions are Dirac measures, making conventional likelihood-ratio verification ineffective (Section 3.2, Eq. 6). SpecFlow overcomes this by constructing a marginal-preserving stochastic realization with non-degenerate Gaussian transitions (Section 3.3, Eqs. 7–10), enabling exact probabilistic verification and correction.

2. No additional model or training: SpecFlow recovers the required score estimate directly from the pretrained FM velocity (Eqs. 8–9; Appendix A.1.3), eliminating the need for an auxiliary score network. It further introduces curvature-adaptive drafting that adjusts prediction order and horizon using stored velocities, without additional drafting network evaluations (Section 3.3.2; Appendix A.1.7).

3. Distributional exactness independent of drafting: Theorem 1 and Appendix A.1.4 prove that the complete committed trajectory exactly reproduces the chosen discretized stochastic target for any valid draft. Thus, drafting quality affects efficiency, not correctness.

Distinction from prior work: FlowCast uses approximate velocity-error verification and does not guarantee distribution preservation. Existing exact stochastic speculative samplers assume an available stochastic target with evaluable transition densities. SpecFlow specifically addresses the deterministic FM setting by constructing such a target from the pretrained velocity model and combining it with training-free adaptive drafting. We build on established ODE/SDE equivalence and Gaussian reflection coupling rather than claiming these components as new individually.

In summary, SpecFlow's contribution is the FM-specific construction that makes exact speculative verification possible, together with a training-free adaptive drafting procedure—not a reinvention of existing acceptance or correction mechanisms.


AuthorsOpenReview ↗

Rebuttal Extended

W3. Why are higher-order ODE solvers and parallel samplers not included as direct baselines?

A3: Higher-order solvers and parallel-in-time methods are relevant, but their acceleration mechanisms differ from SpecFlow. Higher-order solvers such as DPM-Solver++ and UniPC improve numerical integration efficiency, while Picard-style methods parallelize iterative trajectory refinement. SpecFlow instead reduces sequential sampling dependencies through inexpensive drafting, parallel verification, and exact correction, preserving its chosen stochastic target distribution.

Still we take the suggestion and address the request for a higher-order solver baseline. We evaluate DPM-Solver++(2M) on BAGEL using all 553 GenEval prompts, with deterministic Full-50 as the fidelity reference.

MethodSpeedup ↑LPIPS ↓SSIM ↑PSNR ↑
Full-202.46×0.2760.81418.8
DPM-Solver++(2M)-202.46×0.2790.81018.6
SpecFlow2.52×0.0480.96031.7

At comparable speedup, the SpecFlow variant achieves substantially better fidelity to Full-50.

We will include this comparison and clarify the distinction between numerical solver acceleration and speculative execution in the final version.


W3. Why do the speedups shown in the figures differ from those in the main table?

A4: We believe there may be a misunderstanding regarding the experimental settings being compared. Table 1 reports image-generation results on BAGEL and FLUX, whereas Appendix Figures 1 and 2 report image-editing and video-generation results under different tasks and model configurations. Consequently, their speedup values are not expected to be identical. All speedups are measured consistently using end-to-end wall-clock latency relative to the corresponding full-step sampler, as defined in Section 4. Thus, the differences reflect the experimental settings rather than inconsistent speedup calculations.


Q5. What feature representation, sample size, and bandwidth were used for the real-model MMD test?

A5: We clarify the implementation details of our real-model MMD evaluation on BAGEL reported in Appendix A.2 (Table 2).

Feature representation: We use L2-normalized DINOv2 ViT-L/14 CLS embeddings (1024 dimensions) as the primary representation and Inception-v3 pool3 features (2048 dimensions) as a complementary representation.

Sample size: We generate 4,000 independent images per sampler at 1024×10241024\times1024 resolution, with prompts sampled uniformly from the 553 GenEval prompts and independent random seeds. We conduct 20 independent trials, each comparing 200 images per distribution using disjoint subsets. The target--target calibration uses independently generated target samples.

MMD computation: We use the unbiased MMD² estimator with a Gaussian RBF kernel. The bandwidth is selected using the median heuristic on the pooled 400 feature vectors in each trial. We additionally evaluate a combined multi-scale RBF kernel with bandwidths 0.5σ0.5\sigma, σ\sigma, and 2σ2\sigma to assess sensitivity to kernel scale.

Statistical testing: Each trial uses 1,000 label permutations at significance level α=0.05\alpha=0.05. We report mean MMD² ± standard deviation and rejection counts over 20 trials, including an independent target--target calibration control.

As reported in Table 2, the deterministic--stochastic and target--SpecFlow comparisons each yield only 1/20 rejections, matching the target--target control, whereas TeaCache and FlowCast are rejected in 18/20 and 19/20 trials, respectively. These results provide empirical evidence of distributional agreement under the evaluated tests, complementing the theoretical exactness guarantee of Theorem 1.

We will include these implementation details in Appendix A.2 to make the evaluation fully reproducible.

Fina; Remarks

We sincerely thank the reviewer for their constructive feedback and thoughtful evaluation. We hope that the clarifications, theoretical explanations, and additional experimental comparisons adequately address the concerns raised.

Based on these clarifications and new experimental evidence, we request the reviewer to reconsider their assessment and increase their score.

Reviewer_H6ZsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SpecFlow builds on FlowCast’s draft-and-verify approach by introducing SDE-based stochastic transitions and replacing velocity-error checks with probabilistic verification and correction. It preserves the distribution of its chosen stochastic target sampler and reports over 2.5× speedup across image generation, editing, and video generation with comparable quality.

Critical strengths
  1. The method is training-free and requires no auxiliary draft network, making it potentially practical for accelerating existing flow-matching models.
  2. Building on FlowCast, the paper further explores speculative sampling for flow matching. The motivation for moving from approximate velocity checks to probabilistic verification is clear.
Critical weaknesses
  1. The meaning of “lossless” acceleration needs clearer qualification. The guarantee applies to the newly constructed stochastic target, rather than the original Full-50 sampler. Similar aggregate quality scores do not establish that individual images or videos remain unchanged.

  2. Moving from FlowCast to SpecFlow replaces local velocity-MSE checks with probabilistic verification. However, how does this address global deviation from the original FM trajectory? The relationship between local verification and potentially accumulated global error needs clearer explanation.

  3. Like caching methods, SpecFlow exploits temporal redundancy between neighboring denoising steps. The paper should more clearly explain their connections and differences. The caching comparison is also limited to TeaCache, making the advantage over other caching approaches difficult to assess.

  4. Ablations on real backbones are insufficient. Although partial controls are provided in the appendix, the effects of changing the sampler and introducing speculative execution are not sufficiently separated in terms of generation quality and runtime.

  5. The main text lacks an explanatory method diagram, making the drafting, verification, and correction process harder to follow. The appendix also lacks qualitative video comparisons to illustrate the effects of acceleration.

Questions for authors
  1. To clarify the claimed losslessness, could you report LPIPS, SSIM, and PSNR relative to Full-50 outputs, using the same prompts and initial latents?

  2. Regarding Weakness 2, how does local probabilistic verification control global deviation from the original FM sampler? Could you discuss the relevant error bounds and their assumptions?

  3. Could you compare against more advanced caching methods beyond TeaCache? It would also be useful to evaluate SpecFlow’s quality–speed trade-off at higher acceleration ratios.

  4. Could you provide more qualitative examples comparing accelerated and original video generation?

4 comments

We thank the reviewer for the constructive feedback and for recognizing the practical motivation of our training-free approach. We address each concern below, distinguishing results already in the submission from additional experiments.

Q1: What exactly is lossless? Can you report LPIPS, SSIM, and PSNR relative to Full-50?

A1: By lossless, we mean that SpecFlow preserves exactly the probability distribution of the chosen discretized stochastic sampler, rather than merely producing visually similar outputs. Specifically, each speculative proposal is either accepted through probabilistic verification or corrected using the exact residual distribution. This ensures that every committed transition, and consequently the entire generated trajectory, follows the same distribution as serial stochastic sampling, regardless of drafting accuracy or rejection frequency.

In the paper, we formally established this in Theorem 1, with the complete proof in Appendix A.1.4. The Scope of Exactness discussion clarifies that the guarantee is distributional, not pathwise: SpecFlow need not generate identical images to deterministic Full-50. Importantly, Appendix A.2 (Table 2) also empirically evaluates agreement with the original deterministic FM sampler on BAGEL, finding no detectable distributional difference under the reported MMD test (1/20 rejections, matching the independent target--target control).

Thus, losslessness is theoretically guaranteed relative to the stochastic target, while agreement with deterministic FM is additionally supported by empirical evidence, rather than claimed as an exact theoretical guarantee.

Additional experiments: Following the reviewer's suggestion, we measured per-image fidelity on BAGEL at 1024×1024 using GenEval dataset, with identical initial latents across methods. All metrics are computed relative to deterministic Full-50.

MethodSpeedupLPIPS ↓SSIM ↑PSNR ↑
Full-251.98×0.2250.84520.3
Full-202.46×0.2760.81418.8
DPM-Solver++(2M)-202.46×0.2790.81018.6
TaylorSeer-52.11×0.2640.81920.1
SpecFlow2.52×0.0480.96031.7

At 2× speedup, SpecFlow achieves substantially lower LPIPS than direct step reduction and TaylorSeer, indicating better preservation of individual Full-50 outputs at comparable latency.

We will make the distinction between theoretical exactness and empirical fidelity more prominent in the final revision.


Q2: How does local verification control global error? What are the relevant bounds and assumptions?

A2: Our key result is that local probabilistic verification with exact residual correction guarantees global distributional exactness with respect to the stochastic target, rather than merely bounding accumulated speculative error.

In the paper, Theorem 1 (Section 3.3.3) and its complete proof (Appendix A.1.4) establish this guarantee in two steps. First, at each transition, the exact acceptance rule (Eq. 12) preserves the overlapping probability mass between the draft and target distributions, while residual correction (Eq. 14) restores the remaining target probability mass. Thus, every committed transition follows the exact target kernel, whether the proposal is accepted or rejected. Second, applying this argument sequentially across speculative blocks proves that the entire committed trajectory has exactly the same joint distribution as the serial stochastic sampler, regardless of drafting accuracy or rejection frequency.

This result requires valid draft and target densities, exact likelihood-ratio acceptance, correct residual sampling, and the conditional-independence assumptions detailed in Appendix A.1.4. Therefore, speculation introduces zero additional distributional error at every step and over the complete trajectory, rather than an error that accumulates over time.

Relation to deterministic FM: The distinction between stochastic-target exactness and deterministic FM fidelity is addressed in Section 3.3.3 and Appendices A.1.1--A.1.2. Appendix A.1.1 proves that the continuous-time stochastic realization preserves the original FM marginals under the ideal velocity and score assumptions, while Appendix A.1.2 derives the finite-step stochastic transition used by SpecFlow. Appendix B.2 (Table 6) empirically compares deterministic FM, serial stochastic sampling, and SpecFlow, showing close agreement at finer discretizations on the analytic model. Appendix B.4 (Table 8) further shows that learned-velocity error can alter the stochastic target, while exact speculative correction continues to preserve that chosen target.

Thus, the theoretical guarantee eliminates accumulated error from speculative execution relative to the stochastic target, while the separate discrepancy between that target and deterministic FM is explicitly characterized and evaluated in the paper.

AuthorsOpenReview ↗

Rebuttal (Extended)

Q3: How is SpecFlow different from caching? Can you compare against stronger caching methods and higher acceleration ratios?

A3: Although both approaches exploit temporal redundancy, their acceleration mechanisms differ. Caching methods, such as TeaCache, reduce computation by reusing previously computed features, potentially introducing approximation errors. SpecFlow, instead, predicts future sampling states using stored velocities and verifies them through probabilistic acceptance and exact residual correction. This allows SpecFlow to accelerate sampling while preserving the distribution of its chosen stochastic target.

In the paper, Section 2 discusses these differences, and Appendix A.2 (Tables 2–3) compares distributions against TeaCache and FlowCast. On BAGEL, SpecFlow yields only 1/20 MMD rejections, matching the independent target--target control, versus 18/20 for TeaCache and 19/20 for FlowCast. Tables 1, 4, and 5 further evaluate quality--speed trade-offs across sampling budgets, including over 10× acceleration.

Thus, the key distinction is that SpecFlow provides a theoretical distribution-preservation guarantee relative to its stochastic target, whereas conventional feature-caching approaches generally trade approximation error against computational savings. Our existing experiments additionally compare their practical quality and efficiency across multiple operating points.

New experiments: We added caching baselines using the same BAGEL checkpoint, hardware, resolution, prompts, and initial latents.

Each cell reports LPIPS ↓ / CLIP-IQA ↑.

Method2×3×4×6×8×
SpecFlow0.051 / 0.8500.113 / 0.6820.134 / 0.6410.271 / 0.5560.323 / 0.502
TeaCache0.103 / 0.7900.208 / 0.6160.285 / 0.5840.326 / 0.4890.399 / 0.416
DeepCache (adapted to transformers)0.292 / 0.6930.321 / 0.5570.440 / 0.5220.490 / 0.4590.532 / 0.399
TaylorSeer0.250 / 0.7990.499 / 0.5100.556 / 0.4980.754 / 0.4310.763 / 0.419

Hence, SpecFlow performs better across all dimensions.

We will clarify these distinctions in the final manuscript.


W4: The sampler change and speculative acceleration are not sufficiently separated on real models.

A4: We thank the reviewer for raising this point. As far as we understand, the concern is distinguishing (i) the effect of replacing deterministic FM with a stochastic target and (ii) the additional effect of speculative execution. We have explicitly investigated these effects in the submitted paper, including on the real BAGEL backbone.

Real-model validation (Appendix A.2, Table 2): We isolate the sampler change by comparing deterministic FM with serial stochastic sampling, and speculation by comparing serial stochastic sampling with SpecFlow. On BAGEL, both yield only 1/20 MMD rejections, matching the independent target--target control, with mean MMD² values of 0.3×10−40.3\times10^{-4} and 0.1×10−40.1\times10^{-4}, respectively. Neither shows a detectable distributional discrepancy, supporting empirical agreement between stochastic and deterministic FM, while SpecFlow's preservation of its stochastic target is theoretically guaranteed.

Further controlled analysis (Appendix B.2, Table 6): We repeat this separation on an analytic Gaussian-mixture model across 50, 25, and 10 sampling steps. Deterministic and stochastic sampling remain statistically indistinguishable in nearly all trials, while stochastic sampling and SpecFlow are not distinguished in any trial. This further supports the separation between target-sampler effects and speculative execution.

Sensitivity analysis (Appendix B.5, Table 9): We also vary the diffusion scale from g=0.25g=0.25 to g=3.0g=3.0. The results show that larger diffusion scales can reduce target quality even though SpecFlow remains consistent with its chosen stochastic target. This demonstrates that stochastic-target fidelity and speculative correctness are separate considerations.

Clarification: Our experiments separate the distributional effects of sampler change and speculation. If the reviewer instead seeks a decomposition of end-to-end latency and generation quality across deterministic FM, serial stochastic sampling, and SpecFlow on the same real backbone, we would appreciate clarification and would be happy to address it in the revision.

We hope this clarifies the distinction already examined in the paper and thank the reviewer for helping us improve its presentation.


Thanks for your valuable suggestion, we will add a method diagram and qualitative video comparisons in the final version.


Final Remarks

We sincerely thank the reviewer for their constructive feedback and thoughtful evaluation. Based on these clarifications and new experimental evidence, we request the reviewer to reconsider their assessment and increase their score.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer H6Zs,

It is a gentle reminder to acknowledge our rebuttal and, if there are no further concerns, we request to increase the scores.

Best Regards,

Authors

Thank you for the response and additional experiments. I suggest clarifying in the main text that the “lossless” guarantee applies to the distribution of the chosen stochastic target sampler. The new caching comparisons cover only BAGEL. DeepCache is an older baseline, and several methods introduced after TaylorSeer have already been evaluated on FLUX. Could you compare SpecFlow with a more recent method, such as TACache or HiCache, on FLUX at comparable end-to-end speedups? Could you also clarify whether the caching baselines used their original ODE samplers or were adapted to SpecFlow’s SDE sampler? If they were adapted to the SDE sampler, were their caching parameters retuned for it?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.