acceptodds
Under review as a conference paper at ICLR 2027

Correcting Trajectory Drift in Flow Matching Models

Abstract

Flow-matching generative models are typically sampled by numerically integrating a learned velocity field using a small number of deterministic steps. In this regime, fixed-timestep solvers can drive inference trajectories away from the continuous learned trajectory, causing the learned velocity field to be evaluated at increasingly shifted states and potentially compounding error. We study this inference trajectory drift and propose a training-free sampler that combines cached finite-difference acceleration with discrepancy-guided online control. At each step, the method forms an Euler update and a finite-difference acceleration update from consecutive velocity evaluations, and conservatively shrinks the timestep when their magnitude and angular disagreement is large. The sampler uses one network evaluation per step, stores only the previous velocity, and requires no retraining. A scalar gain modulates the acceleration correction, while discrepancy-guided step control provides complementary online stabilization; their relative contributions depend on the model and sampling regime. The strongest gains occur in low-NFE RectFlow sampling, where the method outperforms Euler, standard multistep baselines, flow-prediction DPM-Solver++ 2M, and Flow-UniPC across the tested NFE range. Full-vector trajectory diagnostics further show reduced state deviation and learned-field discrepancy relative to AB2 at matched NFE. These findings highlight that sampler design for learned generative models should account not only for nominal integration order, but also for the behavior of the realized inference trajectory.

Primary area:
generative models
Keywords:
Diffusion Models, Flow Matching, Efficient Sampling, Adaptive Sampling, Generative Models, Trajectory Drift

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_14fmOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes an inference-only sampler for flow-matching models, combining a cached finite-difference acceleration correction with discrepancy-guided step-size modulation. The method requires one network evaluation per step. Experiments demonstrate improvements particularly in low-NFE RectFlow settings, alongside reduced state deviation and learned-field discrepancy relative to a dense numerical reference.

Critical strengths
  1. The method is simple to implement, requires no retraining, and preserves one network evaluation per step.
  2. The low-NFE RectFlow results are promising. The full-vector trajectory diagnostics and paired-bootstrap analysis provide useful evidence beyond endpoint FID.
Critical weaknesses
  1. The novelty beyond existing multistep coefficient modifications needs clearer justification.

The technical contribution appears marginal, given that difference-based correction (often seen in advanced numerical ODE solvers or predictor-corrector frameworks) is widely adopted in prior work, such as ParaFlow [1]. Since ParaFlow is also a training-free method (despite focusing on parallel sampling), a thorough discussion and comparison against existing training-free ODE samplers (e.g., high-order solvers utilizing historical velocity differences) are expected to justify the novelty of the proposed correction.

Moreover, the acceleration update is a direct instance of the established parameterized Adams–Bashforth family. It lacks novel theoretical or empirical insights tailored to learned flows, aside from tuning a correction coefficient. Finally, the default setting of λ=2\lambda = 2 is arbitrarily chosen from a restricted sweep on [1,2][1, 2] using SiT at 30 steps, where the best adaptive FID occurs at the boundary. A more comprehensive sensitivity study is required to validate this hyperparameter choice.[1] ECCV ParaFlow: Parallel Sampling for Flow Matching Models.

[1] ECCV ParaFlow: Parallel Sampling for Flow Matching Models.

  1. The independent benefit of discrepancy-guided adaptation remains insufficiently isolated.

    Table 5 shows that fixed acceleration accounts for most of the RectFlow improvement. Adding adaptation barely changes FID at 4 NFE, although the benefit increases at 8 and 16 NFE. SiT also shows substantial improvements in some PRDC metrics, suggesting regime-dependent benefits. However, Table 2 omits the fixed-acceleration variant with (λ=2), so the trajectory diagnostics cannot determine how much drift reduction comes from adaptation rather than coefficient adjustment.

  2. The writing appears heavily AI-generated. Specifically, in Sections 3.3 and 3.4, each subsection abruptly jumps from a single bolded sentence directly to a formula without proper context or transition. This format is extremely difficult to follow and gives the impression of a superficial, AI-generated summary rather than a rigorous academic derivation.

Questions for authors

Questions

  1. Does increasing λ beyond 2 further improve performance? Could the authors extend the sensitivity study accordingly? Mathematical analysis explaining when an amplified correction reduces trajectory drift would also substantially strengthen the contribution.
  2. Could the authors add the fixed-acceleration variant with λ = 2 and adaptation disabled to Table 2, using the same paired samples, matched NFE, and physical-time evaluation, to isolate the effect of step-size compression on both trajectory-deviation metrics?
3 comments
AuthorsOpenReview ↗

Novelty, Analysis, and Ablations

Thank you for the detailed assessment. We address novelty and serial-solver comparisons, finite-step analysis and λ\lambda sensitivity, AdaptDT's independent contribution, and presentation.

1. Novelty vs. ParaFlow [1] & Strong Serial Solvers (W1)

We agree that finite differences and Adams-Bashforth-type corrections are established numerical techniques; we do not claim these identities as new. ParaFlow [ECCV 2026, Sec. 4.4] uses divided-difference extrapolation to estimate unevaluated future velocities within a parallel-in-time solver. Our method instead uses consecutive velocities along the realized serial trajectory to form a gain-modulated state correction, coupled with sample-dependent magnitude-and-angle disagreement control. It requires one network evaluation per step and no retraining.

Importantly, ParaFlow primarily targets wall-clock acceleration through parallel execution while preserving autoregressive sampling quality. Our method targets improved fidelity and reduced trajectory drift at matched NFE during serial sampling. An FID-only comparison would not capture ParaFlow's primary speed objective or isolate our matched-NFE contribution. We therefore focus our quantitative comparisons on serial solvers and will discuss ParaFlow as complementary related work. Our contribution is the complete inference procedure and its demonstrated behavior in low-NFE learned-flow sampling, not the finite-difference formula alone.

Strong serial baselines are included in Table 1 of the paper. On RectFlow CIFAR-10, our default λ=2\lambda=2 method achieves the lowest FID among evaluated solvers at 4–64 NFE (at 4 NFE: 24.43 versus Flow-UniPC 40.16 and Flow-DPM-Solver++ 2M 46.69). On CelebA-HQ, ours improves over AB2 at 4 NFE (74.44 versus 103.92); an additional Flow-UniPC evaluation gives 115.14.

2. Finite-hh Analysis & Extended λ\lambda-Sensitivity (W1, Q1)

For a smooth field on a uniform fixed-step grid, defining local error as the exact increment minus numerical update gives

=\frac{1-\lambda}{2}h^2\dot v_i +\frac{2+3\lambda}{12}h^3\ddot v_i+O(h^4).$$ Minimizing the squared truncated local error for nonzero material acceleration gives $$\lambda^{*}(h)=1+\frac{5h}{6} \frac{\langle\dot v_i,\ddot v_i\rangle} {\lVert\dot v_i\rVert_2^2}+O(h^2).$$ Thus $\lambda=1$ recovers AB2 on uniform steps without adaptation. Fixed $\lambda\neq1$ is generally first-order, but $\lambda>1$ can reduce finite-step error when derivative alignment is favorable. We observe positive late-time derivative alignment on 256 dense RectFlow trajectories (median $\kappa=\langle\dot v_i,\ddot v_i\rangle/\lVert\dot v_i\rVert_2^2=13.14$ for $t\in[0.75,0.98]$), consistent with this finite-step analysis. This local analysis does not establish stability of the adaptive scheme or predict FID-optimal $\lambda$. We extended the RectFlow CIFAR-10 sweep at 4 NFE: | $\lambda$ | FID ↓ | | ---: | ---: | | 1.5 | 27.59 | | 2.0 | 24.43 | | 2.5 | 22.87 | | 2.75 | **22.61** | | 3.0 | 22.71 | | 3.5 | 23.75 | | 4.0 | 26.01 | The best tested value is $\lambda=2.75$, improving on the submitted default 2.0. At this setting, dense-ODE terminal RMSE is 0.221 versus 0.290 for AB2. We retain $\lambda=2$ for the submitted cross-model benchmarks without retrospective dataset-specific retuning. The RectFlow 4-NFE sweep and SiT 30-step ablation probe different regimes. As $h\to0$, $\lambda^{*}(h)\to1$; we make no second-order convergence claim for fixed amplified $\lambda$ or the complete adaptive method. ### 3. Table 2 Fixed $\lambda=2$ & AdaptDT Isolation (W2, Q2) We independently evaluated the fixed-acceleration $\lambda=2$ variant using the same 256 paired RectFlow samples, matched NFE, and physical-time evaluation as Table 2. Its state-deviation AUC at 4/8 NFE is 0.1174/0.0509, and learned-field discrepancy AUC is 0.2731/0.1300, matching Full to four decimal places. The unrounded state-AUC differences are approximately $10^{-5}$. Thus, the Table 2 improvement over AB2 is primarily driven by amplified acceleration, with negligible additional AUC reduction from AdaptDT on this cohort. AdaptDT provides complementary, regime-dependent benefits. On RectFlow CIFAR-10, adding AdaptDT improves FID from 12.65 to 12.04 at 8 NFE and from 10.23 to 8.91 at 16 NFE; at 4 NFE the improvement is small (24.47 to 24.43). On SiT at 30 steps and $\lambda=2$, AdaptDT improves precision from 69.70 to 81.54 and density from 0.9053 to 1.2292. **Presentation (W3).** We will improve Sections 3.3–3.4 by adding clearer explanations and transitions between equations and the proposed algorithm. As disclosed in the paper, AI assistance was used for writing polish and related-work discovery; the methodology, experiments, and scientific contributions are our own work.

Thank you for the extended λ\lambda sweep, the finite-step analysis, and especially the fixed-λ=2\lambda=2 controls for Table 2. These additions directly address my requested sensitivity study and clarify that λ=2\lambda=2 was not optimal in the new 4-NFE sweep.

The new controls also confirm an important part of my original concern: fixed amplified acceleration accounts for nearly all of the Table 2 trajectory-metric improvements, while the additional reduction from AdaptDT is negligible on this cohort. The Gaussian-mixture experiment demonstrates a complementary benefit under a constructed off-trajectory perturbation, but it does not resolve the attribution on the original RectFlow trajectories. The paper should therefore attribute the Table 2 gains primarily to fixed acceleration and scope the claims concerning discrepancy-guided drift reduction accordingly.

Two central concerns remain. First, the algorithmic novelty still appears to lie mainly in combining an established Adams–Bashforth-type correction with a heuristic discrepancy controller. The local truncation expansion and derivative-alignment measurements explain why finite-step gains are possible, but they do not provide a stability or global-error result for the complete adaptive method. Second, the independent benefit of AdaptDT remains regime- and metric-dependent, rather than a generally demonstrated trajectory-drift improvement.

I therefore maintain my overall rating of 1.

AuthorsOpenReview ↗

Additional Evidence: Independent Contribution of AdaptDT on SD3.5

Thank you for the follow-up. We provide further experimental evidence to clarify AdaptDT's independent contribution and address the remaining concerns.

1. RectFlow: Fixed vs. Full

On RectFlow CIFAR-10, Fixed λ\lambda=2 and Full have nearly identical trajectory diagnostics at 4/8 NFE (state-AUC 0.1174/0.0509 and learned-field discrepancy AUC 0.2731/0.1300, matching to reported precision). This shows that amplified acceleration drives most of the Table 2 trajectory improvements on this cohort. AdaptDT is designed to respond to disagreement between candidate updates, so its benefit depends on the learned-flow regime.

Importantly, the FID ablation demonstrates additional benefits even on RectFlow: adding AdaptDT improves FID from 12.65 to 12.04 at 8 NFE and 10.23 to 8.91 at 16 NFE. Thus, while amplified acceleration accounts for most of the trajectory-AUC improvement, AdaptDT provides additional generation-quality gains at these budgets.

2. Independent Drift Reduction from AdaptDT on SD3.5

To directly examine whether AdaptDT independently reduces trajectory drift on another pretrained generative model, we performed a Fixed-vs-Full ablation on Stable Diffusion 3.5-medium, holding λ\lambda=2 and changing only AdaptDT.

Using the same 64 paired noise/prompt samples, we measured state-deviation AUC against a dense Euler reference at matched physical times and NFE:

NFEFixedFullReduction
40.51770.241353%
80.37720.222341%

Full achieves lower state-AUC on all 64 paired samples at both 4 and 8 NFE, demonstrating consistent improvement across individual trajectories rather than merely in the mean.

Crucially, this also translates to generation quality. At 4 NFE, FID against real images decreases from 149.03 (Fixed) to 82.49 (Full), evaluated using the first 1,000 COCO prompts.

These experiments directly isolate AdaptDT's contribution to both trajectory deviation and generation quality. On RectFlow, fixed amplification already performs well; on SD3.5, AdaptDT substantially reduces the trajectory deviations produced by fixed amplification. The near-identical RectFlow trajectory metrics therefore cannot be generalized to suggest that AdaptDT provides negligible drift reduction. The SD3.5 results demonstrate a substantial independent benefit on real learned-flow trajectories.

3. Cross-Dataset λ\lambda Sensitivity

The extended CIFAR-10 sweep at 4 NFE favors λ\lambda=2.75 (FID 22.61 versus 24.43 at λ\lambda=2). We additionally tested CelebA-HQ at 4 NFE using Full AdaptDT:

λ\lambda1.52.02.75
FID ↓85.3974.4477.79

The best tested values differ across datasets, showing that increasing λ\lambda does not uniformly improve performance. We retain λ\lambda=2 as the shared default across the submitted benchmarks, without dataset-specific retuning.

4. Implications for the Contribution

The additional experiments clarify the complementary roles of acceleration and adaptation. On RectFlow, amplified acceleration drives the main trajectory gains, while AdaptDT provides additional FID improvements. On SD3.5, AdaptDT independently reduces state deviation by 41–53%, consistently across all tested pairs. On SiT at λ\lambda=2, it also improves precision from 69.70 to 81.54 and density from 0.9053 to 1.2292.

Our contribution is the complete training-free serial sampling procedure, combining gain-modulated correction with discrepancy-guided control at one network evaluation per step. The finite-step analysis motivates amplified correction at coarse steps, while the controlled experiments demonstrate when the adaptive component provides substantial additional benefits.

Together, these results support a practical contribution beyond coefficient tuning: a serial sampler that improves low-NFE generation quality and substantially reduces trajectory drift when fixed amplification alone is insufficient.

Reviewer_Ba4DOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

With only a few steps, each numerical step moves the sample slightly away from the continuous trajectory of the learned ODE and the errors can add-up. The authors propose a training-free sampler that uses one NFE per step and stores only the previous velocity. It combines two parts a cached finite-difference acceleration correction and a discrepancy-guided step-size rule that shrinks the step when the Euler update and the corrected update disagree.

Critical strengths
  1. Large gains at low NFE on RectFlow.
  2. The method is simple and practical. It needs no training, uses one NFE per step, and can directly replace a standard sampler.
  3. The evaluation is broad. It covers several model families and many metrics.
Critical weaknesses
  1. There is no analysis of stability and accuracy of the proposed sampler. And the sampler is actually first-order? I think first order method should lose advantages against AB2 as NFE grows?(Correct me if I'm wrong)
  2. At larger scale, such as SD3.5(Table 7) the method roughly ties AB2.
Questions for authors
  1. Could the authors add a toy experiment where the true velocity is known in closed form, for example a Gaussian mixture?
  2. Could the authors report the quantitative trajectory diagnostics(Table 2) at higher NFE, to show whether the method still tracks the learned ODE better than AB2 there?
2 comments
AuthorsOpenReview ↗

Stability and Additional Experiments

Thank you for the constructive review. We address stability and accuracy, the Gaussian-mixture experiment, higher-NFE trajectory diagnostics, and SD3.5.

1. Stability and Finite-Step Accuracy (W1)

Stability. For fixed uniform steps, the method has characteristic polynomial ρ(ξ)=ξ(ξ−1)\rho(\xi)=\xi(\xi-1) and is zero-stable for all fixed λ\lambda. For the linear test equation x˙=μx\dot{x}=\mu x with real μ<0\mu<0, the absolute-stability interval is hμ∈(−2/(1+λ),0)h\mu\in(-2/(1+\lambda),0) for λ>0\lambda>0. Larger gains narrow this interval. AdaptDT restricts steps under candidate disagreement, but we do not claim a global stability guarantee for the complete adaptive method.

Accuracy. At λ\lambda=1, the uniform-step method recovers classical AB2 without adaptation. Fixed λ≠1\lambda\neq1 is generally first-order. For a smooth field, the local error (exact increment minus numerical update) is

eloc=1−λ2h2v˙i+2+3λ12h3v¨i+O(h4).e_{\mathrm{loc}}=\frac{1-\lambda}{2}h^2\dot v_i+\frac{2+3\lambda}{12}h^3\ddot v_i+\mathcal{O}(h^4).

This explains why amplified correction can reduce error at practical coarse step sizes under favorable derivative alignment, even though AB2 can become more accurate as steps decrease. Our objective is improved finite-NFE sampling, not higher asymptotic order.

2. Controlled Gaussian-Mixture Experiment (Q1)

We used a three-mode 2D Gaussian mixture with known conditional-mean velocity v∗v^*. To study robustness to off-trajectory field errors, we evaluate the perturbed field

vθ=v∗+boff,v_\theta = v^* + b_{\mathrm{off}},

where boffb_{\mathrm{off}} is a smooth perturbation that vanishes along the dense reference trajectory. All samplers evaluate the same perturbed field without access to the reference trajectory.

We measure mean final-output error relative to the known true endpoint (lower is better):

NFEAB2Fixed λ\lambda=2Full AdaptDT
41.7862.0411.608
81.2851.3011.187
160.9650.9390.865
320.7400.5730.505
640.4520.4820.447

At 4 NFE, AdaptDT reduces error from 2.041 (Fixed) to 1.608 (Full); at 32 NFE, Full improves over AB2 from 0.740 to 0.505. This experiment demonstrates robustness to off-trajectory field perturbations, measuring final-output error rather than integration error for the perturbed field alone.

3. Higher-NFE Trajectory Diagnostics (Q2)

We extended the paired RectFlow diagnostics against a dense learned-ODE reference at matched physical times:

NFEState-AUC AB2State-AUC FullField-AUC AB2Field-AUC Full
160.02680.01780.06760.0386
320.00530.00280.01440.0077
640.00130.00260.00320.0032

Full reduces both trajectory metrics relative to AB2 at 16 and 32 NFE. At 64 NFE, AB2 achieves lower state-AUC, while field-AUC is tied to reported precision, consistent with the higher-order advantage at smaller steps. Full nevertheless achieves slightly lower CIFAR-10 FID (7.00 versus 7.07). Thus, numerical trajectory accuracy and image-distribution quality need not rank the solvers identically.

4. SD3.5 and Independent AdaptDT Benefit (W2)

At 20 NFE on SD3.5, Full achieves the best reference FID, KID, precision, density, and coverage among the evaluated one-evaluation-per-step solvers (reference FID 6.32 versus 6.45 for AB2 and 6.66 for Euler). These metrics use a dense learned-ODE reference, not real-image FID.

To examine more aggressive sampling budgets, we additionally isolated AdaptDT on SD3.5-medium, holding λ\lambda=2 and comparing Fixed against Full. State-deviation AUC against dense Euler decreases from 0.5177 to 0.2413 at 4 NFE (53%) and from 0.3772 to 0.2223 at 8 NFE (41%). Importantly, Full achieves lower state-AUC on all 64 paired noise/prompt samples at both budgets, demonstrating consistent improvement across trajectories.

This also translates to generation quality: at 4 NFE, real-image FID using the first 1,000 COCO prompts improves from 149.03 (Fixed) to 82.49 (Full).

Together, these experiments establish the complementary roles of amplified acceleration and discrepancy-guided control. The strongest gains occur at aggressive low-NFE budgets, where reducing trajectory drift is particularly important.

Thank you for your rebuttal! I decide to maintain my score.

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

This paper studies inference trajectory drift problem and proposes a traning free sampler that combines cached finite-difference acceleration with discrepancy-guided online control. Experiments are conducted to show the effectiveness of the propsoed approach.

Critical strengths

The paper is well motivated and relatively easy to follow.

Critical weaknesses

W is not defined.

Equation (8) is just a repeat of equation (4), not much information is provided.

Questions for authors

No

1 comment
AuthorsOpenReview ↗

Presentation Clarifications

Thank you for the positive assessment and for identifying two presentation issues.

(1) W2W_2 denotes the 2-Wasserstein distance between probability distributions. We agree that this should be explicitly defined at its first appearance in Section 3.2 and will add the definition.

(2) We agree that Equation (8) repeats the triangle-inequality decomposition in Equation (4). We will remove the redundant equation and clarify that Theorem 1 applies the standard Euler error recursion to establish an O(h)O(h) bound on sampler-induced deviation from the continuous learned ODE. The triangle inequality separates this numerical error from continuous-model error.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Contribution Clarification and Additional Evidence

We appreciate the reviewers' careful evaluation and the opportunity to clarify our contribution. The principal remaining discussion concerns novelty relative to established multistep methods and ParaFlow. We would like to highlight what the paper and additional experiments establish.

1. Contribution Beyond Existing Numerical Methods

We do not claim that finite-difference acceleration or Adams–Bashforth corrections are new. ParaFlow uses velocity extrapolation within parallel-in-time sampling, primarily targeting wall-clock acceleration. Our method instead targets improved fidelity and reduced trajectory drift at matched NFE during serial sampling, combining gain-modulated acceleration with sample-dependent magnitude-and-angle disagreement control.

The contribution is this complete training-free inference procedure, requiring one network evaluation per step and no retraining, together with its demonstrated performance on learned flows.

The submitted experiments show substantial improvements over classical and modern serial solvers, particularly at aggressive low NFE. On RectFlow CIFAR-10 at 4 NFE, our method achieves FID 24.43 versus 40.16 for Flow-UniPC and 46.69 for Flow-DPM-Solver++ 2M, and achieves the lowest FID among evaluated solvers across 4–64 NFE. On CelebA-HQ, FID improves over AB2 from 103.92 to 74.44 at 4 NFE and from 56.91 to 44.40 at 8 NFE.

2. Independent Contributions of Acceleration and AdaptDT

The additional component-isolation experiments clarify when each component contributes.

On the original RectFlow cohort, Fixed λ\lambda=2 and Full have nearly identical trajectory-AUC metrics at 4/8 NFE, showing that amplified acceleration accounts for most of these trajectory improvements. Nevertheless, AdaptDT further improves RectFlow FID from 12.65 to 12.04 at 8 NFE and 10.23 to 8.91 at 16 NFE.

To directly examine independent drift reduction on another pretrained model, we performed additional Fixed-vs-Full experiments on SD3.5-medium, holding λ\lambda=2 and changing only AdaptDT.

Against a dense Euler reference, AdaptDT reduces state-deviation AUC by 53% at 4 NFE and 41% at 8 NFE. Importantly, Full achieves lower state-AUC on all 64 paired noise/prompt samples at both budgets, demonstrating consistent improvement across individual trajectories.

The benefit also translates to generation quality: at 4 NFE, FID against real images improves from 149.03 (Fixed) to 82.49 (Full) using the first 1,000 COCO prompts.

These results establish that AdaptDT is not merely redundant with amplified acceleration. Its contribution depends on the learned-flow regime, becoming substantial when fixed amplification alone produces larger trajectory deviations. Additional SiT ablations also demonstrate improvements in precision and density.

3. Additional Analysis and Experimental Validation

We addressed the requests for extended λ\lambda sensitivity, finite-step analysis, a controlled Gaussian-mixture experiment, and higher-NFE trajectory diagnostics.

The extended sensitivity study finds that the best tested CIFAR-10 FID at 4 NFE occurs at λ\lambda=2.75, while CelebA-HQ favors λ\lambda=2 among the tested settings. We retain one shared default across the submitted benchmarks rather than retrospectively tuning individual datasets.

The finite-step analysis explains why amplified acceleration can improve coarse-step sampling. Higher-NFE diagnostics show improved trajectory tracking over AB2 at 16 and 32 NFE, while the advantage is not uniform at 64 NFE, consistent with the asymptotic-order distinction. The Gaussian-mixture experiment additionally demonstrates robustness to controlled off-trajectory field perturbations.

We have clarified the numerical foundations and the complementary roles of acceleration and adaptation, and will improve the corresponding exposition in the revision.

We respectfully ask that the contribution be evaluated in its intended setting: practical, training-free, low-NFE learned-flow sampling. The combined procedure delivers substantial matched-compute improvements, with controlled experiments demonstrating the independent benefit of adaptive control on pretrained generative models.

Thank you for your consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.