- 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
- The method is simple to implement, requires no retraining, and preserves one network evaluation per step.
- The low-NFE RectFlow results are promising. The full-vector trajectory diagnostics and paired-bootstrap analysis provide useful evidence beyond endpoint FID.
Critical weaknesses
- 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 is arbitrarily chosen from a restricted sweep on 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.
-
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.
-
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
- 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.
- 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
Thank you for the extended sweep, the finite-step analysis, and especially the fixed- controls for Table 2. These additions directly address my requested sensitivity study and clarify that 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.
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 =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 =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:
| NFE | Fixed | Full | Reduction |
|---|---|---|---|
| 4 | 0.5177 | 0.2413 | 53% |
| 8 | 0.3772 | 0.2223 | 41% |
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 Sensitivity
The extended CIFAR-10 sweep at 4 NFE favors =2.75 (FID 22.61 versus 24.43 at =2). We additionally tested CelebA-HQ at 4 NFE using Full AdaptDT:
| 1.5 | 2.0 | 2.75 | |
|---|---|---|---|
| FID ↓ | 85.39 | 74.44 | 77.79 |
The best tested values differ across datasets, showing that increasing does not uniformly improve performance. We retain =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 =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.
Novelty, Analysis, and Ablations
Thank you for the detailed assessment. We address novelty and serial-solver comparisons, finite-step analysis and λ 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 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-h Analysis & Extended λ-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.