acceptodds
Under review as a conference paper at ICLR 2027

Formalizing the Sampling Design Space of Diffusion-Based Generative Models via Adaptive Solvers and Wasserstein-Bounded Timesteps

12 views

Abstract

Diffusion-based generative models have achieved remarkable performance across various domains, yet their practical deployment is often limited by high sampling costs. While prior work focuses on training objectives or individual solvers, the broader sampling design problem, specifically solver selection and scheduling, remains largely governed by static heuristics. We propose SDM, a principled, training-free sampling framework that adapts both the numerical solver and the timestep schedule to the intrinsic properties of the diffusion trajectory. By analyzing the PF-ODE dynamics, we show that velocity variation is small in high-noise stages and increases near the data manifold, identifying intervals where solver order is most consequential. In parallel, we introduce an offline-calibrated adaptive scheduling method that explicitly controls the local Wasserstein discretization error and projects the calibrated trajectory to a prescribed NFE budget. We further extend the formulation to a mixed-transition Wasserstein error bound, providing a unified error-propagation view of adaptive scheduling and solver selection within the overall SDM framework. Across standard benchmarks, with extensions to modern ODE samplers, high-resolution synthesis, and text-to-image generation, SDM achieves improved sample quality compared to baseline methods, attaining an FID of 1.93 on CIFAR-10, 2.41 on FFHQ, and 1.98 on AFHQv2, with a reduced number of function evaluations compared to existing samplers. Our code is available at https://anonymous.4open.science/r/sdm-3128.

Primary area:
generative models
Keywords:
Diffusion Models, Probability Flow ODE, Adaptive Sampling, Wasserstein Distance

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_5CFyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes two techniques to enable effective sampling for diffusion models. The authors first characterize the local stiffness of the ODE in terms of the magnitude of the second-order derivative of the diffusion state xtx_t. They then argue that in the early sampling stage, a low-order ODE solver should be applied. At a later sampling stage, high-order ODE solvers should be applied due to the high stiffness of the ODE to be solved. Therefore, they propose to switch the ODE solvers depending on the local stiffness of the ODE.

They also propose an offline adaptive timestep scheduling method. In particular, they introduce a formula for the error tolerance term ηt\eta_t in terms of the variance scheduler σt\sigma_t. Next, each timestep Δti\Delta t_i is locally searched such that the local Wasserstein distance does not exceed the given upper bound ηt\eta_t.

Critical strengths

The proposed offline adaptive timestep scheduling method is interesting and appears to be theoretically well-grounded. The authors derive an inequality relating the step size Δt\Delta t, the local Wasserstein distance StS_t, and the error upper bound ηt\eta_t. Based on this inequality, they sequentially perform a local search to determine each optimal time step Δti\Delta t_i.

Since this approach cannot guarantee a fixed number of function evaluations (NFEs) during sampling, the authors introduce an N-Step resampling method to produce an adaptive timestep schedule that matches a predefined NFE target.

Experimental results show that the proposed adaptive timestep scheduling method leads to better image quality.

Critical weaknesses

(1) The first proposed technique for switching ODE solvers is mentioned in the main paper, yet it is not clear at all when the solver should be switched.

(2) Table 1 indicates that for Heun's method, the newly proposed method SDM does not bring a significant performance gain. In fact, both EDM and SDM produce roughly the same performance. For the case of an adaptive solver, both EDM and SDM again produce comparable performance. This suggests that the proposed offline adaptive timestep scheduling method is not better than EDM scheduling.

(3) In Table 2, why are only the uniform and SDM schedulers compared? A comparison with EDM is also needed, as the results in Table 1 show that EDM is a very competitive method.

Questions for authors

(1) In the introduction, the authors mention the work of (Hu et al., 2024), which employs Wasserstein-based error estimates for timestep optimization, which is very close to the current work. Why do the author not conduct performance comparison with respect to (Hu et al., 2024)?

(2) What are the main differences between (Hu et al., 2024) and the timestep optimization in the paper?

(3) It is not clear to me how D_{\theta} and D_{\sigma} are computed in Proposition 3.1 by only reading the main text.

5 comments

We thank the reviewer for the constructive feedback. We address the reviewer’s concerns below and provide corresponding revisions in the revised manuscript. All revisions are highlighted in blue in the revised PDF, and those addressing the reviewer’s comments are marked with 5CFy in purple. Please note that the table numbers have changed in the revised manuscript (e.g., Table 8 →\rightarrow Table 2, Table 9 →\rightarrow Table 3). We refer to both the original and revised table numbers in our responses for additional clarity.

Clarification on Interpretation of Table 1 (W2)

We would like to clarify how the results in Table 1 should be interpreted. For Heun's method, the EDM baseline achieves FID of 2.17 for AFHQv2, for instance. Applying SDM scheduling alone improves FID by 2.17$\rightarrow$ 2.13. Applying adaptive solver alone improves FID by 2.17$\rightarrow$ 2.10. Finally, integrating both adaptive solver selection and scheduling methods improves FID by 2.17$\rightarrow$ 1.99.

MethodFID
EDM2.17
SDM scheduling2.13
SDM solver2.10
SDM solver + scheduling1.99

The improvements are more pronounced in few-step generation settings, as demonstrated in Table 2 and 9 (Table 8 and 7, before revision). For instance, on AFHQv2, the FID improves from 3.16 with the EDM baseline to 3.02 with SDM adaptive solver selection, and further to 2.50 when adaptive scheduling is additionally integrated. This indicates that our proposed scheduling method provides particularly additional benefits when the sampling budget is more constrained.

MethodFID
EDM3.16
SDM Solver3.02
SDM Solver + SDM Scheduling2.50

Comparison with AdaFlow (Hu et al., 2024) (Q1, Q2)

AdaFlow additionally trains an auxiliary neural network to estimate the conditional variance of the training loss, which is then used to adjust the step size online during inference based on the current sampling state. In contrast, SDM provides a training-free sampling framework for pretrained diffusion models. Specifically, our scheduling procedure (η\eta-scheduling + NN-step resampling) is performed entirely offline, prior to inference, without requiring an additional variance predictor or online step-size adaptation. To this end, we first derive a Wasserstein-based upper bound on the local discretization error in terms of the velocity variation along the PF-ODE trajectory. We then use this bound to construct an offline-calibrated timestep schedule with an explicit local error budget, and apply NN-step resampling to accommodate a prescribed sampling budget.

Although both methods employ Wasserstein-based error analysis to guide adaptive sampling, they address different settings. AdaFlow is designed for state-conditioned flow-based policies in imitation learning, whereas SDM targets pretrained diffusion models for image generation. In particular, their different application domains, model settings, and evaluation protocols make a direct quantitative comparison inappropriate. We therefore focus on clarifying the methodological distinction between the two approaches.

Comparison with EDM baselines for Flow-based T2I Models (W3)

We would like to clarify why we did not include the EDM schedule in our original comparison in Table 4 (Table 2, before revision). Stable Diffusion 3 and FLUX are rectified-flow-based models, for which timestep schedules are commonly used as the default sampling configurations. In contrast, EDM schedule is defined in terms of the noise level σ\sigma and is not commonly used as a default schedule for these models.

This is also reflected in the resulting schedule distributions. The default schedules of rectified-flow based text-to-image models (e.g., Stable Diffusion 3 and FLUX) generally allocate denser steps to the middle of the sampling trajectory, with relatively larger intervals toward both ends. Our proposed schedules exhibit a similar overall tendency, whereas the EDM schedule produces monotonically increasing intervals as sampling progresses. This indicates that the EDM schedule induces a substantially different allocation of sampling steps compared with native flow-based schedules.

Nevertheless, to address the reviewer's concern, we conduct additional experiments using EDM schedule converted from its noise-level parameterization to the corresponding flow-time parameterization. The results are reported in the table below, where SDM achieves the best FID on both Stable Diffusion 3 and FLUX, outperforming both baselines (Uniform, EDM). These additional experiments demonstrate that SDM remains effective when compared directly with EDM on flow-based image generation models.

Stable Diffusion 3

ScheduleFID
Uniform23.41
EDM23.08
SDM22.06

FLUX

ScheduleFID
Uniform24.41
EDM24.95
SDM23.87

Criterion for Adaptive Solver Transition (W1)

We clarify that the solver transition is determined by the estimated relative local curvature of the sampling trajectory, rather than a predefined switching timestep. Specifically, we introduce a threshold τκ\tau_{\kappa} on the cache-based curvature estimate κ^rel(ti)\widehat{\kappa}_{\mathrm{rel}}(t_i), which measures the relative variation of the vector field between consecutive evaluation points. When the estimate exceeds the threshold, SDM switches from a low-order solver to a high-order solver (e.g., Euler →\rightarrow Heun or DPM-Solver++(1M) →\rightarrow DPM-Solver++(2M)). Importantly, it should be noted that we use the cache-based curvature estimate to introduce no additional NFE during inference.

Clarification of DθD_{\theta} and DσD_{\sigma} (Q3)

Here, Dθ∈RdD_{\theta} \in\mathbb{R}^d denotes the denoiser network parameterized by θ\theta. Specifically, JD:=∂Dθ(x;σ)/∂x∈Rd×dJ_D := \partial D_\theta(x;\sigma)/\partial x \in\mathbb{R}^{d\times d} denotes the Jacobian matrix with respect to the input xx, while Dσ:=∂Dθ(x;σ)/∂σ∈RdD_\sigma := \partial D_\theta(x;\sigma)/\partial \sigma \in\mathbb{R}^d denotes its partial derivative with respect to the noise level σ\sigma, holding xx fixed. For notational clarity and to distinguish the derivative from the denoiser itself, we explicitly clarify these definitions and use JσJ_\sigma instead of DσD_\sigma in the revised manuscript.

Dear Reviewer 5CFy,

Thank you again for your constructive review. We would appreciate your feedback on whether our responses have addressed your concerns.

We clarified the cache-based criterion for adaptive solver switching and the individual contributions of the two SDM components in Table 1. We also added comparisons with EDM schedules for flow-based text-to-image generation (Stable Diffusion 3 and FLUX), where SDM remains achieving the best FID. Finally, we clarified the methodological differences between SDM and AdaFlow and revised the mathematical notation in Proposition 3.1.

We would be happy to address any remaining questions. Thank you for your time and consideration.

Reviewer_1jt4OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SDM, a training-free framework for improving the sampling efficiency of diffusion-based generative models through adaptive solver selection and timestep scheduling. The method consists of two main components. First, it analyzes the second-order behavior of the probability flow ODE and introduces a curvature-based criterion for dynamically selecting numerical solvers. Second, it derives a local Wasserstein error bound to guide adaptive timestep allocation, followed by a resampling procedure to satisfy a fixed NFE budget. The authors evaluate SDM across several benchmarks, including CIFAR-10, FFHQ, AFHQv2, ImageNet, and modern text-to-image models. The results demonstrate improvements in sampling quality and computational efficiency under several configurations.

Critical strengths
  1. Practical motivation and a unified framework. The paper addresses the important problem of improving diffusion sampling efficiency without additional model training. Jointly considering solver selection and timestep scheduling provides a useful perspective on sampling optimization.

  2. Theoretical motivation for adaptive sampling. The analysis of trajectory derivatives and the Wasserstein-based formulation provide interpretable motivations for allocating computational resources differently across sampling stages. The effort to connect numerical analysis with practical sampling decisions is valuable.

  3. Broad experimental evaluation. The experiments cover multiple datasets, diffusion parameterizations, numerical solvers, and modern text-to-image models. Several configurations achieve favorable quality-efficiency trade-offs, demonstrating the practical potential of the proposed framework.

Critical weaknesses

W1:Practical Validity of Wasserstein Error Control. The proposed scheduling algorithm relies on an estimated velocity variation rather than the quantity required by Theorem 3.2. Moreover, the subsequent N-step resampling changes the calibrated step sizes. It remains unclear to what extent the theoretical Wasserstein error guarantee applies to the final sampling procedure.

W2:Limited Evidence for the Complementarity of the Two Components. The joint use of adaptive solver selection and scheduling does not consistently improve the quality-efficiency trade-off. In particular, Table 5 shows that adding adaptive scheduling to the adaptive solver on ImageNet increases FID from 1.45 to 2.25 and NFE from 476 to 503. Similar, though smaller, limitations appear in Table 1. These results raise concerns about the effectiveness of combining the two proposed components.

Questions for authors

Q1: Could the authors clarify how well the estimated Wasserstein error reflects the actual discretization error, particularly after N-step resampling? Is there any empirical evidence supporting the reliability of this approximation?

Q2: In Table 5, why does combining adaptive solver selection with adaptive scheduling substantially degrade performance on ImageNet compared with using adaptive solver selection alone? Could the authors provide further analysis of this behavior?

4 comments

We thank the reviewer for the constructive feedback. We address the reviewer’s concerns below and provide corresponding revisions in the revised manuscript. All revisions are highlighted in blue in the revised PDF, and those addressing the reviewer’s comments are marked with 1jt4 in orange. Please note that the table numbers have changed in the revised manuscript (e.g., Table 8 →\rightarrow Table 2, Table 9 →\rightarrow Table 3). We refer to both the original and revised table numbers in our responses for additional clarity.

Wasserstein Error Control after N-step resampling (W1, Q1)

We provide an explicit Wasserstein bound after NN-step resampling in Appendix D.5. The resulting bound is of the same form as the original Wasserstein-based error bound, up to a factor controlled by the local variation of ϕ\phi. This establishes that the theoretical error analysis remains applicable after NN-step resampling.

Furthermore, as the reviewer mentioned, in practice, our η\eta-scheduling method estimates the supremum of the velocity variation using a single trial step. Figure 2 (right) illustrates the resulting Wasserstein error bounds for the EDM and SDM schedules after NN-step resampling. In particular, SDM effectively reallocates more of the error budget to early sampling stages, where the dynamics are easier to approximate, while maintaining the prescribed error control across the sampling trajectory.

To further assess the practical effectiveness of this estimation and scheduling procedure, we measure the actual local discretization errors, for CIFAR-10, using a high-accuracy reference integration. Specifically, we divide each sampling interval [σi,σi+1][\sigma_i,\sigma_{i+1}] into 64 sub-intervals and integrate the probability flow ODE over them to obtain a reference solution. Rather than using a single reference trajectory over the entire sampling process, we independently initialize the reference integration at the corresponding state xix_i of the original 18-step Euler trajectory for each interval. This allows us to measure the local discretization error within each interval, without including errors accumulated in preceding intervals.

We first examine how well SiS_i captures the actual local discretization errors by computing the Spearman rank correlation within each sampling interval. Specifically, for CIFAR-10, we compute the correlation across 128 noise samples for each of the time intervals and report the average across intervals. SDM achieves a higher mean within-interval Spearman correlation than EDM (0.937 vs. 0.841), indicating that SiS_i more consistently captures the relative variations in actual local discretization errors.

We further evaluate the practical effectiveness of the resulting timestep allocation by measuring the sum of local discretization errors under the same 18-step Euler sampling budget. SDM achieves a lower mean sum of local discretization errors than EDM (6.62 vs. 6.96), averaged over noise samples, with a standard deviation of 1.65 for both methods.

ScheduleMean ∑iei\sum_i e_iStd.Spearman correlation (Avg.)
EDM6.961.650.841
SDM6.621.650.937

This result provides further empirical evidence that our SiS_i-based estimation captures practically relevant variations in local discretization errors and enables more effective timestep allocation. Together with our theoretical error analysis and FID improvements, these results support the reliability and practical effectiveness of our estimation and scheduling procedure. While evaluating additional substeps within each interval for estimating the velocity variation could improve the estimation accuracy, this would introduce additional NFE overhead during scheduling. We therefore adopt the single-step estimate as a practical trade-off between estimation accuracy and computational cost.

Limitations in Stochastic Sampling for ImageNet-64 (W2, Q2)

As the reviewer pointed out, the performance improvement is reduced for ImageNet. We note that ImageNet is the only experimental setting in which stochastic churn is introduced. Since our proposed scheduling method is derived under deterministic PF-ODE sampling, the stochastic perturbations introduced by churn may affect the reliability of the estimated local velocity variation, potentially limiting the effectiveness of adaptive scheduling in this setting. However, we believe that extending our scheduling framework to stochastic sampling dynamics will be an interesting direction for future work.

Meanwhile, to address the reviewer's concern, we conduct additional experiments without stochastic churn, comparing SDM with the corresponding EDM baselines. We reduce the number of sampling steps from 256 to 64 to better assess the effectiveness of our proposed method. As shown in the table below, SDM outperforms EDM baselines under deterministic sampling settings. We have further updated our quantitative results in Table 7 (Table 5, before revision).

SolverScheduleFIDNFE
EulerEDM2.8664
EulerSDM Schedule2.6364
HeunEDM2.26127
HeunSDM Schedule2.25127
SDM SolverEDM2.26121
SDM SolverSDM Schedule2.24121

Thank you for the detailed responses and additional experiments. I appreciate the authors' efforts to address my concerns.

The new analysis of local discretization errors provides useful empirical support for the proposed Wasserstein-based error estimation. In addition, the revised ImageNet experiments under deterministic sampling substantially alleviate my concern about the performance degradation when combining the two components.

I have one minor clarification regarding the revised manuscript. In Section 4.1, Lines 379–381 state that stochastic settings have been removed for ImageNet, while Lines 386–390 still mention using stochastic churn during inference. These descriptions appear inconsistent, and I suggest clarifying the actual experimental configuration in the final version.

Overall, the rebuttal has addressed my main concerns to a satisfactory extent. Although some limitations remain, I find the additional evidence convincing enough to raise my rating from 2 to 3.

We thank the reviewer for the positive assessment and for raising this final clarification. We have corrected Lines 386–390 to clarify that stochastic churn is disabled for ImageNet experiments, ensuring consistency with the deterministic sampling configuration described in Lines 379–381. We appreciate the reviewer's careful reading and constructive feedback.

Reviewer_9VmPOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates the influence of the numerical solver and timestep schedule on the quality of generated samples. The authors analyze the ODE dynamics and observe that the velocity variation is small during the high-noise stages and increases near the data manifold. Motivated by this observation, they propose the Sampling Design Space for Diffusion Models (SDM), a framework that adopts an adaptive numerical solver and schedule. They further introduce an adaptive scheduling method that selects the step size according to the local Wasserstein discretization error. Numerical experiments are conducted to evaluate the performance of the proposed method.

Critical strengths
  1. The paper proposes a dynamic allocation strategy that deploys efficient first-order solvers in the early stages and transitions to stable higher-order solvers only when geometrically required.
  2. The paper derives an adaptive schedule that explicitly bounds the local Wasserstein distance between the discretized trajectory and the true continuous flow.
  3. Experiments on various datasets evaluate the performance of the proposed method.
Critical weaknesses
  1. Proposition 3.1 presents the second derivative of the probability flow ODE under the EDM/VP/VE parameterizations. The conclusion is restricted to these parameterizations, which is reasonable because the ODE dynamics depend on the noise schedule and time parameterization. However, it remains unclear whether the conclusion in Proposition 3.1 holds for arbitrary parameterizations. If not, the result is not an intrinsic property of the ODE. Why, then, does it motivate the authors to design a solver transition rather than to design a more suitable parameterization?

  2. Condition (12) in Theorem 3.2 is only a sufficient condition. Could the authors establish that such a condition involving SiS_i is also necessary, and demonstrate that SiS_i is a key quantity governing sample quality?

  3. In Eq. (15), the design of η\eta remains heuristic, making it difficult to judge whether the current schedule approximates an optimal one. It appears that this section merely provides another schedule with potentially better performance, and the contribution is therefore not significant enough. It would be stronger if the authors could show that the proposed schedule is optimal or near-optimal in some sense.

  4. According to the results on Stable Diffusion in Table 1, the improvement of SDM appears to be mild.

  5. According to Table 6, SDM yields only mild improvement compared with modern ODE solvers.

Questions for authors
  1. In the experiments, the authors choose different NFE values for different datasets. This is reasonable, but how are these values determined?
  2. This work is restricted to ODEs. Could it be extended to SDEs? This could be left as future work.
6 comments

We thank the reviewer for the constructive feedback. We address the reviewer’s concerns below and provide corresponding revisions in the revised manuscript. All revisions are highlighted in blue in the revised PDF, and those addressing the reviewer’s comments are marked with 9VmP in green. Please note that the table numbers have changed in the revised manuscript (e.g., Table 8 →\rightarrow Table 2, Table 9 →\rightarrow Table 3). We refer to both the original and revised table numbers in our responses for additional clarity.

Different NFE values across Datasets (Q1)

The number of discretization steps follows the original EDM configurations (18 steps for CIFAR-10, 40 steps for FFHQ and AFHQv2). We adopt these settings to ensure a consistent comparison with the original EDM baselines. The NFE is determined by the actual number of model evaluations required by each numerical solver. Specifically, Euler requires one evaluation per integration interval, whereas Heun requires an additional evaluation for its predictor-corrector update.

Furthermore, we conduct additional experiments under reduced sampling budgets, as reported in Table 2 and 9 (Table 8 and 7, before revision). These results demonstrate that SDM achieves substantial improvements over EDM baselines in few-step generation settings (e.g., 5.44 →\rightarrow 3.21 for FFHQ and 3.16 →\rightarrow 2.50 for AFHQv2 for 10-step generation).

Motivation of Solver Transition (W1)

We would like to clarify that our analysis is not fundamentally restricted to EDM/VP/VE parameterizations. In Appendix B, we derive a general expression for the second derivative of the PF-ODE in terms of arbitrary differentiable scaling and noise schedules s(t)s(t) and σ(t)\sigma(t) (Eq. 36), while Proposition 3.1 presents its simplified forms under the commonly adopted EDM/VP/VE parameterizations.

Regarding the choice of solver transition, our primary objective is to improve the sampling efficiency of existing pretrained diffusion models without modifying their architectures or training procedures. Although exploring alternative parameterizations is an interesting direction, we consider it a complementary research direction. Our work instead focuses on the two key components of sampling design of diffusion-based generative models, ODE solvers and timestep scheduling.

To further elaborate how trajectory curvature motivates adaptive solver allocation, the local transition error of the Euler solver, for instance, can be estimated as ηiE=Δti22S^(i),\eta_i^{\mathrm{E}} = \frac{\Delta t_i^2}{2}\hat{S}(i), where Δti\Delta t_i denotes the timestep size and S^(i)\hat{S}(i) characterizes the local trajectory curvature, illustrating that the local discretization error depends jointly on the timestep size and the local trajectory curvature. While timestep scheduling can reduce the error by allocating smaller steps to regions with higher curvature, it does not change the underlying trajectory dynamics. Consequently, under a fixed NFE budget, relying solely on finer discretization can become inefficient in regions with high curvature. This motivates our adaptive solver transition, which allocates higher-order solvers to later sampling stages where the trajectory exhibits higher curvature, thereby controlling local discretization errors more efficiently. This design is theoretically motivated by Proposition 3.1.

Potential Extension to SDE sampling (Q2)

In this work, as the reviewer mentioned, we focus on deterministic PF-ODE sampling, which provides an efficient sampling framework with deterministic trajectories while preserving the marginal distributions of the corresponding diffusion process.

We acknowledge that extending SDM to stochastic sampling is an interesting direction. However, additional considerations are required when extending our framework to SDEs. In particular, the stochastic perturbations introduced during sampling affect the estimation of local velocity variations and the corresponding Wasserstein error bounds. This is especially relevant to our adaptive line-search procedure in Appendix E.1, which currently relies on deterministic trajectory evaluations.

We believe that extending the geometric interpretation and Wasserstein-bounded scheduling framework to stochastic trajectories would be a valuable direction for future work.

Sufficiency of Condition (12) and the Role of SiS_i (W2)

We would like to clarify that necessity is not required for our error-control objective. Our goal is to construct a theoretically grounded and tractable criterion that guarantees the local Wasserstein discretization error remains below a prescribed tolerance, rather than to characterize all possible conditions under which this error bound holds. Specifically, Condition (12) is derived by upper-bounding the local discretization error using the supremum of the velocity variation along the true trajectory. By designing the timestep schedule to satisfy this condition, we obtain the desired local error guarantee. The fact that the condition may be conservative does not undermine this guarantee; it only means that some schedules violating the condition may also achieve errors below the prescribed tolerance.

Furthermore, SiS_i is not an arbitrary heuristic quantity: it directly characterizes the velocity variation along the true trajectory and determines the resulting local Wasserstein error bound. Appendix D.3 further extends this analysis to the global Wasserstein error under Wasserstein-stable numerical transitions, establishing how local errors propagate throughout the sampling trajectory. Thus, the theoretical role of SiS_i is to provide a principled basis for controlling local errors and, through their propagation, the global discretization error—not to serve as a necessary condition for achieving a prescribed error tolerance.

To further examine the practical relevance of this error-based criterion, we measure the actual local discretization error, for CIFAR-10, by comparing Euler integration with a high-accuracy reference solution. Specifically, with a batch size of 128, each sampling interval [σi,σi+1][\sigma_i,\sigma_{i+1}] is divided into 64 sub-intervals for reference integration. Importantly, the reference integration is independently initialized at the corresponding state xix_i of the original 18-step Euler trajectory for each interval, allowing us to isolate the local discretization error within that interval.

We then evaluate how well SiS_i captures the variations in actual local discretization errors by computing the Spearman rank correlation within each sampling interval. As reported in table below,, SDM achieves a higher mean within-interval Spearman correlation than EDM (0.937 vs. 0.841), indicating that SiS_i more consistently reflects the relative magnitude of local discretization errors across sampling intervals.

ScheduleSpearman correlation (Avg.)
EDM0.841
SDM0.937

This result provides empirical evidence that SiS_i-based error allocation is a practically meaningful criterion for reducing local discretization errors, supporting the relevance of SiS_i not only to theoretical error control but also to guiding timestep allocation for improved sample quality.

On the Design of Wasserstein Error based Scheduling (W3)

We would like to clarify that the scheduling design is guided by our Wasserstein-based error analysis, rather than being an arbitrary choice. Specifically, the minimum and maximum tolerance parameters (ηmin(\eta_\mathrm{min} and ηmax)\eta_\mathrm{max}) are calibrated based on the range of Wasserstein errors induced by the baseline schedule, thereby preserving the overall cumulative error, as mentioned in Appendix D.4. The parameter pp then controls how the error budget is redistributed across sampling stages, reallocating the error budget to the early stages of sampling where the dynamics are easier to approximate.

Accordingly, our scheduling procedure can be viewed as two steps: (1) η\eta-scheduling, which reallocates the local error budget based on the baseline error profile, and (2) NN-step resampling, which constructs a schedule with a prescribed number of inference steps while retaining the validity of the Wasserstein-based local error bound.

The resulting error-budget redistribution is illustrated in Figure 2. Furthermore, Appendix D.5 provides a theoretical justification showing that the Wasserstein-based local error interpretation remains valid after NN-step resampling.

Marginal Improvement over Baselines (W4, W5)

From Table 4 and Table 10 (Table 2 and Table 10, before revision), it can be observed that, for text-to-image generation, SDM outperforms baselines in FID for both Stable Diffusion 3 (23.41 →\rightarrow 22.06), FLUX (24.41 →\rightarrow 23.87) and SDXL (18.29 →\rightarrow 17.72), while achieving comparable performance in CLIP. We would like to note that our proposed scheduling method is grounded in Wasserstein distance-based optimization, which naturally targets the image distribution quality rather than semantic text alignment.

Furthermore, integrating with modern ODE samplers, SDM outperforms EDM baselines in FID, with more pronounced gains on CIFAR-10 (2.57 →\rightarrow 2.33) and AFHQv2 (2.12 →\rightarrow 2.07), supporting the geometric interpretation in Proposition 3.1. We acknowledge that the improvement on FFHQ is relatively modest, which can be attributed, in part, to the strong baseline already provided by high-order solvers with a relatively large number of timesteps (N=40)(N=40), leaving less room for further improvement.

Importantly, the benefits become more apparent in few-step generation settings, as demonstrated in Table 9 (N=10N=10, Table 7 before revision), Table 2 (N=20N=20, Table 8 before revision), and Table 3 (N=32N=32, Table 9 before revision), where SDM provides a more effective allocation of a limited sampling budget.

Comments on the paper

2 comments

We thank all the reviewers for their constructive feedback and the positive assessments of our theoretically motivated sampling framework, which combines adaptive solver allocation with timestep scheduling, as well as its broad experimental evaluation.

All revisions are highlighted in blue in the revised PDF and tagged with the respective reviewer IDs (9VmP, 1jt4, 5CFy) for reference. Please note that the table numbers have changed in the revised manuscript (e.g., Table 8 →\rightarrow Table 2, Table 9 →\rightarrow Table 3). We refer to both the original and revised table numbers in our responses for additional clarity.

We summarize the major updates as the following:

1. Removal of stochastic settings for ImageNet-64

  • We remove stochastic churn from the EDM sampling configuration on ImageNet-64 to ensure a consistent comparison under deterministic sampling which aligns with the setting considered by SDM [1jt4] - Section 4.1, Table 7, Appendix F.1.
  • We discuss the extension of SDM to stochastic sampling settings as a direction for future work [9VmP, 1jt4] - Conclusion.

2. Baseline Comparisons

  • We reorganize the quantitative results (few-step generation, modern ODE solvers, and high-resolution generation) to better illustrate the effectiveness of SDM compared with corresponding baselines [9VmP, 5CFy] - Section 4.2, Table 2, Table 3.
  • We add comparisons with EDM baselines for flow-based text-to-image generation models [5CFy] - Section 4.2, Table 4.

3. Clarification of mathematical notation

  • We replace DσD_\sigma with JσJ_\sigma to distinguish the partial derivative with respect to the noise level from the denoiser notation and improve mathematical clarity [5CFy] - Proposition 3.1, Appendix B.

Dear Area Chair,

Thank you for handling our submission. Following the discussion, Reviewers 1jt4 and 9VmP have both raised their ratings from 2 to 3. We briefly summarize the discussion and the key evidence supporting our contribution.

Discussion summary

  • Reviewer 9VmP (2 → 3): We addressed the theoretical concerns regarding the generality of Proposition 3.1, the sufficient nature of Condition (12), and the role of SiS_i in local error control. We further provided empirical validation using high-accuracy reference integrations (Spearman correlation: 0.937 for SDM vs. 0.841 for EDM) and clarified the Wasserstein analysis after NN-step resampling.
  • Reviewer 1jt4 (2 → 3): We have addressed the reviewer’s concerns through additional local-error experiments and ImageNet experiments under deterministic settings.
  • Reviewer 5CFy (2): We added comparisons with EDM schedules for flow-based text-to-image generation (Stable Diffusion 3 and FLUX), where SDM remains achieving the best FID. We further clarified the independent roles of SDM Solver and SDM Schedule, and reported the actual offline preprocessing costs of schedule calibration. The resulting schedule is precomputed and reused during inference, introducing no additional runtime overhead.

Our contribution and supporting evidence

  • Theoretically motivated sampling design. SDM connects trajectory analysis and a local Wasserstein error criterion to training-free solver allocation and timestep scheduling.
  • Broad applicability. We demonstrate SDM's effectiveness with modern ODE solvers, in few-step and high-resolution generation, and across text-to-image generation models, showing its versatility across solvers, sampling budgets, and model families.

We sincerely thank the Area Chair for handling our submission and all reviewers for their time and constructive feedback throughout the discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.