- 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 . 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 in terms of the variance scheduler . Next, each timestep is locally searched such that the local Wasserstein distance does not exceed the given upper bound .
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 , the local Wasserstein distance , and the error upper bound . Based on this inequality, they sequentially perform a local search to determine each optimal time step .
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
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 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
| Schedule | FID |
|---|---|
| Uniform | 23.41 |
| EDM | 23.08 |
| SDM | 22.06 |
FLUX
| Schedule | FID |
|---|---|
| Uniform | 24.41 |
| EDM | 24.95 |
| SDM | 23.87 |
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 on the cache-based curvature estimate , 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 Heun or DPM-Solver++(1M) 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 and (Q3)
Here, denotes the denoiser network parameterized by . Specifically, denotes the Jacobian matrix with respect to the input , while denotes its partial derivative with respect to the noise level , holding fixed. For notational clarity and to distinguish the derivative from the denoiser itself, we explicitly clarify these definitions and use instead of 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.
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 → Table 2, Table 9 → 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.
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.
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-scheduling + N-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 N-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.