acceptodds
Under review as a conference paper at ICLR 2027

When Do Extra Evaluations per Step Pay Off in Few-Step Sampling?

Abstract

Few-step samplers for diffusion and flow models spend a fixed budget of network function evaluations (NFE). A multi-stage solver such as Heun spends some of them inside a step to correct it; a first-order solver spends each on a new step. We ask when an extra evaluation is worth more inside a step. Measured on the forward path of a reference sample, the error of one step is a term shared by all solvers plus a correction from the extra evaluations; for Euler the shared term is the step's share of the remaining time times the model's error in predicting the clean sample. An Euler step is a convex combination of the current state and this prediction, so it is non-expansive whenever the prediction is. The correction grows with the ratio of the step length to the distance of its end to the clean image, where the sampling equation becomes stiff. Extra evaluations should therefore pay off only on short steps away from the image, and their cost inside long steps should shrink as the budget grows. A solver menu graph records the one-step error of eight solvers on every interval of a fine grid and plans the timestep and solver of every step for an exact budget by dynamic programming. On three undistilled models, one Heun step never beats two Euler steps once . On GenEval at 10 NFE, re-placing Euler steps raises the score by up to 0.107, while solvers that spend evaluations inside a step lose 0.031 on SDXL and 0.090 on Qwen-Image on average; at 15 NFE the loss falls to 0.010 and 0.028. On FLUX.1-dev and SD3.5-Large, five Heun steps score 0.05 below ten Euler steps, both placed by the native scheduler. An Euler schedule calibrated on a few prompts compares favourably with Align Your Steps and GITS on SDXL at 10 NFE, and paths planned on Qwen-Image carry over to Qwen-Image-Edit, where they bring edits closer to the 50-step result. On the distilled Z-Image-Turbo, whose graph has almost no window, two planned Euler steps score as high as its default eight. In few-step sampling, evaluations should thus go first to placing first-order steps, and into a step only where the step is short relative to its distance from the clean image.

Primary area:
generative models
Keywords:
Fast generative sampling, solver-menu graph, NFE budget, Few-Step Sampling

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_tdPVOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The goal of the paper is to accelerate sampling of diffusion and flow models. Their approach does not change the model weights but rather optimize the sampler of the model.

A planned sampler for a fixed budget of NFE is a time schedule t0,t1,…,tNt_0, t_1,\ldots, t_N and a solver schedule s0,s1,…sN−1s_0, s_1, \ldots s_{N-1}, where sis_i is a step of an existing numerical solver e.g., Euler, Heun, DPM2++, and more, such that the sum of the NFE of the total number of steps steps is equal to the NFE budget. In practice their approach to accelerate the sampling is via a dynamic programing algorithm that plan solvers which they propose.

The main contribution of the paper as I understand them are:

  1. Given a noise ϵ\epsilon which the high accuracy solution maps to a sample x0x_0 (i.e., image), they propose new way to measure local error. Instead of measuring error w.r.t. to the solution the ground truth solution of the ODE, they measure is w.r.t. to the interpolant of ϵ\epsilon and x0x_0: xt∗=(1−t)x0+tϵ.x_t^* = (1-t)x_0 + t\epsilon. Then they provide a bound of the global error using the proposed new local error. This bound is minimized to optimize the planned sampler.

  2. They provide empirical analysis of the local error along the trajectory.

  3. They provide empirical validation of their algorithm on SDXL, Qwen-Image, Qwen-Image-Edit, and Z-Image-Turbo. Notably, the reported results actually shows that the best performance can always be achieved via constate solver schedule s0,s1,…sN−1s_0, s_1, \ldots s_{N-1} (i.e., single solver type).

Critical strengths

Accelerating samplers of diffusion and flow model is a relevant critical problem for many application. And since the approach they propose essentially squeeze the efficiency of the model using existing solvers, without the need for data or to change the model's weights. Hence it could potentially be useful to many users.

Additionally, they provide a principled derivation of the global error bound they optimize, and they provide nice empirical analysis of local error of different solvers on multiple models.

Critical weaknesses

The two main weaknesses which I identify in the paper are significant. First, the proposed local error w.r.t. to the interpolant path is not motivated well and the authors do not discuss any geometrical interpretation of such definition of error. Second, is weak empirical results. In practice the reported results shows that the best performance can always be achieved via a single solver, and the method mostly optimize the time schedule.

Error w.r.t. to the interpolant: I am concerned that the definition of the error is not well justified in the paper. I might be missing something in my understanding, but I will present my current understanding.

Obviously the interpolant xt∗=(1−t)x0+tϵx_t^* = (1-t)x_0 + t\epsilon is not a solution to the ODE. Hence, the fact that the author defines the error w.r.t. to this path without any motivation, geometrical discussion, or what is the meaning of this error in limit of high accuracy solver (clearly it does not approach zero) makes the paper hard to interpret. Note, that even the connection to global truncation error via the bound does not help so much as there is no discussion about how tight the bound is. Even the section "DOES THE LOWEST PATH COST PICK A GOOD PATH?" seems to compare ImageReward without any mention the global error as defined in Proposition 5. Hence, its not clear that minimizing the bound in Proposition 5 actually reduce the global error.

Additionally, since the reference interpolant path is already defined via the map between noise ϵ\epsilon to image x0x_0 of the model, it is already required to solve the ODE with high accuracy (50 steps in the paper). So I would expect the authors to motivate their choice of reference path over defining the reference path to be the peace-wise linear interpolation between every point on the trajectory of the high accuracy solver (This is instead of the linear interpolation between the endpoints as currently in the paper). The latter, appears to be the naive approach to approximate the exact solution of the ODE and it does not require more compute.

Lastly, it seems to me that minimization of the bound proposed by the authors essentially tries to make the trajectories straight. As implied by Proposition 5, If succeeded, then indeed global error should be minimized as well. However, not every map between noise ϵ\epsilon to image x0x_0 allows such a flow. This is discussed in rectified flows[1], they also straighten the trajectory by making the model follow the straight paths connecting ϵ\epsilon and x0x_0. Hence, I think it might be appropriate add a discussion about the connection to rectified flows[1].

Weak empirical results: Tables 1 and 2 shows that the mixed solver schedule is not needed, the proposed method mostly optimize the time schedule. First, it seems that proposed algorithm for finding the best solver schedule is an over kill, and simple grid search over three solvers is enough. Second, it would be appropriate as well as useful to the reader to compare with the popular "time shift"[2] reparametrization on a grid of shift parameter values. It is safe to say that this time transformation is already standard and it is used as the default time schedule for many open source models (including Qwen-Image).

[1] Liu, et al. "Flow straight and fast: Learning to generate and transfer data with rectified flow." [2] Esser, et al. "Scaling rectified flow transformers for high-resolution image synthesis."

Questions for authors

Can the authors please address my concerns regarding the definition of the error w.r.t. to the interpolant? Can the authors motivate why this definition is superior to error w.r.t. the ground truth solution of the ODE which can be approximated by peace-wise linear interpolation between the points on the trajectory of the high accuracy solver?

Can the authors provide empirical evidence that minimizing the proposed global bounded indeed reduce global error?

Perhaps, consider to provide empirical evidence that the proposed bound is better than the naive approach (as I described it)?

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

The paper asks a simple practical question: with a fixed budget of network evaluations (NFE), is an extra evaluation better spent on a correction inside a step (Heun, RK, DPM-Solver-2S) or on one more Euler step? The authors split the one-step error into a part shared by all solvers and a solver-specific correction, and argue that the correction only helps when the step is short compared with its distance to the clean image (small h/kh/k). They then measure one-step errors of eight solvers on a fine time grid and use dynamic programming to choose both the timesteps and the solver for every step under an exact budget. The main message is that at around 10 NFE well-placed Euler steps are usually at least as good as higher-order solvers, and that most of the gain over the native sampler comes from where the steps are placed.

Critical strengths
    1. The question is clear and the answer is simple and easy to interpret. The h/kh/k criterion is checked on several current large models, and Appendix M also looks at the true sampling trajectory of FLUX.1-dev and SD3.5-Large. I liked the result on the distilled Z-Image-Turbo: the graph itself says to use only Euler steps, and two planned Euler steps match the default eight.
    1. There is a useful practical by-product. The Euler layer of the graph needs only the residuals along a reference run, so it is a cheap schedule optimizer. Averaged over 16 prompts it is on par with Align Your Steps [1] and GITS [2] on SDXL at 10 NFE, and schedules planned on Qwen-Image carry over to Qwen-Image-Edit without recalibration.
Critical weaknesses
    1. I am not convinced the error metric is neutral between solvers. Errors are measured against the straight line (1−t)x0+tε(1-t)x_0 + t\varepsilon between the noise and the 50-step sample, not against the ODE trajectory. An Euler step moves straight towards the model's prediction x^0\hat x_0 and is rewarded for that, while even an exact ODE integrator started at xt∗x^{\ast}_t would get a nonzero error under this metric. Appendix M partly answers this: on the true trajectory the advantage of Heun also shrinks as h/kh/k grows. The threshold is not the same, though. On SD3.5-Large, Heun still beats two Euler steps on 16% of the intervals beyond h/k=10h/k=10, whereas on the reference path it never wins beyond h/k≈3−5h/k \approx 3-5. This check is also done on FLUX.1-dev and SD3.5-Large rather than on the models where the window was measured.
    1. The image-quality results support the main message less clearly than the abstract suggests. The large 10-NFE loss of Heun comes from a plan that has to take a Heun step into t=0t=0, which standard samplers avoid (EDM~[4] takes its last step with Euler). Once the last step is Euler, at 15 NFE, Heun is the best sampler on SDXL (GenEval 0.523, ImageReward 0.536) and ties for first on Qwen-Image (0.850). On SDXL at 10 NFE the best sampler is DPM-2S (0.504). The full planner adds little and sometimes hurts: on Qwen-Image at 10 NFE the lowest-cost path (Mix) scores 0.792, compared with 0.830 for the Euler-only member. Many of the reported differences are 0.01-0.03 GenEval with no confidence intervals, while the noise of a single score over 553×4553 \times 4 images is itself around 0.01. Examples are the placement gains on FLUX.1-dev and SD3.5-Large (0.018 and 0.007) and the comparison with Align Your Steps and GITS (0.519 vs. 0.508 and 0.510; with one calibration prompt the planned schedule gets 0.497, below both). There is no FID and no human evaluation, since ImageReward is only a learned proxy. Finally, the general observation that high-order solvers struggle at small budgets and strong guidance is not new (DPM-Solver++ [3], EDM [4]), and the closest method, S3 [5], as well as DPM-Solver-v3 [6], AMED [7] (and other recent methods), are not compared experimentally.

References:

    1. A.Sabour, S.Fidler, K.Kreis. Align Your Steps: Optimizing Sampling Schedules in Diffusion Models. // International Conference on Machine Learning (ICML), 2024.
    1. D.Chen, Z.Zhou, C.Wang, C.Shen, S.Lyu. On the Trajectory Regularity of ODE-based Diffusion Sampling. // International Conference on Machine Learning (ICML), 2024.
    1. C.Lu, Y.Zhou, F.Bao, J.Chen, C.Li, J.Zhu. DPM-Solver++: Fast Solver for Guided Sampling of Diffusion Probabilistic Models. // Machine Intelligence Research, 22(4):730--751, 2025.
    1. T.Karras, M.Aittala, T.Aila, S.Laine. Elucidating the Design Space of Diffusion-Based Generative Models. // Advances in Neural Information Processing Systems (NeurIPS), 2022.
    1. E.Liu, X.Ning, H.Yang, Y.Wang. A Unified Sampling Framework for Solver Searching of Diffusion Probabilistic Models. // International Conference on Learning Representations (ICLR), 2024.
    1. K.Zheng, C.Lu, J.Chen, J.Zhu. DPM-Solver-v3: Improved Diffusion ODE Solver with Empirical Model Statistics. // Advances in Neural Information Processing Systems (NeurIPS), 2023.
    1. Z.Zhou, D.Chen, C.Wang, C.Chen. Fast ODE-based Sampling for Diffusion Models in Around 5 Steps. // IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024
Questions for authors
  • Could you repeat the window measurement on Qwen-Image and SDXL with the error measured against the exact ODE solution from xt∗x^{\ast}_t (fine-step integration)? Does the h/k≈3−5h/k \approx 3-5 threshold survive?
  • For the five Heun steps on FLUX.1-dev and SD3.5-Large, is the final step into t=0t=0 a Heun step or an Euler step? How do the 10-NFE results for Heun change with an EDM-style Euler final step?
  • What about the confidence intervals or paired tests for Tables1, 2 and 15?
  • Why does Mix lose to the Euler-only member on Qwen-Image at 10 NFE, and what does this say about Hypothesis H, given the weak correlation between path cost and ImageReward on this model (−0.25-0.25 to −0.34-0.34, compared with about −0.8-0.8 on SDXL)?
  • How does the window depend on the classifier-free guidance scale?
2 comments
AuthorsOpenReview ↗

To Reviewer uqDx

Dear Reviewer uqDx,

As the discussion period is closing, we would like to check whether our response has addressed your questions. The revised PDF adds the window against a high-accuracy ODE solution, paired 95% intervals, the Euler-ending and guidance experiments, and blind preference tests (Appendix R, changes in blue). The revised abstract and introduction also state the question the paper answers, and Section 5.3 states that every row of Table 1 is planned by the graph, each solver on its own schedule at equal NFE, which bears on your Q4 and W2(iii). We would be glad to clarify anything that remains open. Thank you for your time.

Reviewer_uqDxOpenReview ↗

Reviewer response

Dear authors, thank you for your detailed answer. I'm satisfied with additional comments and experiments presented in your answers. I hope that the comparison against a high-accuracy ODE references will be added to the next revision of this paper. I'll keep my quite positive score, and my remaining concern relates to the absence of the code related to this paper. I strongly recommend to authors to prodive it in the nearest future.

Reviewer_46sBOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Few-step sampling of diffusion or flow models needs to work under a NFE budget. While previous work mainly sought better strategies among whole denoising trajectories from individual solvers, this paper studies the local behavior of each solver. In particular, higher-order solvers like Heun require more than one NFE for one step, so using more Euler steps can be more helpful than one Heun step. The authors study the behavior of higher-order solvers and quantify when does the extra evaluations can help improve sample quality. A solver menu graph identifies the denoising process as a path in a directed acyclic graph and allows mix-matching different solvers along the trajectory.

Experiments show that Heun steps are only useful when the step size is small and the current timestep is close to noise. This is shown among different model types (diffusion and flow) and different tasks (image generation and image editing). Search on the solver menu graph gives planned denoising paths that are better than most random ones. Therefore this paper demonstrates a new method of developing denoising paths that adequately places NFEs to where it is most useful under a fixed NFE budget.

Critical strengths
  • S1: The proposed framework allows local comparison among solvers.

Previous work focused on full trajectories produced by diffusion model solvers. The authors compare a menu of solvers and propose a method to mix-match different solvers in the denoising process and quantify where additional NFEs reduces error the most. This helps practitioners to understand the behavior of each solver locally at timesteps and decide what solver graph to use.

Critical weaknesses
  • W1: The paper contains soundness issues that makes understanding hard.

For example, how does an Euler step "does not amplify an error inherited from earlier steps" when it starts from xtx_t produced from earlier steps (line 90)? In proposition 3, shouldn't CC include a term for the time partial derivative ∂tvθ\partial_t v_\theta from the chain rule? Heun is a 2nd order solver, so O(h2)\mathcal{O}(h^2) terms are not negligible. These kind of issues makes following the paper at first look difficult.

  • W2: The reliablity of the reference path.

The authors set the reference path xt∗x_t^* as the native pipeline's samples at its find schedule and treat it as golden. However, the native solver at its fine schedule still has its own error compared to the theoretical ideal trajectory. Even when a higher-order solver has O(h3)\mathcal{O}(h^3) error, comparing with a O(h2)\mathcal{O}(h^2) reference should make the higher-order solver have O(h2)\mathcal{O}(h^2) error. Therefore, the proposed method fails to measure the error of each solver in a mathematically justified way.

Questions for authors
  • Q1: Realistic value of h/kh/k.

The authors observe that a Heun step beats two Euler steps only when h/k<5h/k < 5. But this means h<5kh < 5k, where the current step size hh is smaller than 5 times the sum of all remaining timesteps. Popular checkpoints use uniform or shifted timestep schedules, where then timestep size remains the same or increase along the denoising process. Therefore I think that the condition is almost always satisfied. If h/kh/k is small almost always, when does a Heun step beat two Euler steps? Is there a better discriminating condition?

4 comments
AuthorsOpenReview ↗

To Reviewer 46sB

Dear Reviewer 46sB,

As the discussion period is closing, we would like to check whether our response has addressed your concerns. In brief, the window keeps its threshold against a high-accuracy ODE solution (Appendix R.2), which answers W2 on the reference; the path cost bounds the true global error, and the planned schedule has the lowest global error against the native and time-shifted schedules (Appendix R.3, R.7), which answers W1. A closely related question on the reference and the global error was also discussed with Reviewer tdPV in the thread above. We would be glad to clarify anything that remains open. Thank you for your time.

Sorry for the late reply, and thank you for the detailed response and the additional experiments. Below I note what is resolved and what remains open.

  • Resolved.

W1b: the expansion around $ (x^*_k, k) $ absorbs the time change into $ r_k - r_t $, so no separate $ \partial _t v _\theta $ term is needed.

W2: Appendix R.2 measures one Heun step against two Euler steps relative to a high-accuracy solution from $ x^*_t $, and the threshold holds.

  • W1a remains open.

I agree that an Euler step is a convex combination of the current state and the model's prediction. However, the prediction itself can amplify an error, so the step is non-amplifying only when the prediction is.

Prop. 2 assumes ∥∂xx^0∥≤1\Vert\partial_x \hat{x}_0\Vert \le 1, and I think this assumption is unrealistic. Under CFG, the guided prediction extrapolates between the conditional and unconditional predictions, so an eigenvalue of its Jacobian can exceed 1, and by the perturbation formula in Appendix I, perturbations along such directions grow along the flow.

The edge weights are also measured only from the reference states. In sampling, each step starts from a state that has already drifted, and at 5–10 NFE this drift is not small (Table 20). The response to uqDx (Q4) gives an example: the Heun edges of Mix are accurate from the reference states but hurt quality in sampling.

If later steps amplify errors, the Bellman recursion should weight each edge by how much the remaining steps amplify it. This weight depends on the number and type of the remaining steps, so it is not a common factor, contrary to the general response, and it can change which path is selected. I think the paper needs to estimate this amplification along the planned paths, or include it in the recursion.

  • W2 (the reference sample).

The response says x0x_0 is not required to be exact, only reachable from ε\varepsilon. Then the bound, the planner and the global errors are all measured against the native sample, and the method finds a few-step path that reproduces that sample. This is a useful goal, but it is weaker than sampling from the model's distribution. The paper should state its claims accordingly.

  • Q1 (the window).

The response clarifies that h/k<5h/k < 5 is a necessary condition, not the criterion. But nearly every step of a standard schedule satisfies it, so it does not tell a practitioner where an extra evaluation pays off. The useful information is inside the window, where the gain varies widely across models and guidance scales. I think the paper should make the properties of this window its main result, rather than lead with h/kh/k.

  • Presentation.

I still find the paper hard to follow. Many paragraphs introduce several definitions at once, and some propositions are first-order heuristics (for example Propositions 3 and 4). The response adds ten appendix sections and plans to move them into the main text. I believe this restructuring goes beyond a camera-ready revision.

I thank the authors again for the detailed answers to my questions, but my score for the paper remains unchanged.

AuthorsOpenReview ↗

To Reviewer 46sB

Dear Reviewer 46sB,

Thank you for the follow-up and for confirming that W1b and the high-accuracy ODE check in W2 are resolved. Addressing your remaining concern, we clarify a central distinction: answering the title's question—when extra evaluations per step pay off in few-step sampling—our paper provides a training-free, interval-wise diagnostic for any previously unseen model; this local analysis requires no error-propagation assumptions, whereas Proposition 5 is a conditional guarantee when extending local measurements to full-trajectory planning.

1. Diagnosing unseen models via the (h,k)(h, k) window (W1a, Q1). Prior discussions of high-order solvers or step ratios rely on qualitative rules of thumb from known models (e.g., avoiding high order at low NFE or ending with Euler), without empirical verification on a new model. Given a previously unseen checkpoint—newly trained, fine-tuned, or distilled—whether and where extra evaluations per step pay off cannot be known from prior heuristics or a fixed h/kh/k rule alone. Anchoring xt∗=(1−t)x0+tεx_t^* = (1 - t)x_0 + t\varepsilon at the model's sample x0(ε)x_0(\varepsilon) lets all solvers start from xt∗x_t^* on [t,k][t, k] against xk∗x_k^* on one scale with zero inherited error, so our exact decomposition (Eqs. 1–2), local bounds (Props. 1, 3, 4), and 400-step ODE check (Appendix R.2) hold independently of Propositions 2 and 5.

On Q1, h/k<5h/k < 5 is only an outer stiffness boundary near t=0t = 0; our primary finding is the empirically measured (h,k)(h, k) window structure across models, which we will foreground ahead of h/k<5h/k < 5 in the abstract and introduction. For instance, Heun beats two Euler steps on 39%, 29%, 15%, and 0% of intervals on Qwen-Image, Qwen-Image-Edit, SDXL, and distilled Z-Image-Turbo, with window interiors shifting under CFG while the upper cutoff stays stable (Appendix R.6). Beyond full 1→01 \to 0 generation, this interval-wise view diagnoses unseen or distilled models directly and informs partial-interval or downstream tasks such as editing—enabling zero-recalibration transfer from Qwen-Image to Qwen-Image-Edit across 272 edits (>0.97>0.97 edge-rank correlation, 96.6% Euler-vs.-Heun agreement; Table 3, Sec. 5.5).

2. Combinatorial intractability and real-CFG validation (W1a). Measuring solvers along rolled-out trajectories conflates local behavior with accumulated error propagation from earlier steps, making local solver comparisons unreliable; conversely, explicitly tracking path-dependent drift and continuation Jacobians across all possible solver-and-timestep sequences is combinatorially explosive and computationally infeasible. This fundamental dilemma is why we decouple drift-free local measurement from global planning, where backward dynamic programming minimizes unweighted cumulative error as a tractable additive surrogate (preserving ~100-NFE Euler calibration) without claiming unconditional optimality.

Because Propositions 2 and 5 give idealized sufficient conditions that do not hold globally under CFG, we tested true endpoint error E∥xN−x0∥2\mathbb{E}\|x_N - x_0\|^2 directly against path cost under standard CFG with real drift:

  • Euler paths: True endpoint error is 0.45–0.48 of path cost at 5–15 NFE (0.29 at 25 NFE), stays below path cost on 96.0%–98.5% of random paths (Appendix R.3), and beats native and time-shifted schedules across all four models (Appendix R.7).
  • Multi-stage paths: Drift sensitivity emerges near t=0t = 0 as stiffness JkJ_k grows (Prop. 3): on 10-NFE Qwen-Image planner, stopping Heun at t=0.31t = 0.31 rather than t=0.12t = 0.12 recovers GenEval from 0.792 to 0.821 (Appendices R.4–R.5).

Error propagation under CFG thus affects surrogate tightness only for multi-stage steps near t=0t = 0, not our local findings or planned schedule gains.

3. Metric scope and camera-ready revisions. Reference-based errors characterize solver behavior relative to the model's own sample x0(ε)x_0(\varepsilon), while generation quality is evaluated independently using GenEval, ImageReward, and preference tests. Clarifying our Nov 16 reply to Reviewer tdPV: we will not move ten appendix sections (R.1–R.10) into the main text or restructure the paper. Core clarifications and paired 95% CI findings are already within the 10-page limit (in blue), and camera-ready edits only add the compact global-error/time-shift summary and 2D schematic (Fig. 9) via the extra page, plus light copy-editing and labeling Props. 3–4 as first-order characterizations.

In summary, our primary contribution is a heuristic-free method to analyze local solver behavior across models on a unified scale; global planning is a further application rather than a prerequisite for its validity. We thank you for your scrutiny of the global bound and respectfully ask that this conditional guarantee (W1a) be weighed separately from the scientific soundness of our core local findings.

Thank you for the clarification. I agree that the local measurements (Equations 1-2, the window, and the ODE check in Appendix R.2) do not depend on Propositions 2 and 5, and that Appendices R.3 and R.7 support the planned Euler schedules under CFG. I also appreciate the commitment to lead with the measured window in the abstract and introduction.

However, this separation should be made clearly in the paper itself, not only in the discussion here. The current paper presents the local analysis, the path cost and the planner as one framework, and the abstract presents the path cost as a bound on the global error. A reader should be able to see which results need the assumptions of Propositions 2 and 5 and which do not.

Moreover, if the core finding is only the local behavior and the planner is a further application, the value and novelty of the paper are significantly reduced. The planner is listed as a contribution, and most of the image quality results (Tables 1-3) are its outputs.

My concern about the planner also remains. Estimating the amplification does not require tracking all paths: measuring it along the few planned paths would show whether the assumption holds where it is used.

Since the response states that the paper will not be restructured, my concern about presentation also remains, and my score for the paper is unchanged.

Comments on the paper

3 comments
Reviewer_tdPVOpenReview ↗

Reviewer response

I thank the authors for addressing my concerns and for the additional experiments.

  1. High accuracy solvers: I thank the authors for the clarification. Indeed I have previously miss interpreted the bound.

  2. Global error minimization: Indeed your experiment shows that minimizing the bound reduce global error. However, if I understand correctly you compare against a "random path" which my guess is that it is a randomly sampled strictly decreasing sequence of time steps from 1 to 0. That sounds like a weak baseline. Why not provide numbers for the native schedule? Or the shift time transformation which I have already mentioned?

  3. Mixed solver schedule: It is my understanding that the mixed solver schedule is a proposition of the authors and not a standard methodology for ODE solvers. In practice authors did not show that the mixed solver schedule improves performance. In the rebuttal the authors claim: "This is the paper's finding, not a failure of the method". If the authors claim their method proves optimality of the solver, that would be a strong result. However I did not see such optimality claim in the paper. Thus, it appears to me as a negative results.

  4. Noise/time schedule: The importance of optimizing the noise schedule for diffusion and flow models have already been shown by numerous previous works [1,2,3,4], and probably there are many more. Thus, the provided comparison only between the native schedule and planned schedule, indeed shows that the planned schedule improves upon the native, but it is not a strong result.

    Furthermore, the provided a grid search for the shift time transformation over 4 values of the shift already yields a better performing noise schedule than the proposed planned schedule. this also emphasize my previous point.

    The authors have rightfully mention that this grid search of the shift time parameter is done using an access to the GenEval benchmark which is considered unfair. I agree that a consistent validation metric or "stop criterion" is a good standard. So I ask the authors again (as in point 2), why not provide the global error of shift time grid search? Given the authors argument, this seems called for.

    If the authors will prove that their method can yield better performance than a grid search over the shift time parameter in a fair and extensive comparison, I am willing to increase my score.

[1] Karras et al., Elucidating the Design Space of Diffusion-Based Generative Models, NeurIPS 2022.

[2] Song et al., Denoising Diffusion Implicit Models, ICLR 2021.

[3] Lu et al., DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps, NeurIPS 2022.

[4] Esser et al., Scaling Rectified Flow Transformers for High-Resolution Image Synthesis, ICML 2024.

Reviewer_tdPVOpenReview ↗

Reviewer response

I am glad the authors found my suggestion excellent.

Regarding the authors comment "The planned schedule reaches this without searching over shift values or scoring any image.": The shit time transformation yields a one-parameter family of time/noise schedules. As discussed in previous results, this parameter can be optimized by minimizing global error. The simplest possible optimization algorithm is a grid search. Since the global error is cheap to evaluate (16 images is cheap) and there is only one parameter to optimize it appears to me like a good choice. Hence, this comment which suggest that suggest that avoiding grid search is somehow preferable is strange to me, if anything, simplicity is an advantage. Additionally, as per the authors wards the the values s=4,4,3s = 4, 4, 3 where selected without scoring any images.

Regarding the results, The table below shows the comparison of the planned schedule vs. shit time schedule.

ModelGlobal error ↓ (shift time)Global error ↓ (planned)GenEval ↑ (shift time)GenEval ↑ (planned)
Qwen-Image0.1160.0980.8420.830
FLUX.1-dev0.4980.4110.6480.644
SD3.5-Large0.3980.3520.6840.683

This show that my initial review was right to call the results presented in the paper weak, since it is currently reports only GenEval, and we see these are on par with the shift time transformation and not better.

On the positive side, we see that aligned with its derivation, the method minimize global error. Probably the authors should consider adding global error to their tables in the papers (the lack of global error is also mentioned in my initial review).

Overall this has the potential of being a worthy paper, so I will increase my score. However, it appears to me that this paper was not fully ready to be submitted at time of the submission:

  1. The small shift time experiment prove that it is clearly missing a more exhaustive comparison to baselines.

  2. The paper should probably put more emphasis in its experiment section on the metric that its method derivation claim to improve i.e., global error.

  3. If the main practical results is the planned schedules, it would be helpful to the reader if the schedules them selves were provided in the paper.

  4. (OPTIONAL) One or two figures illustrating in 2d space the authors definition of error and the bound in proposition 5 could help the reader to understand the method more easly.

AuthorsOpenReview ↗

Update to the summary of our rebuttal

We thank the reviewers for the discussion since our earlier summary. Its current state:

  • Reviewer tdPV checked that the planned schedules lower the true global error (Appendix R.7) and raised the score to a positive rating.
  • Reviewer uqDx is satisfied with the answers and the added experiments, and keeps a positive score.
  • Reviewer 46sB, whose detailed questions over three rounds helped us make the paper clearer, agrees with our answers to the first-round questions on the error measure and on the planned schedules: the local measurements (Equations 1–2, the window, Appendix R.2) do not depend on Propositions 2 and 5, and Appendices R.3 and R.7 support the planned Euler schedules under CFG. His remaining points are answered in our latest reply to him.

The revised PDF states in Section 3.3 which results rely on Propositions 2 and 5 and which do not; the abstract and the introduction now report the measured window ahead of the h/k = 5 threshold, and the abstract states the bound with its condition. Propositions 3 and 4 are labelled as first-order characterizations.

To make our earlier summary more precise: the local analysis and the planner form one method. The analysis shows where an extra evaluation pays off; the planner turns these measurements into samplers that can be compared at equal NFE (Table 1) and that transfer to editing without recalibration (Table 3).

The anonymous repository (https://anonymous.4open.science/r/repo_solver-DF4E) contains the core code, with which the main results can be reproduced directly: the GenEval results of every planned solver in Table 1, the editing results of Table 3, and the planned paths at any NFE budget, either from the released graphs or from a graph built anew. It also contains a two-minute method-overview video.

We thank the Area Chair for handling our submission, and all three reviewers for the time they have given to this paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.