acceptodds
Withdrawn from ICLR 2027

When ODE Fidelity Misleads Diffusion Timestep Search

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Diffusion models can generate images by solving an ordinary differential equation (ODE) that transforms noise into an image. A numerical solver follows this process through repeated model evaluations, using each prediction to update the current state. Fast samplers use fewer evaluations to reduce computation. A schedule specifies when these evaluations occur. We examine choosing among candidate schedules by minimizing the distance from the final state to a reference computed with many more steps. We change the evaluation times while keeping the model, sampling procedure, and total number of evaluations the same. The selected schedules can bring the final state closer to the reference yet score worse on automated preference metrics and on metrics comparing generated and real image distributions. Under strong classifier-free guidance, which increases the text description's influence, the search can leave long gaps in the low-noise region near the end of generation. For a schedule with such gaps, smaller steps early in generation reduce the distance to the reference more, whereas smaller steps near the end improve the learned image-preference score PickScore more. Using fewer evaluations early and more near the end can improve PickScore without increasing the total computation, although the distance to the reference increases. Constrained endpoint search limits the largest low-noise gap and recovers most of the PickScore loss in the tested candidate sets on Stable Diffusion 1.5, Stable Diffusion XL (SDXL), and PixArt-Σ. The failure lies in how the search allocates computation: evaluation times that improve numerical accuracy can leave too few evaluations where they matter more for image quality.

Primary area:
generative models
Keywords:
diffusion models, timestep scheduling, ODE sampling, classifier-free guidance, few-step generation, ODE solvers

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rz8SOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies a failure mode of selecting diffusion sampling schedules by endpoint error, the relative L2 distance between a few-step sample and a high-step reference generated from the same noise. Holding the model, solver, number of function evaluations, and initial noise fixed, the authors show that endpoint error search can favor schedules with large low-noise gaps and worse image quality scores under strong classifier-free guidance. They explain this disagreement through controlled interventions and a two interval models, and propose constrained endpoint search, which limits the maximum low noise gap before minimizing endpoint error.

Critical strengths
  1. This paper presents a rigorous and well-controlled diagnostic analysis. While keeping the model, solver, NFE, prompt information and initial noise constant, only the evaluation timesteps were varied. This analysis goes beyond a mere correlation study, moving the three evaluation points from high-noise regions to low-noise gaps improved the PickScore by (0.124, [0.051, 0.197]), whilst simultaneously increasing the endpoint error. By utilizing held-out prompts, paired confidence intervals, multiple metrics and targeted interventions, the key SD1.5 results were rendered compelling.

  2. This paper proposes a simple and practical early-warning and intervention strategy. Checking low-noise intervals whilst optimizing the distance to the reference point is easy to implement, and constrained endpoint search is able to recover the majority of the observed PickScore loss across multiple test scenarios. Furthermore, the paper candidly explains the settings under which the methods's failure may be mitigated or eliminated.

Critical weaknesses
  1. It remains unclear how often this failure affects practical timestep search methods. The main result selects among 57 fixed candidate schedules using strong guidance on SD1.5. The effect becomes smaller with a larger candidate set, is mixed on SD3.5, reverses without guidance, and is not clearly observed when the OSS adaptations use their own first-order update. Thus, the evidence supports a failure mode under specific conditions, but not yet a general limitation of endpoint-based timestep search. The paper should narrow its claims or demonstrate the issue in more representative published search methods.

  2. The explanation and proposed remedy remain limited. The two-interval model assumes that image quality is more sensitive to errors in later stages, but it neither explains nor empirically predicts when this condition holds in practical diffusion models. Constrained endpoint seach also relies on empirically selected spacing parameter, provides incomplete recovery in some settings, and can underperform alternatives such as AYS. Consequently, its practical advantage over existing schedules or direct quality based selection is not yet clear. Furthermore, the core comparisons are primarily based on implicit endpoint L2 error and learned perference metrics. Distributional metrics provide supporting evidence, but human evaluations are limited to the relatively small scope of the SDXL comparison. The conclusions should therefore be framed around the evaluated fidelity and quality measures rather than ODE fidelity and perceptual quality in general.

Questions for authors
  1. Given that the effect weakens with a larger candidate set, is mixed on SD3.5, reverses without guidance, and is not clearly present for OSS with its own update, in which practical settings do the authors expect published endpoint-based search methods to encounter this problem?

  2. Can the author verify the hypothesis that image quality is more sensitive to low-noise errors through empirical research, rather than relying directly on this hypothesis in the two-interval model? Can this theory predict when inconsistencies will occur?

  3. Can the authors provide broader human evaluation across SD1.5 and PixArt-Σ\Sigma to verify that the reported preference-metric differences correspond to human judgments?

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

This paper studies whether better ODE solving accuracy necessarily leads to better image quality in diffusion sampling. The authors compare different timestep schedules under the same model, solver, and NFE, and find that schedules with smaller endpoint errors could actually produce worse images. They further find that this happens especially under strong CFG, where endpoint-based search tends to leave large gaps in the low-noise region. So they propose a simple constraint on the maximum low-noise timestep gap during schedule search. Results on SD1.5, SDXL, and PixArt show that such constraint can recover much of the image quality loss.

Critical strengths
  • I think this paper studies an interesting problem. The finding that a more accurate ODE solution does not necessarily lead to better images is quite meaningful for diffusion sampling, especially when trying to reduce NFE.
  • The experiments are generally well controlled. The authors fix the backbone, solver and NFE when comparing different schedules. The analysis on different CFG scales, solvers, and models is also appreciated. Besides, they put quite some effort into validating the reference solutions and numerical precision.
  • The proposed constraint is simple and does not introduce extra inference cost. Results show clear improvement over unconstrained endpoint search in some settings. The evaluation covers various metrics, and the additional human preference study is also a plus.
Critical weaknesses
  • Although constrained endpoint search improves over unconstrained search, in Table 8, it performs almost the same as a simple uniform schedule (20.64 vs. 20.63 PickScore), and even worse than AYS (20.73). Considering that endpoint search requires additional computation to generate reference trajectories and evaluate different candidates, I am not sure if such improvement is worth the cost. If the final goal is image quality, why not directly use existing schedules like AYS, or optimize perceptual metrics instead? I think the authors should better clarify the advantage of their method in practical use.
  • The generalizability is also a bit unclear to me. Most of the experiments are conducted on SD1.5 with strong CFG and relatively small NFE. Although SDXL and PixArt are also evaluated, the results on SD3.5 are rather mixed, and the reference convergence under strong CFG is not fully verified. Besides, the proposed constraint uses an empirically chosen threshold, which may not work equally well for different noise schedules. It would be nice to see more experiments on recent flow-matching models to show whether the same observation still holds.
Questions for authors
  • In Appendix A.7, the OSS adaptation with its own update does not show a distinguishable quality drop. Does this mean the problem is more related to the solver or search space, rather than the endpoint error objective itself?
  • Regarding the theoretical analysis, Proposition 1 assumes that late-stage errors have a larger impact on image quality, and then shows that the optimal allocation for image quality would favor later timesteps. I feel like this is more of an explanation under the assumed conditions, rather than really proving why the same thing happens in diffusion models. Although the empirical results do provide some support, it would be interesting to see if the authors could measure such sensitivity directly on real diffusion models.
Reviewer_x25cOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The work explores the impact of when evaluation happens in diffusion sampling schedulers. It observes that schedulers that optimize for small endpoint error often do not allocate many evaluations toward the end of denoising. As a result the output is numerically close to the reference but is not always a good image, as indicated by poor pickscore metrics. To correct this, the paper explores different strategies for when to evaluate. Using pickscore to search for example can lead to better pickscore but worse endpoint error. Guidance also is shown to impact the tradeoff between numerical accuracy and quality. Having the scheduler evaluate more during low noise gaps is shown to significantly improve pickscore, with high noise being more associated with improved end point accuracy. Using the findings the work proposes a constrained end point search. Overall, the work explores a wide range of scheduler evaluation strategies to balance accuracy with quality.

Critical strengths

The work makes simple and intuitive observations on how fast diffusion scheduler evaluations strategies impact the final quality of output images. The observation that endpoint search favors evaluation during early high noise steps (which impacts quality scores) is interesting and informative. Multiple scheduler strategies are evaluated, and all the experiments match the number of evaluations across the different strategies making it clear the importance of the evaluation timing alone.

Critical weaknesses

The evaluations of the different timestep strategies are weak. Most tables in the main text only focus on pickscore and endpoint error without reporting other scores (for example HPS, ImageReward, CLIP, DreamSim, LPIPs, FID etc). Some of these scores are mentioned in section 3.2 but the numerical values are not actually listed there or in any table. No human evaluations are performed and very few qualitative samples are given, which makes it difficult to assess how useful the proposed scheduling strategies are. It is also not clear to me which datasets were used for evaluations (it seems like only COCO captions, in which case I would be curious to know geneval results).

Relatedly, the presentation of the work also makes it difficult to follow. There are many experiments across the schedulers and noise gaps. It would be helpful to more clearly delineate which experiments are for the purpose of characterizing the problem versus addressing it. In addition having a clear methods section would also make it easier to follow all of the experimental setups and evaluation details. Because I found the flow of the paper confusing, it is possible some of these weaknesses I listed are in the

Questions for authors

Can you clarify what dataset(s) were used for experiments? Could the authors also provide more evaluation metrics scores beyond pickscore?

Could the authors clarify what was the final method they proposed? Does it have similar same cost as the endpoint search? Does the method work on bigger models like Flux models?

Could the authors also provide more qualitative image samples or comment on this? Similarly, are there any human evaluation results?

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

Related papers

Discussion (0)

Sign in to comment.