acceptodds
Under review as a conference paper at ICLR 2027

Model-Aware Schedules Improve Generation via Fiberwise Optimal Transport

17 views

Abstract

In generative modeling with diffusion and flow matching, schedules control the signal and noise coefficients that mix data and noise along affine probability paths. Minimizing a kinetic action defined on coefficient paths, motivated by optimal transport, helps explain strong baselines but remains model-agnostic and ignores prediction error. Here we introduce a model-aware schedule construction that uses fiberwise optimal transport to improve generation quality. At a fixed time and state on the probability path, compatible signal/noise decompositions form an affine fiber. We define a fiberwise prediction risk by averaging optimal-transport costs between the true and predictor-induced decompositions within these fibers. On a fixed coefficient curve, combining this risk with coefficient-path kinetic action yields a closed-form optimal time allocation. This construction extends to general linear prediction targets, and the risk profile can be estimated from an early baseline checkpoint. We evaluate denoising diffusion probabilistic models (DDPMs) and flow matching across prediction targets, training configurations, risk-estimation checkpoints, datasets, and architectures. Our model-aware schedules consistently outperform strong baselines, reducing flow-matching FID by 38.6% on CIFAR-10 at 16 function evaluations. Each model-agnostic kinetic baseline determines its own kinetic reference coordinate. In these coordinates, fiberwise-risk profiles from independently trained models in different settings align closely after normalization to unit area. The resulting schedule deformations used in training also align, suggesting empirical universality across the evaluated models and settings. A frozen analytic allocation template retains most of the model-aware gain without further risk estimation or model-specific fitting.

Primary area:
generative models
Keywords:
diffusion models, flow matching, schedule optimization, fiberwise optimal transport, time reparameterization, empirical universality

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_M8NoOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a model-aware scheduling method for diffusion and flow-matching models that improves generation quality by allocating time according to prediction difficulty. The authors define a fiberwise optimal-transport risk that reduces to a weighted prediction MSE and combine it with coefficient-path kinetic action to derive a closed-form time-allocation rule. The risk profile is estimated from a baseline checkpoint, and the resulting schedule is used to retrain the model. Experiments across datasets, architectures, prediction targets, and numerical solvers show improved FID over standard schedules. The paper also reports similar normalized risk profiles across different models and proposes a reusable analytic schedule template.

Critical strengths

A> The fiberwise construction is a simple and easy to implement. The idea to implement a risk evaluation based flow method is novel.

B> The pipeline estimates the risk from a given checkpoint, and then finetunes the pipeline to produce better results.

Critical weaknesses

A> The comparisons provided in the paper are insufficient. There have been prior works like [1] that have performed trajectory correction to perform optimal inference. Furthermore the baselines chosen for both datasets do not perform at SOTA - for instance there are publicly available DDPM models as [2] that perform better than the baseline reported. This might affect the quality of generation and subsequently the FID values, and also provide a larger range for NFE.

B> The paper does not explain the overhead incurred due to the extra computations. The paper discusses 16 function evaluations/NFE - but the paper does not explain how the timesteps are skipped or the intermediate features are passed on.

C> The motivation for the proposed approach is unclear. The authors discuss prior work that optimizes marginal kinetic energy over both the coefficient curve and its traversal in flow-matching settings. They argue that existing kinetic schedules are model-agnostic and ignore prediction error, but do not adequately explain why this constitutes a fundamental limitation or establish why the proposed fiberwise OT risk is the appropriate quantity to optimize. Furthermore, the connection between minimizing the proposed objective and improving generation quality remains insufficiently justified.

[1] Karras, T., Aittala, M., Aila, T., & Laine, S. (2022). Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35, 26565-26577. [2] https://huggingface.co/google/ddpm-cifar10-32

Questions for authors

The paper is a nice read. Please explain how much the overhead for the proposed pipeline will be, if it was to be extended to a larger dataset. Would it affect the training or function executions?

Although results are reported at matched NFE, it is unclear how the discrete timesteps are selected under the proposed modified schedule. For example, with 16 NFE, are the model evaluations distributed uniformly in the modified time coordinate, mapped from the original DDPM timestep grid, or selected according to the solver's discretization rule?

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

The paper describes two quantities for characterizing generative schedules in diffusion and flow matching. Specifically, given an affine interpolant xt=mtx0+stϵx_t=m_tx_0+s_t\epsilon of data x0x_0 and noise ϵ\epsilon, the kinetic action is defined as the integral of m˙t2+s˙t2\dot m_t^2+\dot s_t^2 over time, measuring how quickly the interpolation coefficients change. This quantity is model-agnostic and is minimized by constant-speed traversal of a fixed coefficient curve. On the other hand, the prediction risk at time tt is defined through fiberwise optimal transport and reduces to the noise-variance-weighted epsilon-prediction loss, measuring the model's prediction discrepancy at different noise levels.

Given these quantities, the authors propose to estimate prediction risk using a baseline diffusion or flow model and optimize the time parameterization of the fixed coefficient curve by minimizing kinetic action subject to a constraint on integrated prediction risk. Using a Lagrangian formulation, they derive a closed-form optimal schedule that balances the two objectives. A new diffusion or flow model is then trained from scratch using the optimized schedule, which is also used during inference. Experiments demonstrate improved generation quality on standard benchmarks, including CIFAR-10 and ImageNet-64.

Critical strengths

None.

Critical weaknesses

Unclear relationship between the proposed objective and generation quality. The paper optimizes a combination of kinetic action and prediction risk, but it is unclear whether these quantities jointly provide a meaningful surrogate for generation quality. Although kinetic action is motivated by existing Wasserstein generation-error sensitivity results, the authors do not establish a direct connection between their combined objective and actual generation error or FID. In particular, reducing integrated prediction risk by traversing high-error regions more quickly does not necessarily improve the model's predictions or numerical integration accuracy. A stronger theoretical justification or empirical analysis of the relationship between the proposed objective and generation quality would strengthen the contribution.

Unclear necessity of retraining from scratch. The proposed method requires training a new generative model from scratch under the optimized schedule, but the paper does not compare this approach against applying the same time reparameterization to an existing pretrained model at inference time. In flow matching, such a reparameterization can be implemented through a corresponding transformation of the pretrained velocity field without additional training. It is therefore unclear how much of the reported improvement comes from inference-time schedule optimization rather than retraining. Comparisons with inference-only schedule warping and fine-tuning would help establish whether the additional training cost is justified.

Insufficient disentanglement of the mechanisms behind the improvement. The proposed time reparameterization simultaneously changes the effective training-time sampling distribution, rescales the velocity prediction target (and hence the effective loss weighting), and modifies the numerical discretization used during inference. Although the paper evaluates robustness across different training-time sampling distributions and numerical solvers, these experiments do not isolate the contributions of the individual mechanisms. Consequently, it remains unclear whether the improvements arise from the proposed kinetic-risk optimization principle or from conventional effects of training allocation, loss weighting, and inference-time discretization. Additional ablations that independently control these factors would strengthen the empirical justification.

Limited evaluation on large-scale generative models. The generation experiments are restricted to unconditional CIFAR-10 and ImageNet-64, leaving the effectiveness of the proposed approach on modern, high-resolution conditional generative models unclear. Although the authors analyze prediction-risk profiles of pretrained DiT and 2-Rectified Flow models, they do not evaluate actual generation improvements under the proposed schedules. Experiments on larger-scale models would strengthen the generality and practical significance of the findings.

Questions for authors
  • Could the authors evaluate multiple candidate schedules (e.g., alternative parametric schedules and reversed risk-based allocations) and examine whether the proposed kinetic-risk objective correlates with generation quality (FID) across these schedules?
  • Could the authors compare training from scratch against inference-only schedule warping or fine-tuning to justify the additional training cost?
  • Could the authors disentangle the contributions of training-time sampling, loss weighting, and inference-time discretization to the reported improvements?
  • Could the authors evaluate the proposed schedule on higher-resolution conditional generation benchmarks (e.g., ImageNet-256) to demonstrate its effectiveness beyond low-resolution unconditional generation?
Reviewer_b5w6OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a model-aware time-scheduling strategy for diffusion and flow-matching models. The method combines fiberwise prediction risk with coefficient-path kinetic action to optimize time allocation along a fixed coefficient path. Using risk estimates obtained from a baseline model checkpoint, the authors derive a solution for time allocation and apply the resulting schedule to train new models. Experiments on CIFAR-10 and ImageNet-64 demonstrate improvements in generation quality. Furthermore, the authors find different models exhibit similar normalized risk profiles, and introduce a fixed analytical scheduling template that retains most of the performance gains without requiring additional model-specific risk estimation.

Critical strengths
  1. The paper combines fiberwise optimal transport and practical schedule design, showing that the risk can be expressed as a weighted prediction MSE and that the optimal time allocation admits a closed-form solution. T

  2. The authors demonstrate improvements across prediction targets, architectures, and training settings, and further show that a fixed analytic template retains most of these gains without additional risk estimation, highlighting the method’s practical utility.

Critical weaknesses
  1. The model-aware pipeline requires a baseline training stage for risk estimation followed by training a new model from scratch, but the reported FID comparisons do not account for the full computational cost of both stages. It remains unclear whether the gains persist when the baseline receives the same total computational budget (Section 6.1).

  2. The paper discusses existing adaptive schedule methods, such as Constant Rate Scheduling (Section 2, lines 98–103), but provides no direct experimental comparison with them (Section 6.2, Tables 1–4). The results establish improvements over standard schedules without establishing the advantage of the proposed approach over closely related schedule optimization methods .

Questions for authors
  1. Under the same total computational budget, how does the model-aware pipeline, including baseline training, risk estimation, and training the final model, compare with training the baseline for longer?

  2. How does the proposed method compare with other related method (described in "Schedule and probability-path design." line 98, e.g. Constant Rate Scheduling ) in generation quality and total computational cost when the model architecture, training data, and number of function evaluations are held fixed?

Comments on the paper

1 comment

Thank you for the detailed rebuttal and the additional experiments. I appreciate that the authors have addressed several of my original concerns with meaningful new evaluations.

In particular, the inference-only experiments are informative. The observation that inference-only schedule reparameterization recovers 93.2% of the total improvement for OT-CFM and 73.6% for DDPM substantially clarifies the role of retraining. The additional grid-only control further demonstrates that the benefit extends beyond choosing nonuniform step boundaries. The reversed-risk allocation experiment also provides useful empirical support for the proposed allocation direction. Finally, the new ImageNet-256 and AFHQ-Cat 256 results help address my concerns about generality.

Nevertheless, I have two remaining concerns.

First, the relationship between the proposed kinetic–risk objective and generation quality remains insufficiently established. The reversed-allocation experiment provides a useful sanity check, but comparing three schedules in a single experimental setting does not establish that the proposed objective generally correlates with generation quality or selects a preferable schedule among alternative nonuniform parameterizations. The theoretical justification also remains indirect, as the combined objective is not shown to bound generation or numerical integration error.

Second, the inference-only results suggest that the primary improvements, especially for OT-CFM, arise from the numerical effects of time reparameterization rather than changes in the learned generative model. This makes comparisons with existing inference-time scheduling methods particularly important, including Align Your Steps (Sabour et al., ICML 2024), Entropic Time Schedulers (Stancevic et al., NeurIPS 2025), and Bellman Optimal Stepsize Straightening (Nguyen et al., ICLR 2024). Although these methods are discussed in the related work, direct comparisons would help establish the distinct empirical advantages of the proposed kinetic–risk objective. I acknowledge that exact comparisons may require adaptations due to differences in methodology, but consider their absence an important remaining empirical limitation rather than a fundamental soundness issue.

Overall, the additional experiments have substantially strengthened the empirical evidence, and I consider my concerns about retraining necessity and limited evaluation largely addressed. My remaining reservations concern the justification and distinctiveness of the proposed optimization objective, rather than the existence of empirical improvements. I have therefore updated my assessment accordingly to weak rejection.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.