- 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?
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.