acceptodds
Under review as a conference paper at ICLR 2027

Projection Bottleneck in Adaptive Diffusion Sampling

Abstract

Adaptive timestep selection is widely used to accelerate sampling in continuous-time generative models, yet its effectiveness remains poorly understood. Existing methods generally posit that structured trajectory signals can be leveraged to derive improved discretization schemes. In this work, we show that this assumption can break down for projection-based step-wise adaptation under fixed sampling budgets. We find that trajectory-level signals, including hidden speed, curvature, and topological summaries, exhibit strong and consistent temporal structure across models and datasets. However, directly using these signals for step-wise adaptive control does not consistently improve sampling quality and can lead to severe degradation. Empirically, diverse controllers often produce identical timestep schedules and even identical generated samples. We identify an important mechanism underlying this behavior as a projection bottleneck: under discretization constraints, distinct adaptive proposals can be mapped onto the same feasible schedule through a many-to-one projection. We characterize this effect theoretically through active-set non-injectivity and local loss of controller degrees of freedom under saturation. Our results suggest that effective adaptive sampling depends not only on extracting trajectory information, but also on preserving that information during schedule construction, motivating alternative allocation strategies that avoid projection-induced information loss.

Primary area:
generative models
Keywords:
Diffusion Models, Adaptive Sampling, Timestep Scheduling, Generative Model Inference, Flow Matching, Numerical Discretization, Sampling Efficiency, Adaptive Step Size

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_e3wBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies different strategies of timestep selection in diffusion sampling. The authors argue the existing method often suffer from "projection bottleneck" so different methods may even lead to the same schedule. For example, suppose there's hard constraint minimal step-size = 0.01, then one scheduler proposing 0.001 and the other proposing 0.005 will be adjusted to the same step-size (and thus not useful).

To overcome this issue, the authors propose to use a global allocation strategy, that assigns weight to each timestep, and then normalize the weights to get the step sizes.

Critical strengths

The problem is practically relevant.

The phenemonon that different diffusion models have common speed, curvature, and topological summaries is interesting, and may be valuable for other applications.

The paper is generally well-written.

Critical weaknesses

The paper compares the local control that uses Unified multi-scale score defined in their eqn (10). However, I felt this is different from the common methods used in the existing literature (including those cited in the author's paper), for example, DPM-solver. Thus I felt the specific method where the authors critiqued is not commonly used.

The global allocation strategy does not seem to be much better than the uniform stepsize. Thus I felt the methodological contribution of this paper is slightly overstated.

The handcrafted schedule is somewhat confusing.

I felt the overall writing is fine. However part of the paper seems unnecessarily abstract. For example, the definition of speed, curvature, and topological summaries. The projection bottleneck can also be explained in a simpler way.

In summary, I felt the paper claims 3 contributions: 1. a common phenomenon, 2. pointing out drawbacks of existing methods, 3. propose a new method. I found the first contribution interesting (though could be investigated deeper and more systematically), but am not currently convinced by the other two contributions.

Questions for authors
  1. Can you comment on how often is (10) used in actual diffusion sampling?

  2. Can you explain more on the handcrafted schedule?

  3. Can you comment on Text-to-Image models?

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

This paper studies adaptive timestep selection for continuous-time generative models and identifies a limitation of projection-based step-wise adaptation under fixed sampling budgets. Although trajectory signals such as hidden speed, curvature, and topological statistics exhibit clear temporal structure, directly using them for adaptive control does not consistently improve generation quality and can even cause severe degradation. The authors attribute this behavior to a projection bottleneck, where different adaptive proposals collapse to the same feasible timestep schedule through many-to-one projection. The paper provides both empirical evidence and theoretical analysis of this phenomenon, and argues that effective adaptive sampling should preserve trajectory information during schedule construction rather than lose it through projection.

Critical strengths
  1. The problem is insightful. The paper highlights that structured trajectory signals do not necessarily remain informative after timestep schedule construction, providing a useful perspective for understanding why adaptive sampling can fail.

  2. The diagnostic evidence is intuitive. Analyses of saturation rates, realized step sizes, and identical generated samples clearly illustrate how different controllers can collapse to the same timestep schedule under certain settings.

Critical weaknesses
  1. The core causal explanation is not yet sufficiently established. The severe degradation may arise from factors such as controller strength, step-size bounds, or the specific budget-correction procedure, rather than solely from projection-induced information loss. The paper lacks systematic ablations disentangling these effects. Moreover, the global allocation formulation only enforces positivity and a unit-sum constraint, without clearly imposing the same per-step lower and upper bounds as the step-wise method. Therefore, the recovered performance cannot be attributed solely to avoiding repeated projection.

  2. The theoretical contribution is relatively limited. The two main theorems largely characterize standard properties of Euclidean projection onto a bounded simplex, namely its many-to-one nature and the reduction of effective degrees of freedom under saturation. However, the paper does not establish a quantitative connection between these properties and sampling error or generation quality degradation. It is also not fully clarified whether the clipping, normalization, and budget-correction procedures used in practice are exactly equivalent to the Euclidean projection analyzed theoretically.

  3. The practicality and computational cost of global allocation are unclear. The method appears to require sensitivity scores along the full trajectory, but these scores depend on intermediate states that have not yet been generated. The paper does not clearly explain whether this requires an additional pilot trajectory, offline statistics, or iterative refinement, nor whether such costs are included in the matched-NFE comparisons. This makes the practical efficiency and fairness of the comparison difficult to assess.

  4. The experiments are insufficient to support broad generality. The evaluation mainly uses conventional U-Net models, DDPM/DDIM, and early Rectified Flow settings, without comparisons on modern large-scale text-to-image models or mature adaptive ODE solvers. In the main experiments, global allocation mostly recovers performance close to hand-designed schedules rather than showing a consistent advantage. The positive AFHQ-Cat result additionally uses a different scheduling method and only 500 generated images per seed, which provides limited evidence for the proposed approach.

Questions for authors

Refer to Weakness.

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

This paper studies how to choose sampling timesteps for diffusion and flow models under a fixed NFE budget. The motivation is that trajectory signals may show clear patterns, yet using them to adjust timesteps does not always improve sample quality. The authors examine hidden-feature speed, curvature, and topological summaries, and report cases where different step-wise controllers produce the same schedules and outputs after feasibility constraints are applied. They study this behavior using properties of Euclidean projection and propose a global allocation method that sets all step sizes jointly. The main comparison on CIFAR-10 shows that global allocation recovers sample quality comparable to the handcrafted schedule. Additional experiments examine trajectory signals across datasets and evaluate a different phase-aware schedule on AFHQ-Cat.

Critical strengths
  1. The paper reports a concrete example of how adaptive timestep control can fail. In the reported saturated setting, different curvature weights lead to identical schedules and byte-identical samples. Examining the actual schedules alongside FID helps explain what the controllers are doing and makes the negative result easier to investigate.

  2. I find the distinction between measuring changes along a trajectory and using those measurements to set timesteps useful. The global allocation experiment also provides a constructive alternative, recovering sample quality comparable to the handcrafted baseline under the reported NFE budgets.

Critical weaknesses
  1. It is still unclear how much of the drop in sample quality comes from the projection step. Section 4 studies Euclidean projection, but Eq. (13) describes clipping, normalization, and budget correction without showing that these implement the same operation. The theorems explain how different proposals can produce the same schedule, but do not establish when this leads to worse FID. Figure 3 also shows that much of the remaining integration interval is assigned to a few large steps. The experiments do not separate the effects of the original proposals, the constraint-handling procedure, and the resulting allocation. Global allocation changes several of these components together, and it is unclear whether it enforces the same step-size bounds. In addition, several RschedR_{\mathrm{sched}} values in Table 3 exceed the 2/N2/N bound implied by Eq. (29). This needs clarification before the metric can support the explanation quantitatively.

  2. The paper could more clearly establish which practical methods face this problem and when global allocation would help. It does not identify existing schedulers that use the full procedure being analyzed, and the discussion of related schedule-optimization methods such as AYS and GITS is limited. On CIFAR-10, global allocation matches the handcrafted baseline, but the experiments do not show how much the trajectory signals contribute beyond the base schedule and the allocation rule. The results on other datasets mainly examine trajectory structure, while AFHQ-Cat uses a different scheduler, so they do not directly validate the same remedy more broadly. The paper also leaves unclear how the full sequence of trajectory scores is obtained and how much additional computation this requires. These details matter for judging the method's usefulness in practice.

Questions for authors
  1. What exact algorithm is used to enforce the sampling constraints, including the step-size bounds and budget correction? Does it implement the Euclidean projection in Eq. (20), and does global allocation enforce the same bounds?

  2. How is RschedR_{\mathrm{sched}} calculated in Table 3? For positive step sizes summing to one, Eq. (29) gives Rsched<2/NR_{\mathrm{sched}}<2/N, but several reported values exceed this bound.

  3. What base schedule is used for global allocation? How are the trajectory scores computed and normalized, and are they obtained online or from reference trajectories? Please clarify any additional model evaluations or calibration costs.

  4. How many generated samples and which reference statistics are used for each FID evaluation? Were the handcrafted results in Tables 1 and 2 obtained under the same evaluation protocol?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.