acceptodds
Under review as a conference paper at ICLR 2027

ACQUIRE ONCE,DEPLOY MANY:ZERO-TARGET-FIT TEMPORAL MEASURES FOR FEW-STEP FLOW MODELS

Abstract

Few-step flow samplers choose both the number and timing of field evaluations. Existing scheduling methods often repeat this decision for each target model. We instead acquire a temporal measure once and reuse it. At a budget of eight function evaluations (NFE = 8), explicit-midpoint second-order Runge–Kutta (RK2) has four executable intervals, represented by four positive widths. We fit these widths on an inexpensive unconditional CIFAR-10 Diffusion Transformer (DiT) using projected terminal Fréchet risk, freeze them, and copy them unchanged to ImageNet receivers. Deployment uses no target trajectory, feature, gradient, extra parameter, auxiliary network, or additional field evaluation. Canonical Fréchet Inception Distance (FID), evaluated with the Ablated Diffusion Model (ADM) suite over 50,000 samples (FID50k), falls from 8.3674 to 7.1696 on Scalable Interpolant Transformer (SiT-XL/2) and from 17.9714 to 14.6264 on Self-Flow; paired 95% bootstrap intervals for the FID reductions (uniform minus the acquired measure) are [1.1088, 1.2975] and [3.1486, 3.4903]. A second, independently seeded acquisition also improves fresh paired FID10k on SiT, Self-Flow, and REPresentation Alignment (REPA). A one-parameter power measure retains 83.7–91.8% of the flexible four-interval gain, with a receiver-dependent residual. The frozen measure transfers to an unseen Ralston RK2 tableau, and the allocation ordering agrees between ADM FID and independent DINOv2 metrics. Applied separately to forecasting, the same acquire-then-freeze protocol improves ensemble continuous ranked probability score in all six fresh receiver cells across electricity and short-horizon Lorenz forecasting at NFE = 8. The evidence supports amortized post-training temporal allocation within stated objective, source family, receiver family, solver, and budget contracts.

Primary area:
generative models
Keywords:
flow matching, generative models, numerical ODE solvers, temporal allocation, few-step sampling, sampling schedules, zero-target-fit transfer

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_BtTXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper does something with a few-step sampler. It tries to tackle some existing problems in scheduling (probably time-step) methods and tries to propose a new method that is more efficient with less number of function evaluations. It does this by formulating temporal allocation as a positive measure that fully specifies a finite-step solver action.

Critical strengths

None.

Critical weaknesses
  • The writing and presentation of the paper need to be improved a lot. Even in the Introduction, it is challenging to understand what the authors want to say. There are too many notations/concepts that are undefined, which makes it much more challenging to read.
Questions for authors

I'm gonna ask clarification for writing only for the Introduction, although there are a lot more throughout the paper.

  • line 36. "A numerical solver samples that field at selected times" <- Is this a correct explanation of flow matching?
  • line 46. "Our deployed object is a positive temporal measure" <- What do you mean?
  • line 46. "At midpoint NFE=8" <- What is midpoint here and you did not define what NFE is here yet...
  • line 49. "ImageNet" <- No citation?
  • line 49. "No target trajectory, feature, gradient, quality metric, or selection query enters acquisition." <- Need a lot more explanation on this.
  • line 51. "FID50k by 14.3% on SiT and 18.6% on Self-Flow" <- Define or cite FID50k, SiT, or Self-Flow.
  • line 53. "A directly acquired power measure recovers most of the gain. Transfer follows the temporal allocation, not one optimizer trajectory." <- Can you explain what you want to say here?

Some questions for the method section (below are questions I have while reading Section 3.1 only)

  • What is "corresponding time-only dynamics" in Equation 2. Is it derivative of x w.r.t t? How did you get the Equation 2?
  • What is xλx_\lambda and xIx_I? Did you define the difference between the two?
  • What is nn in tnt_n in physical grid?
  • What is Dhclock\mathcal D_h^{clock}? Did you define what it is?
  • Also same for Dhgrid\mathcal D_h^{grid}
  • What is nonlinear clock?
  • I cannot understand Equation 3 at all since there is no definition of D\mathcal D
13 comments
Comment 1

Weakness. The writing and presentation need substantial improvement. Even in the Introduction, it is hard to follow what the authors mean. Too many symbols and concepts are undefined.

Response 1

We made targeted revisions to the paper PDF to address the reading difficulties you raised.

I will comment here on the Introduction, although there are further issues elsewhere in the paper.

Comment 2

Question. Line 36: “A numerical solver samples that field at selected times.” Is this a correct explanation of Flow Matching?

Response 2

The sentence you question describes inference after training. Flow Matching learns a velocity field for continuous transport. At inference, the ODE solver evaluates this field to advance the trajectory. The sentence describes standard numerical sampling and contains no conceptual error.

Comment 3

Question. Line 46: “Our deployed object is a positive temporal measure.” What does this mean?

Response 3

Section 3.2 defines the measure you ask about.

It consists of four positive interval widths satisfying

mj>0,∑j=14mj=1.m_j>0,\qquad \sum_{j=1}^{4}m_j=1.

Their cumulative sums give the integration nodes. The four widths fully determine the midpoint computation at NFE = 8.

“Positive temporal measure” is our finite-dimensional representation of time allocation. It has four positive interval masses that sum to one. It has both a mathematical definition and a numerical implementation.

Comment 4

Question. Line 46: “At midpoint NFE = 8.” What is “midpoint”? You have not yet defined what NFE is.

Response 4

NFE stands for Number of Function Evaluations, a standard measure of computational cost in Flow Matching and diffusion sampling.

Section 4 states the following.

Here, NFE denotes the number of function evaluations, with one NFE equal to one neural-field evaluation.

“Midpoint” means the explicit second-order midpoint Runge–Kutta method (RK2). The abstract introduces it. Section 3.2 gives the full update.

Each interval uses two velocity-field evaluations, so NFE = 8 gives four intervals.

We suggest you read the paper carefully. It explains both terms.

Comment 5

Question. Line 49: “ImageNet” — no citation?

Response 5

You are right that the citation is missing here. We will add the original ImageNet reference in the final version.

Comment 6

Question. Line 49: “No target trajectory, feature, gradient, quality metric, or selection query enters acquisition.” This needs more explanation.

Response 6

The sentence you quote summarizes the procedure in Section 3.2.

  • We optimize the measure on a low-cost CIFAR-10 source model.
  • Before acquisition, we fix the source model, calibration noise, real-data feature bank, projection matrix, solver, and NFE budget.
  • We select the measure using a preset source search budget and an independent source validation stream.
  • We freeze the four interval widths and transfer them unchanged to the target.
  • We use no target trajectories, features, gradients, quality metrics, or selection queries to acquire or modify the measure. Target samples enter only the final evaluation.

We acquire the measure once on the source and deploy it directly on the target without target-side fitting. This distinguishes our method from target-specific optimization.

Your objection overlooks the full procedure given in Section 3.2.

Comment 7

Question. Line 51: “FID50k increases by 14.3% on SiT and by 18.6% on Self-Flow.” Please define FID50k, SiT, and Self-Flow or provide references.

Response 7

The manuscript provides definitions or references for the terms you identify.

  • FID50k. The abstract expands FID as Fréchet Inception Distance and specifies 50,000 generated images under the Ablated Diffusion Model (ADM) evaluation setup. Section 4 cites Heusel et al. (2017) and Parmar et al. (2022).
  • SiT. The abstract gives the full name, Scalable Interpolant Transformer. Section 4 cites Ma et al. (2024).
  • Self-Flow. This is the model’s full name. Section 4 and the reference list cite Chefer et al. (2026).

Your description of these terms as undefined or uncited is inaccurate. They are established terms, metrics, and model names widely discussed or cited in the field.

Comment 8

Question. Line 53: “A directly acquired power measure recovers most of the gain. Transfer follows the temporal allocation, not one optimizer trajectory.” Can you explain what you mean?

Response 8

The “directly acquired power measure” is a single-parameter power schedule optimized on the source. The source search determines its exponent.

Section 5.3 reports that it retains 91.5%, 83.7%, and 91.8% of the four-interval gain on the three targets, respectively.

The second sentence refers to two independent Sobol/GP searches. They gave different interval widths, but both schedules had the widest interval last. Both beat uniform allocation in all eight cross-stream evaluations.

The recurring late-loaded allocation transfers. Transfer does not require a particular optimizer trajectory or identical interval widths across runs.

Your question overlooks the independent acquisition and cross-model transfer experiments reported in Section 5.3.

I also have several questions about the Methods section. The following arose while I was reading Section 3.1 alone.

Comment 9

Question. In Equation 2, what are the “corresponding time-only dynamics”? Is this the derivative of x? How do you derive Equation 2?

Response 9

Yes. You are asking about the derivative of the transformed state xλ(t)x_\lambda(t) with respect to physical time tt.

Equation 2 follows directly from the chain rule.

Let the original dynamics be:

dy(τ)dτ=uθ(τ,y(τ)),\frac{dy(\tau)}{d\tau} =u_\theta(\tau,y(\tau)),

and define:

xλ(t)=y(Cλ(t)),x_\lambda(t)=y(C_\lambda(t)),

Then, by the chain rule:

dxλ(t)dt=Cλ′(t)uθ(Cλ(t),xλ(t)).\frac{dx_\lambda(t)}{dt} =C_\lambda'(t)u_\theta(C_\lambda(t),x_\lambda(t)).

This is exactly Equation 2.

Immediately after Equation 2, the paper gives the relation between the exact solutions,

xλ(t)=xI(Cλ(t)),x_\lambda(t)=x_I(C_\lambda(t)),

and their agreement at the endpoint,

xλ(1)=xI(1).x_\lambda(1)=x_I(1).

The equation thus follows from the chain rule applied to the change of time variable.

Comment 10

Question. What are DhclockD_h^{\mathrm{clock}} and DhgridD_h^{\mathrm{grid}}? Have you clearly distinguished between them?

Response 10

You will find both definitions in the paragraph immediately preceding Equation 3.

  • DhclockD_h^{\mathrm{clock}} is the numerical endpoint obtained by integrating the transformed dynamics in Equation 2 on the uniform physical-time grid.
  • DhgridD_h^{\mathrm{grid}} is the numerical endpoint obtained by integrating the original velocity field on the induced grid τn=Cλ(tn)\tau_n=C_\lambda(t_n).

They are endpoints of two finite-step implementations of the same continuous time transformation.

A nonlinear time transformation generally changes the internal Runge–Kutta stages and their quadrature weights. This accounts for the difference.

This is the central issue studied in Section 3.1.

Comment 11

Question. What does hh denote on the physical-time grid?

Response 11

Here, hh is the step size on the uniform physical-time grid, standard notation in numerical ODE integration. For KK equal intervals over normalized time,

\qquad n=0,\ldots,K.$$ After the time transformation, the corresponding induced grid is: $$\tau_n=C_\lambda(t_n).$$ You are asking for clarification of standard numerical notation. The uniform grid, the numerical endpoint $D_h$, and the limit $h\to0$ establish the meaning of $h$. ##### Comment 12 > **Question.** What is $C_\phi(t)$? Have you defined it? **Response 12** You will find the definitions of the time transformations immediately before Equation 2. First: $$I(t)=t,$$ where $I$ denotes the identity time transformation. $C_\phi(t)$ denotes **a learned monotone time-transformation function with fixed endpoints**. ##### Comment 13 > **Question.** Likewise, what is $C_\lambda(t)$? **Response 13** For your question on interpolation, the definition is $$C_\lambda(t) =(1-\lambda)t+\lambda C_\phi(t), \qquad \lambda\in[0,1].$$ Here, $\lambda$ **sets the degree of interpolation between the identity and the learned transformation**. ##### Comment 14 > **Question.** What is a nonlinear clock? **Response 14** The nonlinear clock you ask about is **a non-affine time transformation**. It changes the mapping from the solver’s physical time to the original flow time. Time reparameterization and mesh redistribution are established concepts in numerical analysis. Section 2 cites the following work. - Lindberg (1977) - Estep (1995) - Carey and Dinh (1985) - Becker and Rannacher (2001) - Giles and Süli (2002) - Huang et al. (2026), *Adaptive Reparameterized Time (ART)* The paper gives the background and implementation. Your description of these concepts as entirely undefined is inaccurate. ##### Comment 15 > **Question.** I cannot understand Equation 3 at all because D has not been defined. **Response 15** Your assessment that D is undefined is inconsistent with the manuscript. The paragraph immediately before Equation 3 defines $D_h^{\mathrm{clock}}$ and $D_h^{\mathrm{grid}}$, as explained above. Equation 3 compares these endpoints, $$D_h^{\mathrm{clock}} \neq D_h^{\mathrm{grid}},$$ and: $$\left\| D_h^{\mathrm{clock}}-D_h^{\mathrm{grid}} \right\| \rightarrow 0, \qquad h\rightarrow0.$$ Your claim overlooks the text immediately before the equation.

In summary

The review primarily raises concerns about presentation and notation, but does not explain how these concerns support the lowest soundness rating. In particular, it does not identify a specific mathematical error, flaw in the experimental design, or central conclusion contradicted by the evidence.

We recognize that unclear presentation can hinder technical verification. However, it is important to distinguish difficulty verifying an argument as presented from an identified technical flaw. We would appreciate clarification on which of these underlies the soundness assessment.

We take the readability concerns seriously and have made targeted revisions to improve the clarity of the presentation. We address the specific questions about terminology, notation, and derivations in our point-by-point responses.

We respectfully request that you reconsider the soundness assessment and overall rating in light of these clarifications and the methodological and experimental evidence presented in the paper. If substantive technical concerns remain, please identify the relevant claims or results so that we can address them directly.

Reviewer_LymkOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary
  • The paper studies the temporal allocation during few-step inference in flow-based generative models, identifying how the quality can depend on how a fixed number of solver evaluations are distributed over time.
  • They propose that a non-uniform temporal measure (integration-interval widths) could be learned on a cheaper source model and be transferred 'zero-target-fit' to unseen receiver models without changing receiver weights or inference NFE. This would essentially amortize the cost of the width search on the initial search model across the receivers.
  • The authors also provide a conditional asymptotic motivation for why non-uniform temporal allocation can be beneficial, while the actual finite-NFE intervals are learned through optimisation from a terminal quality objective.
  • The approach is validated across multiple receiver models, solvers, NFE budgets, acquisition methods and domains/tasks. The results of which suggest that temporal allocation can transfer across compatible receivers, but it is not universal - i.e. different domains, objectives, solvers and budgets can favour different temporal allocations.
Critical strengths

Strength 1 - the work contains extensive and well-targeted experimental validation. The method is evaluated across multiple receiver models, solvers, NFE budgets, acquisition methods, domains/tasks and metrics, including negative results showing where transfer fails. This establishes the robustness and boundaries of the claim and method. The experiments also support the proposed finite-step interpretation: clock/grid differences generally contract as NFE increases, as predicted by the numerical argument in Section 3.1.

Strength 2 - the method is simple and practical for post-training intervention. The method outputs a small set of temporal interval widths that can be acquired on a cheaper source model and transferred zero-target-fit to compatible receivers without requiring retraining or additional inference NFEs. Therefore deploying this method is straightforward as a one-off acquisition cost to be amortized across related receivers, while improving the output quality at the same inference budget.

Critical weaknesses

Weakness 1 - The scope of transfer is contract dependent and the compatibility boundary is not clearly defined. The strongest transfer results are between compatible flow-model receivers. Change in the domain, objective, solver or NFE budget can require a new source-side acquisition. For example the speech, text and forecasting experiments do not reuse the image-derived measure, requiring each domain/task to acquire their own measure. This could limit how broadly one can interpret the claim of "acquire once, deploy many". The paper does not provide a clear criterion to determine when a new receiver is sufficiently compatible with the source for transfer to be expected to work, and when reacquisition is required. The method is particularly attractive in scenarios when many compatible receivers are available over which the one-off cost can be amortizes, rather than as a universally reusable schedule. However it is unclear as to what defines a 'compatible receiver'.

Weakness 2 - The practical advantage over simpler temporal allocations are not always clear. The magnitude of improvement varies substantially across tasks: ETTm1 forecasting improves CRPS by only about 0.16%-0.29%, while the Lorenz improves by roughly 19%-22%. A simpler one-parameter power schedule retains approximately 83.7-91.8% of the gain of the flexible four-interval measure across the main image receivers. Therefore it is not always clear as to when the additional flexibility of the four-interval measure is practically worthwhile relative to the much simpler one-parameter temporal family. A clearer total-compute/amortization analysis, including when the one-off source acquisition becomes worthwhile relative to simpler alternatives would make the practical significance easier to assess.

Questions for authors

What determines whether a source and receiver are compatible enough for the same temporal measure to transfer successfully? The main image experiments reuse a CIFAR-10-acquired measure across different ImageNet receiver models, whereas the non-image experiments acquire a new measure for each domain/task. Is there a practical way to determine when an existing measure should transfer and when reacquisition is needed?

3 comments

Question. What determines whether the same temporal measure transfers from a source to a receiver? The image experiments reuse the CIFAR-10 measure across ImageNet models. The non-image experiments acquire a new measure for each domain or task. How should we decide whether to transfer or reacquire?

Response 1

Our claim is that multiple target models may reuse a source schedule under validated task and numerical conditions. We do not claim transfer across all models and tasks. We will clarify this in the paper.

We distinguish operator admissibility from goal alignment.

A. Operator Admissibility

The source and target must share the same noise-to-data evolution semantics, endpoint conventions, solver structure, and inference budget. These checks require no target-model queries.

Operator admissibility is necessary for correct deployment. It does not guarantee better generation quality.

B. Goal Alignment

The source schedule must remain useful for the target model’s terminal objective. A shared architecture or dataset alone does not establish compatibility.

Our CIFAR-10-to-ImageNet experiments validate transfer across datasets, resolutions, and training configurations within image generation.

Speech, text, and forecasting change the data representation or terminal objective. We therefore acquire a measure for each new domain or task. These experiments test reuse of the acquisition and deployment procedure. They do not establish that the image interval widths transfer to other domains.

C. Practical compatibility assessment and deployment decisions

We conducted prospective validation experiments and used them to define three deployment steps.

(a) Check operator admissibility.

Check the time-evolution semantics, endpoint conventions, solver structure, and NFE budget. The solver name alone is insufficient. Its time definition, integration rule, and implementation matter. Do not directly reuse the measure outside its validated range.

(b) Conduct pilot statistical validation.

For each configuration outside the validated range that meets the basic operator conditions, test candidate schedules on 512 paired pilot samples. Use a fixed projected terminal-risk metric before independent FID confirmation.

Accept transfer only if the bootstrap confidence interval for improvement lies entirely above zero after correction for multiple comparisons. Otherwise, use uniform allocation for now. Screening leaves all interval widths fixed.

We tested seven configurations across four checkpoints.

  • For the original source measure, the pilot and independent evaluation agreed on the direction of improvement in six configurations.
  • Two configurations passed screening. Both improved FID in the independent evaluation.
  • Self-Flow at NFE = 16 improved in the final evaluation but did not pass screening.
  • On LightningDiT, the four-interval schedule improved the FID point estimate. The confidence interval under the primary reference was inconclusive, so this configuration did not pass.

The pilot provides empirical evidence for transfer. It does not guarantee success. A failed pilot means the evidence for transfer is insufficient. Transfer might still succeed.

(c) Decide whether to transfer, fall back, or reacquire.

The first two steps lead to three choices.

  1. Direct transfer. If the operator conditions hold and the pilot passes, reuse the frozen source measure without target-side optimization.
  2. Temporary fallback. If the operator conditions hold but the pilot fails, use uniform allocation for now. A failed pilot alone does not establish incompatibility.
  3. Reacquisition. If changes in the data distribution, terminal objective, or solver make the measure unsuitable, or if uniform allocation falls short of the required quality, acquire a task-specific measure on a suitable low-cost source model and transfer it to the target.

Changing the target model alone does not require reacquisition. Reacquisition is warranted when the existing measure no longer applies and the fallback fails to meet task requirements. Acquisition still takes place on the source model, followed by deployment on the target.

The optional pilot adds target-side validation and selection costs. Count these separately. The original zero-target-fitting experiment used no target information to select or modify the measure.

Three-step deployment decision table

Decision stepCriterionAction
1. Operator admissibilityWhether the time-evolution semantics, endpoint conventions, solver structure, and NFE satisfy the validated applicability conditionsIf not, do not transfer directly
2. Pilot validationWhether the multiplicity-adjusted Bootstrap improvement confidence interval lies entirely above zeroIf it passes, accept transfer; otherwise, temporarily fall back to uniform allocation
3. Final deployment decisionWhether the existing measure meets task requirements and whether uniform allocation reaches the required performance levelReuse or fall back if adequate; otherwise, reacquire a task-specific temporal measure

Compatibility determines whether to try transfer. Pilot evidence determines whether to accept it. Task requirements determine whether to reacquire. These decisions are related but distinct.

Comment 2. The practical advantage over simpler time-allocation schemes

Weakness. The practical advantage over simpler allocations is unclear.

Forecasting CRPS improves by about 0.16%–0.29% on ETTm1 and 19%–22% on Lorenz. On the main image receivers, a single-parameter power schedule retains about 83.7%–91.8% of the four-interval gain. When is the extra flexibility worthwhile? A clearer account of total cost and amortization would help establish when one-time source acquisition is worth the cost compared with simpler alternatives.

Response 2

We agree that the comparison should distinguish three schedules. One is fixed in advance. The other two are acquired on the source model, using either a single power parameter or four flexible intervals.

The single-parameter power schedule is a restricted case of our acquisition framework. Our contribution does not require four intervals to outperform the power schedule on every target. Their added value depends on the model and task.

A. The four-interval schedule provides repeatable additional gains on some target models.

On Self-Flow, the original and independently reacquired four-interval schedules lower FID by a further 0.5585 and 1.0735, respectively, relative to the acquired power schedule.

The respective paired bootstrap 95% confidence intervals are

[0.4626, 0.6627][0.4626,\ 0.6627]

[0.8566, 1.2809][0.8566,\ 1.2809]

Both intervals exclude zero, supporting an additional gain from four intervals on Self-Flow.

B. The four-interval schedule is not advantageous on every task.

The added LightningDiT checkpoint experiment gave the following results.

MethodFID10k ↓
Uniform9.5334
Four-interval9.3298
Acquired Power9.1874

The power schedule performed better here. For the four-interval schedule, the confidence interval for improvement under the primary reference includes zero.

Forecasting CRPS improved by about 0.16%–0.29% on ETTm1 and 19%–22% on Lorenz. In the added paired comparisons, four intervals outperformed the power schedule for one of three ETTm1 seeds and all three Lorenz seeds.

The added value depends on the task and receiver. A single power parameter retains most of the gain. Four intervals improve performance further on some receivers.

C. Source-side acquisition cost comparison

We timed three runs on the same hardware, each in a fresh process.

Starting from a pretrained base checkpoint, we included source adaptation, preparation of the validation metric, and acquisition. The table gives median times.

Acquisition schemeMedian computation time
Flexible four-interval scheme248.9 seconds
Acquired power-schedule scheme242.2 seconds

Both searches used 131,072 source model–sample function evaluations. Neither increased inference NFE.

The timings exclude base-model pretraining.

Direct acquisition on SiT took 1,357.2 seconds in the recorded run. This measures cost alone and does not establish equal generation quality.

D. Amortizing computational costs when deploying to multiple target models

For MM compatible targets with matched generation costs, source acquisition has a lower total cost when

Asource+∑r=1MVr<∑r=1MAtarget,r{ A_{\mathrm{source}}+ \sum_{r=1}^{M}V_r < \sum_{r=1}^{M}A_{\mathrm{target},r} }

The terms in this inequality are defined below.

  • AsourceA_{\mathrm{source}} is the one-time cost of acquisition on the source model.
  • VrV_r is the optional validation cost for target model rr.
  • Atarget,rA_{\mathrm{target},r} is the cost of independent acquisition on target model rr.
  • MM is the number of compatible target models.

In the original deployment procedure, selection uses no target-side information, so Vr=0V_r=0. When the optional target-side pilot is used, its validation cost must be included.

At NFE = 8, our four-schedule pilot costs 16,384 logical model–sample function evaluations per configuration. Include this cost whenever the pilot is used.

Amortization alone cannot make source acquisition cheaper than a fixed, unfitted schedule, which has no acquisition cost.

Compared with a fixed schedule, source acquisition trades a one-time cost for better generation quality. Compared with separate target searches, it spreads that cost across compatible targets.

We report all source acquisition results for budgets of 8, 16, 32, and 64 candidates. Target results did not determine the search budget.

The choice depends on the task. The simpler power schedule retains most of the transferable gain. Four intervals are worthwhile when the extra quality justifies the slightly higher source search cost.

Reviewer_Mw2aOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an acquire-once, deploy-many approach to sampling schedules for few-step flow models. The central idea is to learn a temporal allocation on an inexpensive source model and reuse it on other receivers without target-side fitting. A temporal measure describes how a fixed evaluation budget is distributed over sampling time, and the allocation is optimized for terminal distributional quality without modifying receiver weights. Experiments report improved image quality when transferring from CIFAR-10 to ImageNet models and examine the limits of reuse across solvers and budgets. A separate forecasting study applies the same principle within individual tasks and reports gains in probabilistic prediction.

Critical strengths
  1. Amortized Temporal Allocation. The paper proposes formulating temporal scheduling as a positive measure acquired once on a source model and reused across receivers. This decouples schedule optimization from target deployment without requiring receiver-side gradient updates.

  2. Lightweight Deployment. Deploying the acquired measure adds no model parameters or auxiliary networks and requires no target-side trajectory fitting. It can be used directly to adjust the time grid of compatible flow model solvers.

Critical weaknesses
  1. Unclear benefit over a preset schedule.

Section 5.3 identifies "the late-loaded family" as the reproducible feature of the acquired schedules. The one-parameter power measure also retains most of the four-interval measure's gain across the three ImageNet receivers. As specified in Section 4, the exponent of this power schedule is selected by a search on the CIFAR-10 source model; it is not fixed in advance. The ImageNet comparisons omit a simple nonuniform schedule fixed before acquisition, such as a power schedule with a preset exponent. This leaves unclear how much source optimization adds beyond choosing a broadly useful allocation pattern in advance.

  1. Unclear quality tradeoff of avoiding target-side fitting.

The GITS-style and LD3 baselines adapted to the receiver use different fitting objectives, so they leave open the quality tradeoff between transferring the source measure and running the same terminal-objective search on a receiver. Without this comparison, it is difficult to assess what sampling quality is gained or sacrificed by avoiding receiver-specific acquisition.

Questions for authors
  1. On SiT using the midpoint solver at NFE = 8, how does a simple power schedule with its exponent fixed before acquisition compare with the Sobol/GP measure on the same paired streams?

  2. On the same receiver, how does the transferred source measure compare with a measure optimized directly on that receiver using the same terminal objective and a matched search budget? This would help quantify the quality tradeoff of avoiding receiver-specific acquisition.

2 comments

Comment 1. Comparison with a prespecified power schedule

Question. On SiT with the midpoint solver at NFE = 8, how does a fixed-exponent power schedule compare with the Sobol/GP measure on the same flow?

Response 1

We agree that the schedule acquired on the source model should be compared with a simple, unfitted nonuniform schedule. We therefore evaluated a quadratic schedule fixed before the experiment.

tj=(j/4)2t_j=(j/4)^2

This gives the following four interval widths.

[0.0625, 0.1875, 0.3125, 0.4375][0.0625,\ 0.1875,\ 0.3125,\ 0.4375]

We fixed the exponent at 2 before the experiment. We did not sweep exponents or tune the schedule.

All methods used the same SiT-XL/2 checkpoint, CFG scale of 1.5, explicit midpoint solver at NFE = 8, and 10,000 paired noise and class inputs.

Time allocationAcquisition methodADM FID10k ↓
Uniform allocationNone10.9824
Fixed quadratic schedule (γ = 2)None10.2584
Four-interval scheduleSource-model acquisition only9.8544

The fixed quadratic schedule improves generation quality. The source-acquired four-interval schedule lowers FID by a further 0.4041.

The paired bootstrap 95% confidence interval is [0.3076, 0.4951]. After Bonferroni correction for the two primary comparisons, it is [0.2947, 0.5076]. Both intervals exclude zero.

We held the checkpoint, schedules, generation streams, and reference statistics fixed when calculating these intervals.

Thus, the prespecified nonuniform schedule helps, and source-side optimization improves on this unfitted baseline.

Our contribution is to use a low-cost source model to acquire a four-interval schedule that improves generation quality and transfers directly to the target model. Deployment requires no target-model fitting and uses the same inference budget (NFE = 8).

Comment 2. Comparison with direct acquisition on the target model

Question. On the same receiver, how does the transferred measure compare with one optimized directly on the receiver using the same terminal objective and search budget? This would quantify the trade-off.

Response 2

To compare source transfer with direct target optimization, we ran an additional acquisition experiment on SiT.

We used the same four-interval parameterization, Sobol/GP search, and 64-dimensional projected terminal Fréchet objective as in source acquisition. We used real-data feature statistics from the target dataset.

Each search evaluated 128 candidates on 128 calibration samples at NFE = 8, for 131,072 logical model–sample function evaluations.

We compared the frozen source schedule with three schedules acquired independently on SiT using different seeds. We used three fresh paired generation streams of 10,000 samples each and one fixed SiT checkpoint. All schedules were frozen before the final FID evaluation.

Frozen time-allocation schedulePrimary ADM reference statistics, FID30k ↓Secondary validation50k reference statistics, FID30k ↓
Source-acquired schedule7.51714.4383
Mean of direct target-model acquisitions7.85304.5837

For the secondary evaluation, we changed only the real-data reference statistics. The generated images and ADM features were the same.

With the primary ADM reference, the three target searches had a mean FID of 7.8530. The source schedule scored 7.5171.

With the secondary reference, mean target FID was 4.5837 and source FID was 4.4383. One target schedule performed slightly better than the source schedule.

Under these conditions, direct target search did not consistently improve generation quality. The source schedule remained competitive without a separate search for each target.

Each target search took about 21.6 minutes. Repeating the source search took 45.5 seconds with an already adapted source model.

Once frozen, the source schedule requires no target-side acquisition calls and retains the NFE = 8 inference budget.

These experiments support acquiring a schedule once on the source and transferring it to compatible targets. Generation quality remains competitive without repeated target-specific optimization. We do not claim that source acquisition always outperforms direct target optimization.

Reviewer_Mw2aOpenReview ↗

Response

Thank you for running both comparisons I asked for. On SiT at NFE = 8, the fixed quadratic schedule already improves on uniform sampling, and the source-acquired schedule brings FID10k down further, from 10.2584 to 9.8544. This answers my first question about what source-side optimization adds beyond choosing a simple nonuniform schedule in advance.

The matched-budget comparison also helped me judge the cost of skipping target-side fitting. The transferred schedule achieved a FID30k of 7.5171, compared with 7.8483 on average across five independent searches on SiT. The confidence interval for the difference includes zero, so my takeaway is that the transferred schedule was competitive at this budget while saving a separate search on the target. These experiments address my two main concerns, and I am raising my rating from 2 to 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.