acceptodds
Under review as a conference paper at ICLR 2027

Learning Inference Clocks for Finite-Step Generative ODEs

Abstract

Finite-step sampling from generative ODEs depends on where a solver evaluates the frozen vector field. We present Generative Inference Clock Optimization (GICO), an offline policy-optimization method for utility-specific allocation with the generator and solver rule fixed. Each clock is a normalized flow-time density whose inverse CDF yields a grid at the requested evaluation budget. A surrogate fitted to completed-solve utilities weights reference densities to anchor a policy conditioned on context, solver, and budget, and guides regularized refinement without further generator evaluations or gradients. An exact finite-budget adjoint identity explains utility sensitivity; smooth asymptotics yield an allocation surrogate. Experiments cover temporal sequences, molecular 3D coordinates, and images. Temporal errors decrease by up to 10.0% and Kabsch RMSD by 5.4–7.9% relative to uniform grids. Deterministic GICO attains the lowest mean FID among evaluated methods in 10/12 low-resolution image settings; its stochastic extension leads in the remaining two. FID reductions relative to fixed-reference envelopes reach 30.9%. With the full observation budget on frozen SANA, GICO policies outperform observation-matched Bayesian optimization and PPO-style clock policies on ImageReward and VQAScore at all tested NFEs. Held-out evaluations and ablations further support generalization across contexts and inference budgets without refitting.

Primary area:
generative models
Keywords:
Generative ODEs, Sampling Schedule Optimization, Offline Policy Optimization, Amortized Optimization

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_dAsWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Generative Inference Clock Optimization (GICO), an offline method for learning sampling-time allocations for frozen generative ODEs. It represents schedules as normalized densities, uses measured terminal utilities to train a surrogate, and learns an anchored clock policy conditioned on context, solver, and NFE. Experiments cover temporal, molecular, and image generation, including held-out-budget evaluations.

Critical strengths

The density representation separates allocation from node count, enabling a shared formulation across inference budgets. Utility-based supervision targets downstream quality directly, with theoretical analysis explaining how local errors affect terminal utility. Experiments span multiple domains and include representation, context, and reference-selection controls.

Critical weaknesses
  1. The surrogate determines reference weights, guides refinement, and supports checkpoint selection. Its calibration assumptions are not verified at the anchor and refined allocations. Accuracy on held-out reference clocks does not establish reliability at policy-generated clocks, where optimization may exploit prediction errors.
  2. For temporal and molecular tasks, each context is primarily assigned one reference clock plus a uniform anchor; ImageNet similarly assigns each class to one density. The surrogate subsequently scores the full reference pool for each setting. The paper does not sufficiently validate these unobserved context-clock predictions using a directly measured multi-clock evaluation panel.
  3. The pool is explicitly specified but remains manually configured. Its composition, exponents, reversals, and cross-model schedule mappings may affect both surrogate coverage and policy anchors.
  4. In the full text-to-image comparison, transfer combines 12,800 observations from source NFEs, whereas direct fitting uses 6,400 observations. Excluding the target NFE establishes a valid transfer setting, but superiority over direct fitting cannot be attributed solely to budget generalization.
  5. Its unbiased estimator provides a reasonable motivation, and evaluation with other metrics is useful. However, unbiasedness does not establish low variance or reliable clock rankings. The paper does not adequately assess supervision noise, ranking stability, or sensitivity to the choice of training metric.
  6. The distinction between distributions over whole clocks and densities allocating evaluations over time could also be presented more clearly.
Questions for authors
  1. Can you measure surrogate calibration and ranking accuracy at the weighted anchors and policy-generated clocks? How do these errors relate to realized utility gains, and how sensitive are the results to the tilt temperature and refinement coefficient?
  2. Please evaluate multiple reference clocks on the same held-out contexts or classes.
  3. What motivated the reference-pool composition and parameter choices?
  4. Please compare transfer and direct fitting using equal total observation budgets.
  5. Why was KID selected as the primary image supervision metric?
  6. Can you report repeated-estimation uncertainty, clock-ranking stability, and agreement with final FID rankings, or compare alternative supervision metrics under matched collection costs?
Reviewer_xAF1OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how to improve generation quality under finite NFEs by optimizing the inference time-allocation nodes. The authors propose GICO, which collects the terminal performances of reference clocks offline, trains a surrogate, computes the anchor, and subsequently refines a conditional policy using KL regularization. Experiments across temporal sequences, molecular structures, and image generation demonstrate solid performance gains in multiple settings.

Critical strengths

Strength 1: Optimizing inference step allocation without altering the pretrained model or solver is a direction of value, especially under low NFE constraints.

Strength 2: The paper covers diverse scenarios (including temporal sequences, 3D molecules, and images) and shows good performance gains across benchmarks.

Critical weaknesses

Weakness 1: The heavy theories in Sections 3–5 and Appendix A (such as adjoint analysis and asymptotic allocation rules) feels like Over-packaged because the actual algorithm does not use them. Formalisms like the forward-KL centroid are essentially just weighted arithmetic means. Part of the derivations might neither drive the algorithm nor guarantee its success, which means extensive math appears decoupled from the final method design.

Weakness 2: Given that prior work (e.g., CVPR 2026 James–Stein Scheduler) already learns conditional schedules via terminal rewards, GICO's core novelties rely on offline surrogate training and density parameterization. However, the paper fails to empirically decouple the independent contributions and prove the effectiveness of the method's core designs. More ablation studies should be included.

I am really open to raise my score if the authors can address my concerns.

Questions for authors

Q1. Since the actual algorithm does not use the derived adjoints or asymptotic densities, what value do these theories give beyond intuition?

Q2. Can you provide an ablation comparing an Anchor-only and No-KL baseline against Full GICO, to prove whether the policy refinement and KL regularization actually help?

Q3. The comparison with the James-Stein Scheduler changes both the optimization paradigm and the schedule representation. Can you isolate the factors to clarify where GICO's gains originate?

Q4. What is the total compute cost considering offline data collection, reward scoring, surrogate training, policy training? This could be a fair comparison with online baseline approaches.

Q5. Appendix G.7 shows that 12 reference clocks perform nearly as well as 25. Can you justify the necessity of using 25 clocks?

Q6. Why does the extra text conditioning in GICO-prompt frequently underperform GICO-global across some settings?

Q7. Given that text-to-image experiments rely on ImageReward and VQA score on both training and evaluation, can you report independent metrics like FID to rule out reward-model overfitting?

Q8. Why does focusing solely on VQA lead to a drop in the final VQA score?

Q9. Why does Bayesian Optimization collapse exactly to the result of uniform schedule on SANA?

1 comment
Area_Chair_pTACOpenReview ↗

please engage

Dear reviewers,

Please join the discussion with the authors.

Best, AC

Reviewer_oFx8OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GICO, an offline method for learning inference-time schedules for frozen generative ODEs. It represents schedules as normalized flow-time densities, fits a surrogate from utilities measured on reference schedules, and learns deterministic or stochastic policies conditioned on context, solver, and NFE. The paper also provides a finite-step error analysis and evaluates the method on temporal, molecular, and image generation tasks.

Critical strengths
  1. The density representation is clean and allows one policy to produce schedules for different NFE budgets, making the held-out-NFE experiments relevant.

  2. The evaluation is broad and carefully documented. In particular, the low-resolution image experiments show consistent improvements over fixed-reference envelopes.

Critical weaknesses
  1. The experiments do not clearly distinguish the value of schedule-ranking information from the effect of mixing candidate schedules. If schedule rankings are stable from calibration to test, selecting the best calibration schedule should be a strong baseline, and the paper should establish the additional benefit of density aggregation and refinement. If the rankings differ substantially while GICO still improves, the gain may instead come from regularization through averaging rather than from learning a transferable utility-based ranking. The paper does not report calibration-test rank correlations, the gap between the calibration-selected and test-oracle schedules, or systematic comparisons among top-1 selection, the weighted centroid, and refinement.

  2. The selection and evaluation principles are not fully aligned. On COCO, both GICO and the selected-fixed baseline are optimized using a balanced IR/VQA utility, but the test results report only raw IR and VQA rather than the scalarized objective actually used for selection. This permits metric trade-offs without establishing that GICO is better under one prespecified objective. Other tasks instead use metric-wise test reference envelopes, so the meaning of "outperforming fixed schedules" is not consistent across experiments.

  3. The evidence for prompt-dependent schedules is limited. Each prompt is observed with only one nonuniform reference schedule plus uniform, while the surrogate subsequently predicts utilities for all unobserved prompt-schedule pairs. No complete schedule-panel evaluation on held-out prompts is provided to validate these counterfactual rankings, and the prompt-conditioned policy does not consistently outperform the global policy.

Questions for authors
  1. Can the authors report the locked-test balanced utility for GICO and the selected-fixed baseline, together with the test-oracle best result over the same schedules and the calibration-test schedule-ranking correlation?

  2. Can the authors provide systematic comparisons of calibration top-1, surrogate top-1, the weighted density centroid without refinement, and the final refined policy under the same observations and test objective?

  3. Can the authors evaluate all reference schedules on a subset of held-out prompts to verify prompt-specific surrogate rankings and clarify whether prompt conditioning provides value beyond a global learned schedule?

1 comment
Area_Chair_pTACOpenReview ↗

please engage

Dear reviewers,

Please join the discussion with the authors.

Best, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.