acceptodds
Under review as a conference paper at ICLR 2027

POST-TRAIN THE MAP YOU SOLVE: DEPLOYMENT-ALIGNED SPECIALIZATION FOR FEW-STEP FLOW SAMPLING

Abstract

Flow matching trains a continuous velocity field, but few-step sampling deploys a solver- and budget-specific finite map. We post-train that exact map by differentiating a terminal distribution objective through the deployed solver. A shared field–clock repair is followed by tiny solver-specific residual clocks, activated only when held-out target alignment predicts a gain. At 12 field evaluations on a public conditional-flow receiver trained with C²OT (conditional optimal transport), shared repair lowers mean CleanFID10k from 46.824 to 7.704; a 120-parameter residual package reaches 7.334 and is favorable in 10/10 retrainings for each active solver. On the official Scalable Interpolant Transformer (SiT) checkpoint, a frozen finite-stage operator improves Heun and fourth-order Runge–Kutta in every paired 50,000-sample ImageNet Fréchet Inception Distance (FID) evaluation and transfers without target-budget fitting to 8 and 24 field evaluations. B-series analysis and measured error orders trace these gains to tableau-dependent finite-step error. The alignment rule also predicts held-out directions on a second architecture and non-image receivers, while an alternative objective and independent feature representations corroborate the effect. The method keeps one field across solvers and adds no field evaluations; held-out evidence determines which residual clocks are stored.

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_vr7vOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper post-trains the exact few-step sampling map of a flow-matching model for a given solver and NFE. In Stage I, a terminal distribution loss is differentiated through the deployed solver; the loss is a Fréchet distance on projected Inception pool-3 features, plus FM anchor terms. This updates a late subset of the network and a 16-logit monotone time reparameterization ("clock") shared across Heun, RK3, and RK4. In Stage II, the network is frozen, small solver-specific residual clocks are learned, and a held-out alignment statistic decides whether each one is deployed. On a C2OT CIFAR-10 model at NFE=12, mean CleanFID10k drops from 46.8 to 7.70 after Stage I and to 7.33 after Stage II. On SiT-XL/2, a finite-stage temporal operator improves Heun and RK4 FID50k by 0.56 and 0.14. Further experiments cover B-series analysis, transfer across budgets and RK2 tableaus, composition with STORK, and four non-image domains.

Critical strengths
  • Optimizing the deployed solver- and budget-specific map is well motivated. The method adds no NFE at deployment and keeps a single field across solvers.
  • The B-series analysis and the NFE-decay study (5.586% → 2.513% → 0.250%) support the claim that the gains come from finite-step error.
  • Statistical practice is careful (paired streams, multiple seeds, rules frozen before confirmation, 50k-sample FID), and the coverage is broad, including non-image domains.
Critical weaknesses
  • The Stage I objective is a Fréchet distance on CleanFID Inception features, essentially the same network as the evaluation metric. The headline C2OT gain (46.8 → 7.7) may therefore partly reflect metric fitting.
  • Quality is not competitive: BézierFlow reaches 6.19/5.08/3.96 FID50k at NFE=8/12/24, versus 20.56/9.43/5.49 for the proposed method (Sec. 5.9). Optimized schedules and bespoke solvers are not compared.
  • Stage II adds considerable machinery for small gains (4.8% on C2OT, none for Heun), and heavy non-standard terminology makes the paper hard to follow.
Questions for authors
  1. Can you report the main C2OT and SiT results with features independent of the training objective (e.g., DINOv2 FD and CLIP-FID for all rows)?
  2. How does the method compare with an optimized timestep schedule (Sabour et al., 2024) or bespoke solvers (Shaul et al., 2024) at the same NFE and a similar adaptation cost?
  3. On SiT, how much benefit is lost if Stage II is removed?
5 comments

Comment 1. Metric fitting in Stage I

Question. Stage I optimizes Fréchet distance in CleanFID Inception features, and final evaluation uses the same feature network. Does this cause metric fitting?

The C2OT reduction from 46.8 to 7.7 may partly reflect fitting to the evaluation metric. Please add independent metrics such as DINOv2 FD and CLIP-FID.

Response 1

Thank you for raising the possible dependence between the training objective and evaluation metric.

We added DINOv2-L/14 FD and CLIP ViT-B/32 Fréchet distance to the main C2OT and SiT experiments, using the same saved images as the Inception evaluation.

The C2OT NFE12 results are below.

Solver / ArmInception FID50kDINOv2 FDCLIP-FID
Heun Original60.7317449.933018.0311
Heun Shared / Selected5.2044276.09424.6713
RK3 Original33.6879346.874912.2536
RK3 Shared6.4163268.09547.1348
RK3 Selected5.4275264.41085.8238
RK4 Original39.9393360.125114.0622
RK4 Shared5.2778260.10316.1451
RK4 Selected5.1480259.34055.8425

All shared-adaptation arms improve DINOv2 FD and CLIP-FID over the original model. Both active residual arms also improve these independent distances over their shared starting points.

The main gains therefore extend beyond the Inception feature space used in training.

We also tested training-feature and projection-dimension sensitivity using Inception, DINOv2, and CLIP features with 16-, 32-, and 64-dimensional projections.

All nine training configurations outperform the matched zero-fit control in the two feature spaces excluded from their training objective.

We will add these independent-feature evaluations and clearly state the training features, evaluation features, and sample counts for FID10k and FID50k.

Comment 2. Generation quality and external baselines

Question. Generation quality remains less competitive than methods such as BézierFlow. Comparisons with optimized timestep schedules and specialized solvers are missing.

Response 2

Thank you for identifying these gaps. We considered the results of BézierFlow, optimized schedules, and specialized solvers, then added controlled comparisons and mechanism analysis on CIFAR-10 and ImageNet-256.

The new experiments show that performance depends strongly on the training solver, deployment solver, and their interaction with the adapted velocity field. We address quality, solver interaction, and deployment applicability below. Code is available in the anonymous repository.

A. The original CIFAR-10 quality gap

We agree with your observation about the original Heun configuration. It did not achieve a better FID than BézierFlow at NFE8, NFE12, or NFE24.

We then tested Midpoint training and Midpoint deployment with the same initial checkpoint, NFE12, training objective, model capacity, and 64 optimization updates.

MethodFID50k ↓DINOv2 FD ↓CLIP-FID ↓
BézierFlow Fixed Compatibility Artifact5.0717382.84165.1506
Ours: Midpoint Training → Midpoint Deployment4.0793 ± 0.0562374.32202.5748

All three distances improve.

  • Inception FID50k decreases from 5.0717 to 4.0793, about 19.6%.
  • DINOv2 FD decreases from 382.8416 to 374.3220.
  • CLIP-FID decreases from 5.1506 to 2.5748.

All three metrics improve in all nine paired conditions formed by three independent training seeds and three independent generation streams.

The corresponding Heun training and deployment results in the same CIFAR-10 diagnostic are below.

MethodFID50k ↓
BézierFlow Fixed Compatibility Artifact5.0717
Ours: Heun Training → Heun Deployment8.9208 ± 0.1920
Ours: Midpoint Training → Midpoint Deployment4.0793 ± 0.0562

The original gap is concentrated in the Heun configuration. It does not establish an overall lack of competitiveness for the adaptation framework on CIFAR-10.

B. Further comparison with BézierFlow

We directly compared against the BézierFlow SiT Compatibility Port at NFE12.

MethodADM-FID50k ↓DINOv2 FD ↓CLIP-FID ↓
BézierFlow SiT Compatibility Port3.184069182.9421663.379921
Ours: Shared Adaptation (RK4)2.853464168.8646493.274840

Our RK4 configuration achieves lower values on all three distances in this comparison.

We also checked different BézierFlow training configurations.

BézierFlow ConfigurationFID
Same trajectory, 125 updates3.2237
Same trajectory, 500 updates3.2243
Increased to 200 training pairs3.1513
Ours: Shared Adaptation (RK4)2.8535
Ours: Shared Model + Midpoint2.8420

Increasing BézierFlow updates from 125 to 500 on the same trajectory did not improve FID. Increasing the number of training pairs to 200 reduced FID, but it remained above Shared RK4.

Deploying the same shared model with Midpoint gives FID 2.8420.

C. Optimized schedules and specialized solvers

We added ImageNet-256 comparisons with training-free samplers, optimized schedules, and learned specialized solvers at NFE12.

MethodADM-FID50k ↓DINOv2 FD ↓CLIP-FID ↓
STORK Order-23.131723182.8701903.293903
STORK + Frozen Clock3.087859181.4417073.272954
Bespoke RK2 Full-port2.884906179.9712073.155823
Bespoke RK2 Resource-cut2.911586181.3701973.167371
Ours: Shared Adaptation (Heun)3.993210197.7557783.990408
Ours: Shared Adaptation (RK4)2.853464168.8646493.274840

Shared RK4 has lower ADM-FID50k and DINOv2 FD than both Bespoke configurations, while Bespoke has better CLIP-FID.

Shared RK4 also has lower ADM-FID50k and DINOv2 FD than STORK, with slightly higher CLIP-FID.

The supplement reports full results for Uniform Native Midpoint and the AYS MC128 Resource Port.

These results support competitiveness in the tested low-NFE ImageNet configurations.

D. Quality at similar adaptation cost

To address your request, we measured acquisition costs on ImageNet-256.

MethodAcquisition Cost (seconds)
Uniform Native Midpoint0
AYS-flow-grid Heun (MC128 Port)7,463.81
Bespoke RK2 Full-budget Port15,034.45
Bespoke RK2 Resource-cut Port9,571.74
Ours: Shared Adaptation (Stage I only)7,822.59
Ours: Full Pipeline (Stage I + Residual Candidates + Selection)9,574.85

The full procedure costs 9,574.85 seconds, only 3.11 seconds more than Bespoke Resource-cut's 9,571.74 seconds. At this similar cost, Shared RK4 has lower ADM-FID50k and DINOv2 FD, while Bespoke has better CLIP-FID.

Stage I alone costs 7,822.59 seconds. Stage II adds 1,752.26 seconds for candidate acquisition and selection.

In this SiT setting, the prespecified development selection rule chooses the exact Stage-I fallback for both Heun and RK4. Removing Stage II preserves final deployment quality and saves its acquisition cost.

E. Interaction between training and deployment solvers

We added a full 2×2 experiment on CIFAR-10, using Heun and Midpoint as both training and deployment solvers.

The table summarizes FID50k for all four combinations.

Training solverHeun deploymentMidpoint deployment
Heun8.9048 ± 0.19046.4433 ± 0.2545
Midpoint36.4151 ± 1.17834.0793 ± 0.0562

The benefit of a training solver depends on the deployment solver. Heun training gives lower FID with Heun deployment, while Midpoint training gives lower FID with Midpoint deployment.

The original CIFAR-10 Heun result therefore does not establish the framework's performance limit. The controlled Midpoint experiment achieves better quality, and the full cross experiment supports a substantial interaction between training and deployment solvers.

We will report FID50k, DINOv2 FD, and CLIP-FID for all four combinations in the supplement and account for the interaction when interpreting solver performance.

F. Method scope and contributions

BézierFlow focuses on absolute quality for few-step generation. Our work addresses adapting pretrained Flow Matching models to practical low-NFE deployment.

Our method has three main features.

(a) Shared field-and-clock adaptation

We jointly adjust part of the velocity field and a shared time parameterization to adapt finite-step generation without an external teacher.

(b) Reuse across configurations and budgets

On CIFAR-10, one shared adaptation costs 281.55 seconds and is reused across eight prespecified solver/budget configurations.

On ImageNet-256, the frozen finite-stage temporal operator acquired at NFE12 transfers directly to NFE8 and NFE24 without refitting.

(c) Conditional solver-specific refinement

Residual refinement is retained only when validation supports it. Otherwise, deployment returns exactly to the shared adaptation.

Our contribution is a deployment-aware post-training framework for pretrained Flow Matching models. Shared adaptation, reuse across configurations, and optional specialization improve low-step deployment. We do not claim superiority to BézierFlow across all datasets and NFE settings.

Comment 3. Stage-II complexity and benefit

Question. Stage II adds complexity for a small gain. How much performance is lost if Stage II is removed on SiT?

Response 3

Thank you for this question.

A. Direct comparison on SiT-XL/2 ImageNet-256

SolverStage-I FIDRaw Residual FIDFinal selection
Heun3.9932104.000501Stage-I Fallback
RK42.8534642.855850Stage-I Fallback

Neither raw residual candidate improves on Stage I. The development selection rule, fixed before formal evaluation, chooses the exact Stage-I fallback for both solvers.

Thus, removing Stage II causes no deployment-quality loss in this SiT setting.

Removing residual training and selection reduces the shared acquisition cost from 9,574.851 to 7,822.585 seconds, saving 1,752.266 seconds. This is a one-time shared acquisition and selection saving. Inference speed is unaffected, and the saving is not counted separately for each solver.

B. Larger, repeatable gains on C2OT–RK3

The table shows NFE12 FID50k results on the original paired generation stream.

Stage-I modelStage IWith fixed Stage IIRelative FID reduction
Original selection model6.41635.427515.41%
Additional model 16.32714.996821.03%
Additional model 25.91794.708520.44%

We will clarify that Stage I is the core of the method, and Stage II is optional solver-specific refinement. Its value depends on the model, solver, and deployment budget. Extra residual parameters do not guarantee better quality.

Comment 4. Nonstandard terminology

Question. Too many nonstandard terms make the method hard to understand.

Response 4

Thank you for pointing this out. We will use consistent terminology across the Abstract, Introduction, Methods, tables, and figure captions, and reduce repeated descriptions of the framework while retaining the mathematical definitions and reproducibility details. Please see the updated PDF.

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

This paper proposes a deployment aligned post-training method that repairs the exact few step solver budget sampling map of flow models via shared field clock adaptation and selective residual clocks, improving low NFE generation without adding field evaluations.

Critical strengths

1.The paper is well-structured and presents a comprehensive methodological framework. 2.Extensive theoretical derivations and analyses support the proposed method.

Critical weaknesses
  1. The full two-stage pipeline is mainly validated on CIFAR-10 receivers, while ImageNet SiT experiments only evaluate a frozen finite-stage temporal operator.
  2. The experiments mainly rely on internal controls and lack sufficient matched comparisons against representative few step sampling and solver optimization baselines.
  3. The introduction is overly compressed and fails to clearly convey the core motivation and intuition.
  4. The first figure in the appendix, which presents qualitative sample comparisons under fixed noises, is blurry and difficult to interpret.
Questions for authors

See the Weakness.

6 comments

Comment 1. Full two-stage validation on ImageNet SiT

Question. The ImageNet SiT experiments validate only the frozen temporal operator. The full two-stage procedure is mainly tested on CIFAR-10 receivers, so the current evidence does not establish transfer of the full method to ImageNet.

Response 1

Thank you for identifying this gap. We added full experiments on SiT-XL/2 ImageNet-256.

Stage I updates 28,222,880 prespecified field parameters and 8 shared-clock logits for 192 updates. From the same Stage-I model, we then construct Continued Shared, Solver-specific Residual, and Shared-capacity Control.

The residual and shared-capacity controls each have 14 extra clock degrees of freedom, with different function families.

At NFE12 and CFG1.5, every condition uses 50,000 images from the same paired generation stream. Each cell gives ADM-FID50k / DINOv2 FD / CLIP-FID.

MethodHeunRK4
Original Uniform5.660788 / 233.418433 / 4.7513163.626882 / 195.753637 / 3.665255
Frozen Finite-stage5.012748 / 223.237866 / 4.4152693.443654 / 192.274793 / 3.551778
Shared Adaptation3.993210 / 197.755778 / 3.9904082.853464 / 168.864649 / 3.274840
Continued Shared4.009560 / 198.105315 / 3.9988502.866288 / 169.146239 / 3.280756
Raw Residual4.000501 / 198.031288 / 3.9931772.855850 / 168.951369 / 3.275201
Shared Capacity3.995885 / 197.673120 / 3.9918052.858884 / 168.908310 / 3.278103

Shared Adaptation substantially improves all three distances for both solvers. Raw Residual Candidates give no further improvement.

The experiment supports two conclusions.

  1. Stage-I field-and-clock adaptation improves low-NFE generation on ImageNet SiT.
  2. Stage-II residual refinement does not always add a gain for every architecture and solver.

We will present the original frozen-operator experiments and the new full field-adapted experiments separately.

Comment 2. Direct comparisons with few-step sampling and solver optimization

Question. Direct comparisons with representative few-step sampling and solver optimization methods are missing.

Response 2

Thank you for pointing out this gap. We added ImageNet-256 comparisons covering training-free samplers, optimized timestep schedules, and learned specialized solvers. The implementation is available in the anonymous code repository.

We also accounted for acquisition and adaptation costs consistently to compare methods at the same NFE and similar adaptation budgets. All main quality comparisons use NFE12 and the same initial pretrained checkpoint.

A. ImageNet-256 generation quality

The comparison includes the following methods.

  • Native sampler. Uniform Native Midpoint.
  • Optimized schedule. AYS MC128 Resource Port.
  • Training-free solver. STORK Order-2.
  • Solver with clock. STORK + Frozen Clock.
  • Learned solver. Bespoke RK2 Full-port and Resource-cut.
  • Specialized few-step method. BézierFlow SiT Compatibility Port.
  • Ours. Shared Adaptation with Heun and RK4.

Each condition uses 50,000 generated images. Lower is better for ADM-FID50k, DINOv2 FD, and CLIP-FID.

MethodADM-FID50k ↓DINOv2 FD ↓CLIP-FID ↓
BézierFlow SiT Compatibility Port3.184069182.9421663.379921
STORK Order-23.131723182.8701903.293903
STORK + Frozen Clock3.087859181.4417073.272954
Bespoke RK2 Full-port2.884906179.9712073.155823
Bespoke RK2 Resource-cut2.911586181.3701973.167371
Ours: Shared Adaptation (Heun)3.993210197.7557783.990408
Ours: Shared Adaptation (RK4)2.853464168.8646493.274840

Shared RK4 is competitive on ImageNet-256.

  1. Against BézierFlow. ADM-FID50k decreases from 3.184069 to 2.853464, DINOv2 FD from 182.942166 to 168.864649, and CLIP-FID from 3.379921 to 3.274840. All three distances are lower in this SiT compatibility-port comparison.
  2. Against Bespoke. Shared RK4 scores 2.853464 and 168.864649 on ADM-FID50k and DINOv2 FD. These are lower than Full-port's 2.884906 and 179.971207 and Resource-cut's 2.911586 and 181.370197. Bespoke has better CLIP-FID.
  3. Against STORK. Shared RK4 has lower ADM-FID50k and DINOv2 FD than both STORK configurations, with slightly higher CLIP-FID.

These results support competitive performance across multiple metrics in this low-NFE ImageNet deployment setting.

B. Acquisition and adaptation costs

We measured ImageNet-256 acquisition costs to compare methods at similar adaptation budgets.

MethodAcquisition Cost (seconds)
Uniform Native Midpoint0
AYS-flow-grid Heun (MC128 Port)7,463.81
Bespoke RK2 Full-budget Port15,034.45
Bespoke RK2 Resource-cut Port9,571.74
Ours: Shared Adaptation (Stage I only)7,822.59
Ours: Full Pipeline (Stage I + Residual Candidates + Selection)9,574.85

The zero cost for Uniform Native Midpoint means no extra model adaptation. Inference still requires computation.

  • Stage-I Shared Adaptation costs 7,822.59 seconds, similar to the AYS MC128 Port's 7,463.81 seconds.
  • Stage I costs less than Bespoke Resource-cut's 9,571.74 seconds.
  • The full two-stage procedure costs 9,574.85 seconds, less than Bespoke Full-budget's 15,034.45 seconds.
  • The full procedure and Bespoke Resource-cut differ by only 3.11 seconds, or about 0.0325%.

At this nearly equal measured acquisition cost, Shared RK4 has lower ADM-FID50k and DINOv2 FD, while Bespoke has better CLIP-FID. This compares both quality and adaptation cost.

C. Cross-budget transfer and reuse across configurations

Our method also addresses reuse of an adapted pretrained model across solvers and budgets.

On CIFAR-10, one shared adaptation costs 281.55 seconds and is reused across eight prespecified solver/budget configurations. Adaptation uses Heun NFE8/16. Evaluation includes NFE12, NFE24, and other prespecified solvers without acquiring a new shared model for each configuration.

On ImageNet-256 SiT, we acquire the finite-stage temporal operator at NFE12, freeze it, and apply it directly at NFE8 and NFE24 without refitting or selecting it for those budgets.

For Heun NFE8, ADM-FID50k decreases from the Continuous Control's 14.271673 to 11.947994. At NFE24, it decreases from 2.616144 to 2.562640. For RK4 NFE8, it decreases from 7.610743 to 7.471832. The RK4 NFE24 gain is 0.000362.

The learned time parameterization thus transfers to multiple NFE settings without further adaptation to the target budget.

Comment 3. Introduction and method intuition

Question. The Introduction is too compressed and does not adequately explain the motivation or method intuition.

Response 3

We reorganized the Introduction around the gap between continuous-time models and finite-step deployment before introducing the optimization objectives and numerical analysis. Please see the updated PDF.

Comment 4. Blurred qualitative comparison

Question. The first fixed-noise qualitative comparison in the appendix is blurred and hard to read.

Response 4

Thank you. We are temporarily unable to retrieve the generated image due to device issues. We will remake and replace the figure later.

Dear Reviewer,

The discussion deadline is approaching. Please review the authors’ responses and clearly indicate whether they adequately address your concerns. Kindly update your score if appropriate.

Best regards,

AC

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

This paper addresses the mismatch between continuous-time flow-matching training and few-step deployment by directly optimizing the discrete solver–budget map. Stage I jointly adapts a restricted subset of velocity-field parameters and a shared monotone clock across solvers using a terminal distribution objective with flow-matching regularization. Stage II freezes the field and learns lightweight solver-specific clock refinements. A held-out alignment statistic is proposed to guide their deployment, without additional field evaluations at inference. On a public C2OT-trained class-conditional CIFAR-10 model at NFE12, mean CleanFID10k across Heun, RK3, and RK4 decreases from 46.824 to 7.704 after shared adaptation and to 7.334 with selected refinements. A separate finite-stage temporal operator improves NFE12 ADM-FID50k over grid and continuous-clock controls on frozen SiT-XL/2 for both Heun and RK4. Further analyses examine solver dependence, residual capacity, diminishing gains at higher NFE, and transfer limitations.

Critical strengths

1. A practically useful deployment-aligned design supported by low-NFE results.

The paper directly optimizes the discrete solver–budget map used at inference. Its two-stage design combines shared field–clock adaptation with lightweight solver-specific clock refinement, allowing multiple solvers to reuse one adapted velocity field without additional field evaluations. Table 1 demonstrates substantial improvements from shared adaptation over the untouched C2OT-trained CIFAR-10 model at NFE12. Separately, Table 2 shows that a finite-stage temporal operator improves NFE12 ADM-FID50k over grid and continuous-clock controls for both Heun and RK4 on a frozen SiT-XL/2 checkpoint. These results support the practical value of adapting sampling behavior to the deployment configuration.

2. Informative analysis of solver-dependent calibration and its limitations.

The paper examines matched versus mismatched deployment solvers, residual capacity, equal-compute refinement, and diminishing gains at higher NFE. These experiments show that the gains from solver-specific adaptation and the preferred residual-clock capacity vary across solvers and receivers, and that retaining a shared clock can be preferable. The inclusion of negative results and independent-feature evaluations helps characterize the scope of the method. This analysis is consequential because it provides evidence about when specialization is useful, rather than relying solely on aggregate improvements.

Critical weaknesses

1. The incremental benefit of the selected residual-clock package could be isolated more convincingly.

Most of the headline C2OT improvement comes from shared field–clock adaptation. In the residual-capacity study, positive-capacity candidates receive additional refinement updates, whereas the zero-capacity action retains the Stage-I clock. A continued-shared-clock comparison under matched optimization budgets would help assess whether the selected (0,32,8) package offers gains beyond additional shared-clock optimization; isolating the effect of increased capacity would require an appropriate capacity control. The SB-CFM equal-compute control provides useful supporting evidence, but its FID gains are concentrated in RK3: Heun and RK4 worsen in FID on average, while the pooled FID50k improvement is 0.805% (supplementary Table 39). In the accompanying DINOv2 evaluation, however, conditioning improves FD in 8/9 cells and KID in 9/9 cells. Moreover, supplementary Section 30 explicitly limits the selected C2OT package's validation to one frozen receiver, with no selected-action FID50k or second-backbone confirmation. These limitations qualify the attribution and robustness of the selected residual package, rather than the demonstrated benefit of shared adaptation.

2. Some transfer and robustness statements need clearer distinctions between proxy and FID evidence.

Section 5.2 describes SiT cross-budget improvements as FID gains, but the reported values match the projected Fréchet-risk differences in supplementary Table 52, whose surrounding text refers to a future FID evaluation. Similarly, the 10/10 residual-retraining result in supplementary Table 46 measures proxy improvement on a fixed Stage-I receiver and evaluation stream, rather than repeated FID confirmation. The main text should identify these metrics and statistical units explicitly. These reporting issues limit the strength of the cross-budget transfer and robustness claims, while leaving the directly reported C2OT and SiT NFE12 FID improvements intact.

Questions for authors

1. Attribution of the selected residual package’s gains.

For the C2OT (0,32,8) package, how does refinement compare with continuing to optimize the shared clock under matched optimization budgets, batches, noises, and solver exposure? Is an appropriate capacity-controlled comparison available for this package to distinguish solver-specific parameterization from increased clock capacity? These controls would clarify how much of the incremental gain is attributable to specialization.

2. Validation scope and robustness of the selected package.

Supplementary Section 30 limits the selected package’s validation to one frozen C2OT receiver, while supplementary Table 46 reports 10/10 proxy improvements for each active solver on a fixed evaluation stream. Are FID-level comparisons available for this specific package across independently adapted Stage-I models or repeated refinement runs? Please distinguish refinement-seed robustness, Stage-I adaptation variability, and generation-stream variability when interpreting the 10/10 result.

3. SiT cross-budget metrics.

Section 5.2 describes the NFE8 and NFE24 improvements as FID gains, but the reported values match the projected Fréchet-risk differences in supplementary Table 52, whose surrounding text refers to a future FID evaluation. Are corresponding ADM-FID50k results available? If not, please clarify that this SiT cross-budget result demonstrates transfer in projected risk and revise the associated claims accordingly.

4 comments

Comment 1. Attribution of the residual-clock gain

Question. How do you show that the selected residual clock gains from solver-specific specialization rather than additional shared-clock optimization?

In the capacity experiment, positive-capacity candidates receive extra updates, while the zero-capacity candidate retains the Stage-I clock. The gain may therefore come from the extra optimization budget.

Please add two controls.

  • Continue optimizing the shared clock with the same update budget.
  • Control for capacity to distinguish added parameters from solver-specific parameterization.

Response 1

Thank you for raising this issue. We added Continued Shared Clock and Shared-capacity Control experiments. All controls start from the same frozen Stage-I model and use paired training batches, noise, and solver exposure.

We tested 48 updates per candidate, 96 updates across the two retained candidates, and 192 updates matching the original four-capacity experiment. The 192-update controls use the same four blocks of random numbers for 48 updates each, with continuous optimizer state.

The table shows the 192-update controls at NFE12 using 50,000 paired generated images. Each cell gives Inception FID / DINOv2 FD / CLIP-FID. Lower is better for all three.

SolverSelected ActionContinued Shared (192 updates)Shared Capacity (192 updates)
Heun5.2044 / 276.0942 / 4.67135.5399 / 274.5299 / 3.30625.7395 / 278.1588 / 3.3396
RK35.4275 / 264.4108 / 5.82385.6606 / 266.4934 / 6.24836.0504 / 269.9290 / 7.0170
RK45.1480 / 259.3405 / 5.84254.8606 / 259.4046 / 5.14624.8296 / 260.6999 / 4.5061

These results support an independent contribution from solver-specific specialization, especially for the RK3.

The selected C2OT residual package contains 40 active residual logits, giving 38 softmax degrees of freedom. We added a corresponding Shared39 capacity control, plus Shared32 and Common-K8 controls. Different solvers implement different finite-step numerical maps. Their local sensitivity directions and full responses to time reparameterization therefore differ, so the gain depends on the solver.

Comment 2. Validation scope and robustness of the selected residual package

Question. Do the 10/10 residual retraining results establish robustness across models and random seeds?

Supplementary Section 30 limits validation to one frozen C2OT receiver. The 10/10 results in Supplementary Table 46 are proxy improvements with a fixed Stage-I receiver and evaluation stream.

Please provide FID comparisons across independent Stage-I models or independent refinement runs. Distinguish refinement-seed robustness, Stage-I adaptation variability, and generation-stream variability.

Response 2

We completed additional FID50k experiments with three independent Stage-I models, ten residual refinement seeds, and three independent generation streams. All 105 experimental conditions and 315 distance measurements passed the full validation checks.

The fixed residual package now has formal confirmation across Stage-I models, independent refinement runs, and generation streams. atability across refinement runs and generation streams on the fixed Stage-I receiver.

A. FID50k across independent Stage-I models

We fixed the selected residual package and refinement seed, then evaluated three independent Stage-I models on three independent generation streams each.

The table reports the arithmetic mean of the three FID50k values. Each stream is evaluated separately. The images are not pooled into an FID150k evaluation.

Stage-I ModelRK3: Shared → Residual PackageRK4: Shared → Residual Package
20281951 (original selection model)6.4031 → 5.42845.2799 → 5.1531
20281952 (additional independent model)6.3536 → 5.03584.7986 → 4.7162
20281953 (additional independent model)5.9291 → 4.73704.5624 → 4.4295

The two additional models, 20281952 and 20281953, were not used to select residual capacity. They provide validation beyond the original selection model.

  • RK3 improves FID50k on all three Stage-I models.
  • RK4 also improves FID50k on all three models.
  • For both solvers, FID, DINOv2 FD, and CLIP-FID improve on every generation stream of every Stage-I model.

The gain thus transfers to two independently adapted Stage-I models beyond the model used for selection.

B. FID50k across ten independent refinement runs

We fixed the original Stage-I model, ran residual refinement independently ten times, and evaluated each refinement seed on three independent generation streams.

SolverMean FID reduction ± SD across refinement seedsSeed × stream conditions improving all three distances
RK30.958958 ± 0.07195130/30
RK40.137326 ± 0.02055230/30

FID reduction is shared FID minus residual-package FID, so a positive value means improvement.

For each refinement seed, we first averaged the FID differences across its three streams. We then calculated the sample standard deviation across the ten seed means.

Both RK3 and RK4 improve all three distances in all 30 seed × stream conditions. This provides FID-level evidence for repeatability across refinement runs and generation streams on the fixed Stage-I receiver.

C. Three sources of variability

(a) Stage-I adaptation variability

With the package and refinement seed fixed, both solvers improve FID50k on all three Stage-I models. All three distances improve on every stream. Two models were excluded from capacity selection, providing independent validation across Stage-I adaptations.

(b) Refinement-seed variability

With the original receiver fixed, ten independent refinement runs each improve all three metrics on three generation streams. Each solver achieves 30/30 improvements. The mean FID reduction and between-seed standard deviation quantify refinement-seed variability.

(c) Generation-stream variability

Both experiments use three independent generation streams. The observed improvements therefore extend beyond the earlier fixed-stream evaluation.

D. Revised interpretation of the original 10/10 result

We will revise Supplementary Table 46 and the related text. The original 10/10 result refers only to ten refinement-seed proxy improvements on a fixed Stage-I receiver and a fixed 2,048-sample evaluation stream.

The new experiments add the following evidence.

  • FID50k confirmation on three independent Stage-I models.
  • FID50k confirmation across ten independent refinement runs.
  • Repeated evaluation on three independent generation streams.
  • Improvements in FID, DINOv2 FD, and CLIP-FID for RK3 and RK4 in every corresponding condition.

We will report the original proxy result alongside the new model, refinement-seed, and generation-stream evaluations. These provide formal FID-level evidence of the added gain for RK3 and RK4.

Comment 3. Metric labels in the SiT cross-budget experiments

Question. Were the SiT cross-budget results incorrectly described as FID improvements?

The NFE8 and NFE24 values in Section 5.2 match the projected Fréchet risk differences in Supplementary Table 52. Please clarify the metrics and provide corresponding ADM-FID50k results.

Response 3

Thank you for identifying this reporting error. The projected Fréchet risk differences were incorrectly labeled as FID improvements. We corrected this in the updated PDF.

The original values are projected risk differences.

SolverNFE8NFE24
Heun0.37070.0165
RK40.03110.0054

They evaluate the proxy training objective. We also completed ADM-FID50k, DINOv2 FD, and CLIP-FID evaluations for 16 cross-budget conditions.

A. Heun

NFEBranchADM-FID50k ↓DINOv2 FD ↓CLIP-FID ↓
8Uniform14.690812357.7595078.502115
8Grid14.436746358.1019288.400083
8Continuous14.271673350.6696978.387267
8Finite-stage11.947994327.9906007.455960
24Uniform2.620245169.6184303.091858
24Grid2.628747170.0161433.088179
24Continuous2.616144169.2325693.085041
24Finite-stage2.562640167.7929273.030234

B. RK4

NFEBranchADM-FID50k ↓DINOv2 FD ↓CLIP-FID ↓
8Uniform7.769353272.3278435.687817
8Grid7.719096271.6101095.669960
8Continuous7.610743270.7984425.608978
8Finite-stage7.471832269.4937475.478290
24Uniform2.262377157.3530092.819068
24Grid2.253491157.1188742.820448
24Continuous2.254748157.1959332.819543
24Finite-stage2.253128157.1529082.813731

C. Representative comparisons

Solver / NFEControlControl ADM-FID50kFrozen Finite-stage ADM-FID50kFID Improvement
Heun / 8Continuous14.27167311.9479942.323679
Heun / 24Continuous2.6161442.5626400.053503
RK4 / 8Continuous7.6107437.4718320.138911
RK4 / 24Grid2.2534912.2531280.000362

These results show how the cross-budget gain varies with solver and budget.

We will clearly distinguish projected risk, ADM-FID50k, and independent-feature evaluation. We will retain the near-zero RK4 NFE24 result and the adverse changes in the relevant metrics.

Thank you for the detailed responses and the additional experiments.

Q2 and Q3 are resolved. The new FID50k results across Stage-I models (including two not used for capacity selection), refinement seeds, and generation streams address the robustness question. The corrected SiT metric labels, together with the new ADM-FID50k results, address the reporting issue. I appreciate that the near-zero RK4 NFE24 result will be retained.

For Q1, the new controls support solver-specific specialization for RK3, which improves on both controls in all three distances. For RK4, however, the selected residual is worse than both the continued-shared clock and the shared-capacity control in Inception FID (5.1480 vs. 4.8606 and 4.8296) and CLIP-FID (5.8425 vs. 5.1462 and 4.5061), while DINOv2 FD marginally favors the residual (259.34 vs. 259.40 and 260.70). Please state in the revision that, under these controls, the specialization benefit is supported for RK3 but not for RK4.

Please also reconcile the residual parameter count. The paper reports 120 active parameters (96 for RK3 and 24 for RK4), whereas the response states 40 active logits, and the Shared39 control appears to be sized from the latter. The paper’s figures coincide with the “Scalars” entries for K = 32 and K = 8 in Supplementary Table 44. Please clarify which count is correct and, if needed, correct the abstract, Section 3.3, and Supplementary Section 30.

These points qualify the residual-clock claims but do not change my overall assessment. I will keep my rating.

Comments on the paper

1 comment
Area_Chair_NgE7OpenReview ↗

Reminder: Discussion Deadline — November 18

Dear Reviewers,

Thank you for your valuable time and contributions to ICLR. This is a friendly reminder to review the authors' responses and engage in discussions with them. The discussion deadline is November 18.

Thank you for your continued efforts!

ACs

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.