acceptodds
Under review as a conference paper at ICLR 2027

Training-State Shaping for Low-NFE Flow Matching: Gains Across Solvers and Better Couplings for Distillation

Abstract

Flow matching trains on prescribed interpolation paths, while generation follows trajectories of the learned velocity field. We develop training-state shaping (TSS), a post-training approach that improves few-step sampling and the data efficiency of teacher-generated pairs. In pixel-space models that obtain velocity from clean-image prediction, we reshape training inputs using an exponential-moving-average model while retaining clean targets. We extend TSS to native velocity prediction by pairing each shaped path with its exact tangent. On ImageNet, TSS improves low-NFE generation across JiT model scales and solvers. With JiT-H/16 and DPM++2M, it achieves FIDs of and at 10 and 20 NFE, respectively. In latent space, deterministic reparameterization of training time explains most of the gain. With fixed pools of 200k pairs per source and matched student recipes, shaped-teacher data lets reflow and flow-map students reach target FID with fewer updates, equivalently fewer sample presentations at fixed batch size. We measure this advantage across multiple target FIDs and two teacher pairs. The original flow-map extensions reach similar one-call EMA quality at 100k updates. A separate pool-prefix study finds better student quality from fewer TSS-generated pairs than from a larger vanilla pool. An endpoint-FID-matched control shows that endpoint FID alone does not rank downstream data value.

Primary area:
generative models
Keywords:
Flow matching, Few-step generation, Generative model post-training, Training-state shaping, Time reparameterization, Synthetic training data

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_2n8hOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper suggests a post-training method (TSS) to better align training with the sampling trajectories. For JiT (x-prediction), TSS-X shifts the noisy training input slightly toward the EMA model's clean-image prediction, while keeping the true clean image as target. For velocity models, the path is bent and the model is trained on its exact derivative (full TSS), or on a pure time warp (TSS-T). The authors argue that this improves sample quality at low NFE and produces better (noise, sample) pairs for distillation, which train reflow and flow-map students in fewer updates.

Critical strengths
  • The gain in FID is real for low NFE

  • Experiments cover three model sizes, two solvers and three NFE budgets, although L/16 and H/16 are single runs

Critical weaknesses
  • Paper seems to be about exposure bias, which addresses the misalignment problem between training and sampling. It is a common problem in diffusion models in general, but the paper never mentions it or cites paper related to that. In addition, the main method (TSS-X) is only shown to work on JiT. The velocity version (TSS-T / full TSS) is a different method and only helps at 10 NFE on LightningDiT (Table 10). To be thoroughly tested, it should be applied to various diffusion models and parameterisation. At best it would work on different parameterisation, or then it would expose a failure mode of x-prediction that would then require an explanation on why their method works here and not else where. The field would then gain deeper understanding of JiT like that. Results on other models seem unlikely given that the TSS-X transfer to LightningDiT was negative (A.4.6), but that limitation should be made clear upfront. Also, TSS-T, a pure time warp, gives most of the latent gain, so schedule baselines (e.g. time-shift sampler, vanilla on the warped sampling grid) are necessary. The gains are modest, which is ok in my opinion if the paper is able to explain better its motivation and the failure modes, so that we have a better understanding of the mechanism.

  • A gain of 1 FID over 10-20 NFE is real but quite modest. Table 2 has no error bars and does not say which rows are multi-seed (L/16 and H/16 are single runs), so it is hard to really quantify how substantial the contribution is. Also the slight degradation at higher NFE is slightly worrying (within noise for Heun-50, but positive in all three seeds for DPM++2M at 50 NFE, Table 8). The method should in the worst case not degrade at higher NFE. I am willing to buy that there is no degradation if error bounds are provided.

  • The distillation gain can come from better targets or better couplings. Better targets is a weak argument: generating the pairs is a one-time cost, so saving a few NFE is not really worthwhile. In addition, vanilla at 14 NFE already matches TSS-X at 10 NFE (FID 8.11 vs 8.22, Table 26), so the whole procedure does not even buy 4 sampling steps. The natural baseline is vanilla at a much higher budget (e.g. 50-100 NFE), which is missing. Vanilla19 is used for the 47.4% cost saving claim, but its endpoint FID is never reported.

  • Better coupling is a solid hypothesis, and it is in the title, but it is never tested. The only evidence is the transport cost and cosine between noise and endpoint (Table 26), and the two do not agree on the ordering (Vanilla14). A cosine of 0.06-0.09 is close to orthogonal; a reference point is needed to know if it is meaningful (e.g. shuffled pairs vs. minibatch OT pairs, vanilla at higher NFE).

Questions for authors
  • How does TSS-X compare to exposure-bias methods (e.g. DDPM-IP, epsilon scaling, time-shift sampler) ?

  • How do students trained on vanilla pairs generated at 50-100 NFE compare?

  • Can you test the coupling hypothesis directly (e.g. the straightness measure of Liu et al. 2023, or cosine and transport cost compared to shuffled and minibatch OT pairs)?

  • Why does TSS-X transfer negatively to LightningDiT?

2 comments

Thank you for the exposure-bias references and the questions about downstream data value.

1. Exposure-bias methods and scope. We will expand the related work to include DDPM-IP, epsilon scaling, Time-Shift Sampler, and MixFlow. The training/sampling mismatch is established prior motivation. TSS-X's particular intervention uses a detached EMA clean-image estimate to choose a sample-dependent displacement, concentrated around an anchor, while retaining the clean-image target. This differs from perturbing inputs with additional noise, scaling predictions at inference, or changing inference query times.

We completed a common-environment FID50K evaluation of existing single-seed JiT MixFlow adaptations. All four arms use a clean-image head and 100-epoch continuations from the same epoch-600 source. DPM++2M results are:

MethodNFE10NFE20NFE50
Vanilla9.59727.19953.7298
TSS-X8.38876.08093.8053
MixFlow, native time sampling9.53466.68924.2198
MixFlow, Beta time sampling10.29937.36604.5415

TSS-X improves on both adaptations; vanilla remains best at NFE50. This tests two JiT adaptations, not official MixFlow reproduction or exhaustive tuning.

2. Effect sizes and high-NFE regression. We will identify B/16 as three paired seeds and L/H as single pairs. Paired Student-tt intervals from the original FID50K files give:

DPM++2M NFEB/16 mean TSS-X minus vanilla95% interval
10-1.3557[-1.7813, -0.9301]
20-1.0688[-1.2969, -0.8408]
50+0.1070[-0.0195, +0.2335]

All three differences favor TSS-X at 10/20 NFE and vanilla at 50 NFE. We retain the high-NFE regression. These three-seed intervals are not multiplicity corrected. H/16 improves from 7.6652 to 6.6524 at 10 NFE and 4.9841 to 4.2417 at 20 NFE (13.2% and 14.9%), with a +0.0584 regression at 50 NFE. The practical claim is low-budget sampling improvement.

3. Higher-quality vanilla pools and costs. The 19-call vanilla endpoint pool's measured FID is 7.0070, compared with 9.4154 for vanilla at 10 calls and 8.2179 for TSS-X at 10 calls. We will make this value explicit next to the student table.

Despite its better endpoint FID, the 19-call pool gives one-call reflow-student EMA FIDs 32.3649/31.8548, versus 29.7561/29.6124 from TSS-X in the two matched student streams. At 15 student calls, vanilla-19 instead wins: 13.2008/13.2715 versus 13.5930/13.5491. The 14-call, endpoint-FID-matched control also produces worse students than TSS-X in both streams. This establishes that endpoint FID alone does not predict downstream performance. We have not trained the proposed 50/100-call vanilla-pool students, so these controls do not establish superiority over such a source.

The efficiency measurements concern student updates and pool size. Teacher post-training and pool generation cost extra; fewer generation calls alone do not establish full-pipeline savings.

4. Coupling geometry and straightness. Your distinction is important. Endpoint distance and cosine do not establish straighter solver trajectories or isolate pairing from image quality. We subsequently measured 8,192 trajectories per model for each of three matched JiT seeds. At 50 DPM++2M calls, mean path-length/chord ratios are 1.0096 for vanilla and 1.0127 for TSS-X; midpoint normalized chord residuals are 0.0421 and 0.0500. Thus these diagnostics do not support straightening as the explanation. We will replace that interpretation with the measured downstream data-source result.

A within-class permutation or matching on the exact same endpoint/noise sets would isolate coupling more directly. Such a student comparison is still missing; neither the small endpoint cosines nor the scalar FID match substitutes for it.

5. Latent transfer. The negative TSS-X transfer jointly changes the prediction head, MSE-plus-cosine loss, denominator floor, time distribution and shaping profile; it does not isolate a cause. A new three-seed Euler-NFE10 control gives 13.9057 for frozen vanilla with exact inference-time reparameterization versus 13.9658 for TSS-T training. There is no demonstrated training advantage over that baseline at this point. We will distinguish these recipes and retain their high-NFE limitations.

I acknowledge and thank the authors for their response and will raise my score to 2.

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

This paper proposes Training-State Shaping (TSS), a post-training method for flow matching aimed at improving low-NFE sampling and the data efficiency of teacher-generated noise-image pairs. For pixel-space x-prediction models (JiT), TSS-X shifts training inputs toward a detached EMA prediction while keeping the clean-image target. For native velocity-prediction models (LightningDiT), the authors introduce a transport-consistent variant that modifies both the training path and its tangent target, along with an EMA-free time-reparameterization special case (TSS-T). Experiments on ImageNet test low-NFE generation across solvers and evaluate downstream reflow and flow-map students trained on teacher-generated pairs.

Critical strengths
  1. TSS-X provides consistent low-NFE improvements on JiT without additional inference calls. Table 2 shows gains across B/16, L/16, and H/16 with DPM++2M at 10 and 20 NFE, and with Heun at 19 NFE. The three paired B/16 runs strengthen this evidence.
  2. The downstream study includes useful controls, including shared student initialization, matched training recipes, and an endpoint-FID-matched teacher. The finding that slightly better teacher endpoint FID can still lead to worse students is useful. Longer training runs and negative results also help clarify the limits of the observed gains.
Critical weaknesses
  1. The full method’s benefit over simpler time changes is unclear. Most latent-space gains come from TSS-T, while adding EMA gives mixed results. The paper also lacks a comparison with applying the same time change only at inference. The main JiT gains come from TSS-X, but the theory covers a different training rule that has not been tested under the same JiT setup.
  2. The distillation gains depend strongly on the training setup. On the 200k pools, longer training largely closes the one-call gap, and two-call results favor vanilla. Direct endpoint fine-tuning also removes the advantage. Larger pools retain some gains, but benefits for stronger students remain unclear. Matching teacher FID also does not show that better noise–image pairing causes the gains, since the generated image distributions still differ.
Questions for authors
  1. How much does TSS-T gain over applying the same time change at inference without retraining? Please compare at the same NFE, changing both the model’s query times and the velocity scale.
  2. Why do state shaping and target correction work differently in JiT and LightningDiT? Does the full transport-consistent method improve JiT under the same training setup?
  3. Do students learn faster because of better noise–image pairing or differences in the generated images? Can you test this by keeping the image set fixed and changing only the pairing?
1 comment

Thank you for the specific questions. They help separate the training intervention, the inference schedule, and the source of the downstream benefit.

1. Exact inference-time reparameterization. The appropriate frozen-model baseline is dzdt=ϕ′(t)vvanilla(z,ϕ(t)),ϕ(t)=t+(1−a)sDb(t).\frac{dz}{dt}=\phi'(t)v_{\rm vanilla}(z,\phi(t)),\qquad \phi(t)=t+(1-a)s_D b(t). It changes both the model's query time and the velocity scale. Our SnormS_{\rm norm} experiment retains the original query times, so it does not answer this question.

We completed this control on three matched LightningDiT seeds at Euler NFE10, with the same 50,000 noise/class IDs, guidance, hardware and evaluator. The training profile is fixed without tuning; sD=1s_D=1. Mean FIDs are:

VanillaVanilla + inference transformTSS-T training
14.645813.905713.9658

The paired TSS-T-minus-inference difference is +0.0600+0.0600, with 95% interval [−0.4220,+0.5421][-0.4220,+0.5421]. We therefore find no additional training benefit at this operating point. The interval does not establish equivalence.

The existing full-variant evaluation gives 13.6358. Its paired difference from the new inference baseline is −0.2700-0.2700 [−0.8975,+0.3576][-0.8975,+0.3576], a post-hoc comparison after checking that the vanilla/TSS-T bridge results agree exactly. All three point differences favor full TSS, but the interval crosses zero. The previously reported −0.3300-0.3300 increment over TSS-T alone does not establish superiority over this stronger inference baseline. At DPM++2M NFE10, full minus TSS-T is −0.0327-0.0327 [−0.2880,+0.2226][-0.2880,+0.2226].

2. What the JiT and latent controls establish. TSS-X retains clean-image supervision on shifted inputs. If z~t=zt+δt\tilde z_t=z_t+\delta_t, its velocity target is u−δt/(1−t)u-\delta_t/(1-t), whereas the shaped-path tangent is u+δ˙tu+\dot\delta_t. The consistency proposition concerns the latter construction. We will state this distinction at the introduction of the methods.

The three-seed JiT D+N control also retains the clean-image objective. At 19-NFE Heun, vanilla/TSS-X/D+N FIDs are 8.2543/7.1112/7.2824; at 99 NFE they are 3.6851/3.7163/3.8556. This supports retaining the full TSS-X direction in JiT, but it is not a matched test of transport-consistent tangent supervision. We do not have a completed, verified full transport-consistent JiT comparison under that same protocol to report.

Likewise, the unsuccessful latent TSS-X transfer changes more than output parameterization: the native loss includes cosine supervision, and the denominator floor, time distribution and shaping profile differ. Those results establish recipe dependence; they do not identify the prediction head as the cause. The transport-consistent experiments supply a workable latent alternative with low-NFE gains; regressions at 20 NFE and Euler 50 NFE remain reported. DPM++2M at 50 NFE instead improves slightly.

3. Pairing versus the image marginal. Keeping the image set fixed while changing its pairing with noise is the right counterfactual for isolating coupling. Our existing endpoint-FID control does not hold that marginal fixed. It answers the narrower question of whether endpoint FID predicts downstream usefulness.

The 14-call vanilla pool has slightly better endpoint FID than the TSS-X 10-call pool (8.1094 versus 8.2179), yet its reflow students are worse in both matched training streams: one-call EMA FIDs 32.7024/32.1380 versus 29.7561/29.6124. The 19-call vanilla pool has still better endpoint FID, 7.0070. Its one-call students are worse than TSS-X's, while its 15-call students are better. These results show that endpoint FID alone does not rank data sources across consumers and sampling budgets; they do not isolate coupling as the cause.

A clean pairing intervention should preserve the exact image and noise sets, class labels, student initialization and budget, changing only a within-class permutation or matching. We will distinguish that untested causal claim from our measured data-source benefit. Transport cost and endpoint cosine will be presented as descriptive statistics, not as a ranking theorem.

The supported benefit is a learning-curve improvement for the tested FM-initialized students. We retain the convergence of the 200k-pool results at 100k updates, the loss of the advantage after direct one-step fine-tuning, and the failed adaptation of mature pMF. They delimit which consumers benefit and prevent the update ratios from being interpreted as a universal one-step quality or total-compute result.

Reviewer_Yu2wOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Training-State Shaping (TSS), a post-training approach for flow matching models aimed at improving few-step sampling quality and the downstream utility of synthetic teacher-generated pairs. For pixel-space models parameterized by clean-image prediction, the authors shift intermediate training states toward a detached exponential-moving-average (EMA) prediction while retaining the clean target, and further introduce a transport-consistent variant that pairs shaped states with their exact path tangents for native velocity-prediction models. In addition, they derive a deterministic time-reparameterization special case and an inference-only velocity rescaling companion. Experiments on class-conditional ImageNet across pixel-space and latent-space backbones show that the method improves low-NFE generation and allows downstream reflow and flow-map students to reach target FID scores with fewer updates or smaller synthetic data pools.

Critical strengths
  1. Studying how post-training input states and path tangents affect few-step sampling and downstream teacher-student couplings is well-motivated. The decomposition of the shaping direction into along-path (time reparameterization) and normal components provides a clear analytical lens to separate schedule effects from EMA-guided state corrections.
  2. The empirical evaluation is extensive and transparent. The authors provide multi-seed paired comparisons across JiT model scales (B/16, L/16, H/16) and solvers (Heun, DPM++2M, Euler), include an informative endpoint-FID-matched control (Vanilla14) showing that teacher endpoint FID alone does not determine downstream coupling quality, and openly document negative transfers and boundary conditions.
Critical weaknesses
  1. Theoretical soundness of TSS-X and limited incremental novelty of the EMA-guided mechanism:
    • Marginal velocity mismatch in TSS-X: In Section 3.1 (Eq. 1), TSS-X shifts the training input from ztz_t to z~t=zt+β(t)(x^a−zt)\tilde{z}_t = z_t + \beta(t)(\hat{x}_a - z_t) while keeping the original clean-image target xx and conditioning on the unshifted time tt. At inference time, JiT converts the network output into a velocity via (fθ(z,t)−z)/(1−t)(f_\theta(z, t) - z)/(1 - t). Because the supervision does not include the derivative of the perturbation β′(t)(x^a−zt)−β(t)ut\beta'(t)(\hat{x}_a - z_t) - \beta(t)u_t, the learned field no longer corresponds to the exact tangent of the shaped path z~t\tilde{z}_t (and thus does not satisfy the continuity equation for the marginal distribution of z~t\tilde{z}_t). Conversely, when the authors enforce exact transport consistency on JiT (such as the D+ND+N variant in Table 5 or the target-corrected ablations in Table 31), performance is worse than TSS-X. It remains theoretically unclear why an inconsistent velocity field outperforms the transport-consistent formulation in pixel space.
    • Combinatorial nature and dominance of simple time reparameterization: The transport-consistent formulation in Section 4 closely follows auxiliary-path flow matching (AuxPath-FM) combined with a heuristic EMA anchor and trust bound. More importantly, in the latent velocity-prediction experiments (Tables 3, 10, and 12), the EMA-free deterministic time reparameterization (TSS-T) accounts for almost all of the low-NFE gain: with DPM++2M at 10 NFE, TSS-T achieves 9.6469 vs. 9.6142 for full transport-consistent TSS (and at 9 NFE in Table 12, TSS-T is 11.2987 vs. 11.6026 for full TSS). This suggests that in velocity space, the core benefit essentially comes from warping the training time schedule near a=0.5a=0.5 rather than from EMA-guided state shaping.
  2. Experimental anomalies at medium/high NFE and practical limitations in downstream distillation:
    • Regressions at 20 and 50 NFE: While TSS improves 10-NFE generation, it consistently degrades latent-space generation at 20 NFE across both Euler (7.1748 to 7.3549) and DPM++2M (5.9126 to 6.3069) in Table 10, as well as Euler at 50 NFE (5.9242 to 6.2054) and JiT-B/16 / H/16 at 50 NFE under DPM++2M in Table 2. Since 20–50 NFE is a standard practical operating regime for flow matching solvers, this trade-off limits the method's utility as a general post-training upgrade.
    • Transient downstream gains and large gap to state-of-the-art one-step models: While the authors note in the text and appendix that the 200k-pool flow-map advantage narrows by 100k updates and that TSS-X student curves actually worsen after 40k updates (Figure 3 and Table 21, where 2-call EMA FID at 100k even favors Vanilla10 by +0.5604/+0.1510), framing this primarily as a 2.3x update-efficiency gain in the main text overstates the practical benefit when the asymptotic quality on the fixed pool converges and degrades. Furthermore, across all downstream experiments (Tables 4, 15, and 21), the student 1-call FID remains around 19.7–31.7, and vanishes under direct 1-step MSE+LPIPS fine-tuning (15.27 vs. 15.38 in Appendix A.4.8) or when adapting the mature official pMF checkpoint (where FID degrades from 3.15 to 24.76 in Appendix I.1). It is unclear whether TSS-generated couplings can improve a competitive few-step distillation pipeline that reaches strong absolute FID.
    • Unisolated experimental factors in control comparisons: Although the paper acknowledges in Section 5.5 and Appendix A.4.11 that the 640k-pool experiment changes pool size, storage precision (fp16 vs. fp32), LR schedule (cosine vs. constant), EMA decay (0.9999 vs. 0.999), and update budget (60k vs. 20k/100k) simultaneously—and that the guided flow-map comparison (Arm F) uses different initial weights (EMA vs. Online) and optimizer states—leaving these factors unisolated weakens the conclusions drawn from these comparisons.
  3. Writing, organization, and presentation:
    • Several parts of the paper (and especially the 36-page appendix) read like unedited internal experiment logs or rebuttal audit reports rather than a cohesive scientific paper. The text is filled with internal project jargon and provenance details (e.g., "Gate-2B", "Gate-2A", "arm F", "O0 calibration failure", "parent-20k bridge evaluation", raw seed numbers like 20260921, and hardware migration logs between Blackwell, H100, and H100 NVL). Streamlining the exposition and unifying the evaluation hardware/protocols would make the paper much easier to follow.
Questions for authors
  1. In TSS-X (Eq. 1), the model is trained on shifted inputs z~t\tilde{z}_t with the original clean target xx at time tt, and inference recovers velocity via (fθ(z,t)−z)/(1−t)(f_\theta(z,t)-z)/(1-t). From my understanding, this velocity does not match the tangent of the shaped path z~t\tilde{z}_t and introduces a time-dependent bias relative to the marginal flow. Could the authors clarify the theoretical justification for why this biased velocity field improves low-NFE sampling on JiT, whereas enforcing the exact path tangent (Table 31 and D+ND+N in Table 5) performs worse?
  2. In the latent LightningDiT experiments (Tables 3, 10, and 12), deterministic time reparameterization (TSS-T) achieves virtually the same (or better) DPM++2M performance as full transport-consistent TSS without using any EMA prediction. Does this imply that the primary bottleneck in low-NFE latent flow matching is simply the training time-sampling/schedule density around t=0.5t=0.5? Have the authors compared TSS-T against standard non-uniform time sampling (e.g., logit-normal time weighting) or inference time-step rescheduling during post-training?
  3. Why do both TSS-T and full transport-consistent TSS degrade FID at 20 NFE on LightningDiT (Table 10, e.g., DPM++2M degrading from 5.91 to 6.31)? Since the shaping profile b(t)b(t) is supported on [0.25,0.75][0.25, 0.75] with b(0)=b(1)=0b(0)=b(1)=0, Proposition 1 guarantees that the exact marginal flow has the same endpoint distribution—why does a finer 20-step solver suffer more than a 10-step solver?
  4. In the downstream flow-map experiments on the 200k pool (Figure 3 and Table 21), why does the TSS-X student's 1-call and 2-call FID start degrading after 40k updates while the Vanilla student continues to improve or remain stable up to 100k updates? In addition, since the simplified JVP student recipe only reaches ~20 FID at 1 NFE, could the authors demonstrate whether TSS-generated pairs still provide data-efficiency or final-FID gains when used in a full-featured distillation recipe (e.g., with perceptual/LPIPS loss or distribution matching) that achieves competitive (<5) 1-step FID?
1 comment

Thank you for recognizing the controlled experiments and the state/tangent decomposition. Your comments identify distinctions that the presentation should make more clearly.

1. What TSS-X optimizes. Let δt=β(t)(x^a−zt)\delta_t=\beta(t)(\hat x_a-z_t) and u=x−ϵu=x-\epsilon. Away from the denominator floor, retaining the clean-image target induces vX∗=x−z~t1−t=u−δt1−t,z~˙t=u+δ˙t.v_X^*=\frac{x-\tilde z_t}{1-t}=u-\frac{\delta_t}{1-t},\qquad \dot{\tilde z}_t=u+\dot\delta_t. Thus the discrepancy is −δt/(1−t)−δ˙t-\delta_t/(1-t)-\dot\delta_t, with δ˙t=β′(t)(x^a−zt)−β(t)u\dot\delta_t=\beta'(t)(\hat x_a-z_t)-\beta(t)u. Your observation is correct: TSS-X optimizes clean-image prediction on shifted inputs; the transport-consistency proposition applies to the construction with the corrected tangent. It does not justify TSS-X as exact flow matching on that shaped path.

The JiT D+N ablation also retains the clean-image objective; it changes the direction, not target consistency. Its worse result therefore does not show that a correct tangent is harmful. Our present evidence establishes which recipe works under the tested budgets, rather than a theorem explaining its advantage. Additional diagnostics on 8,192 trajectories per model over three training seeds find slightly more curved TSS-X trajectories. We will remove the suggested straightening explanation and separate this observation from endpoint coupling geometry.

2. Time transformations and novelty. We credit AuxPath-FM for the general auxiliary-path construction. Our contribution is the EMA-conditioned post-training rule, its decomposition into time and normal components, and the controlled sampling/downstream evaluation. The continuity argument is a correctness requirement for the transport-consistent variant.

We completed the requested inference-only control: w(z,t)=ϕ′(t)vvanilla(z,ϕ(t))w(z,t)=\phi'(t)v_{\rm vanilla}(z,\phi(t)). It changes both query time and velocity scale, using the fixed training profile. At Euler NFE10, three matched seeds with the same 50,000 noise/class IDs, guidance and evaluation environment give mean FIDs:

VanillaVanilla + inference transformTSS-T trainingFull transport-consistent TSS
14.645813.905713.965813.6358

TSS-T minus the inference baseline is +0.0600+0.0600 (paired 95% interval [−0.4220,+0.5421][-0.4220,+0.5421]): this experiment shows no additional benefit from TSS-T training. Full TSS reuses the reported evaluation; its post-hoc difference from the inference baseline is −0.2700-0.2700 [−0.8975,+0.3576][-0.8975,+0.3576], favoring full TSS in each seed but insufficient to establish superiority. We will present the time component as a derived control, with the train-versus-inference comparison explicit. Ordinary time-density/reweighting training remains a separate, untested baseline.

3. Endpoint preservation and high-NFE behavior. Endpoint preservation concerns the prescribed probability path and its exact marginal velocity. Learned-field error and numerical integration error remain. It therefore does not imply that the trained model improves at every NFE. We retain the 20/50-NFE regressions and describe the method as a low-NFE intervention. The proposition should not be read as a guarantee about finite-model FID.

4. Downstream benefit and experimental matching. The 2.3x number measures student updates to specified target FIDs on fixed pools, using the reported milestone interpolation. It is not a full-pipeline speedup or an asymptotic quality claim. The 100k results, post-40k deterioration and negative consumers remain part of the evidence. The 640k experiment compares two data sources under one common recipe; because several settings differ from the 200k experiment, it cannot isolate pool-size scaling. It does provide a matched endpoint comparison: one-call EMA FID 21.0873 to 20.4418.

We have also completed the initialization/optimizer-matched comparison requested by your fifth concern. All three arms start from the same epoch-600 EMA backbone, fresh AdamW and their own initial EMA, on eight H100s. Each receives approximately 87,590 training seconds; stopping overshoots are below 1.1 seconds. Final EMA FIDs are:

NFEVanilla continuationTSS-XGuided flow-map post-training
521.581619.30528.6660
109.81078.37206.6960
503.75593.72243.9487

Direct flow-map post-training wins at 5/10 NFE. This is a matched JiT recipe comparison, not a reproduction of official pMF training. Our data-source results concern training students from FM weights; transfer to a competitive sub-5-FID one-step system remains unestablished.

Finally, we will reorganize the appendix by scientific question, moving job labels, migration records and operational logs out of the research presentation while retaining protocols and unfavorable results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.