acceptodds
Under review as a conference paper at ICLR 2027

PT-Flow: One-Step Schrödinger-Bridge Generation by Proximal Tilting

11 views

Abstract

One-step generative models usually compress a many-step integration of a velocity field into a single step, using distillation, consistency conditions, trajectory straightening, or averaged fields. This compression tends to give up the optimality of the learned map or an explicit transport objective. PT-Flow takes a different route. We choose the governing law of the field (a viscous potential flow, i.e. the Schr\"odinger bridge) so that its time evolution has an exact closed-form solution through the Cole–Hopf transform. The single step is then not a shortened trajectory but the closed-form solution of the flow. The one hard part is a single Gaussian expectation whose naive estimator has variance that grows sharply as the flow becomes sharp. We remove this obstruction with a prox-tilted importance estimator: we center the proposal at the proximal point and scale it by the local curvature, which cancels the leading variation and makes the remaining variance shrink as the viscosity decreases. The estimator therefore becomes more accurate exactly where generation is sharp, and the same network that builds the proposal is the one-step generator itself, so sampling is a single network evaluation with no teacher, no adversarial training, and no ODE solver at inference. The learned scalar potential also defines a normalized Schr\"odinger-bridge marginal. On ImageNet , PT-Flow reaches an FID of 1.54 with one function evaluation. It combines one-step sampling with teacher-free, adversarial-free training and optimal-transport consistency at convergence.

Primary area:
generative models
Keywords:
Generative Models, One-Step Generation, Schrödinger Bridge, Optimal Transport, Flow Matching

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

This paper proposes PT-Flow one-step generative modeling framework, which utilizes Schrödinger bridge to construct analytical target generation function. They parameterize the data-side potential to approximate data distribution and the proposal mean to reduce estimation variance and serve as generator. The experiments are primarily conducted on ImageNet 256 × 256 and get good results.

Critical strengths
  • It's quite interesting to explore generative modeling paradigms beyond diffusion/flow-based models, so any exploration in this direction should be encouraged.
  • This paper presents detailed derivation of PT-Flow and gives substantial complexity analysis, which is novel and interesting.
Critical weaknesses
  • Although PT-Flow parameterize proposal to reduce estimation variance, training this one-step generative model seems potentially unstable: there is no theoretical guarantee for the alternating optimization scheme will converge to optimal solution, and this type of two-player optimization is to be challenging to train such as GAN.
  • The experimental evaluation is limited to ImageNet and the results is not sufficiently convincing to demonstrate the general effectiveness of PT-Flow.
Questions for authors
  • I have generally understand the design principle of proposal, but it seems just a heuristic method and no guarantee for reducing importance weight variance. Can you give some theoretical evidence for this?
  • I do not understand the statement in lines 195–197. How do you calculate "0.14" and what is the meaning of "the unit data scale"? It's quite confusing to encounter this statement at this place.
  • In section 4, why do you present the ablation study first then show the main results? I think this ablation study should not shown in the main text because this is just how you tune the hyperparameters.
  • As you state that "training is expensive", you should report the training time of your main experiments and compare it with other baselines.
1 comment
AuthorsOpenReview ↗

Response to Reviewer LFxk

Thank you for asking what the proposal guarantees and what evidence supports the alternating training procedure. We address these questions separately below and report a completed second-dataset experiment.

W1/Q1. First, the potential network uθu_{\theta} and the proposal network mηm_{\eta} do not have opposite goals as in a GAN. The potential network learns the bridge potential. The proposal network predicts the center of the importance proposal during training and serves as the one-step generator at inference. During each generator update, the potential is held fixed, and the generator is trained by regression to satisfy the proximal condition (Eq. 15). The generator loss does not use the value of the potential loss (Appendix I.2). A better picture is a half-Sinkhorn step: one side of the bridge has a closed-form expression, and the other is fit by gradient steps. The potential changes between updates, so the two training steps are still coupled.

Second, the variance fix is not just a heuristic. Theorem 3.9 gives an O(ε) bound on the log-weight variance of the centered diagonal proposal under condition (A6) and the stated remainder bound. This is a small-ε result. At our ImageNet setting, ε=0.01 and εd≈41, outside that regime. Table 7 checks how the proposal works at this setting: on the same frozen checkpoint, mean ESS/K is 0.0768 for the naive proposal, 0.9158 after recentering, and 0.9895 for the defensive mixture. Most of the observed gain comes from recentering.

On training stability, we completed the same 160,000-step schedule for the B/2, M/2, and L/2 models. This shows that the alternating procedure could be trained at all three sizes. A general convergence guarantee for the trained networks remains open, as we discuss in Appendix I.4.

W2. Evaluation beyond ImageNet. On FFHQ-256, PT-Flow-B/2 achieved FID-50k of 2.12 at 1 NFE. We used the same 256×256 latent pipeline and generated 50,000 images with 1 NFE. We computed FID using the 60,000 FFHQ training images as the real-image reference. On ImageNet-1K, the 1-NFE FIDs were 1.73, 1.62, and 1.54 for B/2, M/2, and L/2, respectively.

Q2. Origin and interpretation of “0.14.” This line was too compressed. The 0.140.14 comes from Proposition 3.4, with the scale of the bound written as a fraction of the data scale. The proposition bounds the W2W_2 distance between the distribution of the exact conditional-mean output and the bridge marginal by the square root of the expected conditional covariance trace. The Laplace approximation gives 2εd1+λˉ\sqrt{\frac{2\varepsilon d}{1+\bar{\lambda}}} for this quantity.

“Unit data scale” refers to assumption (A5): we scale the data to unit variance per dimension, so the root-mean-square distance from the data mean is d\sqrt{d}. Dividing by d\sqrt{d} cancels the dimension:

2εd/(1+λˉ)d=2ε1+λˉ.\frac{\sqrt{2\varepsilon d/(1+\bar{\lambda})}}{\sqrt{d}} = \sqrt{\frac{2\varepsilon}{1+\bar{\lambda}}}.

At ε=0.01\varepsilon=0.01, neglecting the curvature correction gives 0.02\sqrt{0.02}, about 0.140.14. This is an approximate scale for the distribution-level bound. We will rewrite the passage to show this calculation, define (A5) there, and make clear where we switch from an absolute distance to a relative one.

Q3. Order of the experiments. We agree. We moved the main results table (Table 3) before the ablation, and pushed the tuning details to the end of the section. (Another reviewer asked us to move the algorithm into the main text; we did both in the same revision.)

Q4. Training time. All three ImageNet models followed the same 160,000-optimizer-step schedule. The wall-clock estimates use the measured time per step on each configuration.

ModelTraining GPUsTime/stepPeak memory/GPUTraining time
B/22×H20010.8 s60.2 GB480 h
M/24×H20011.2 s56.2 GB498 h
L/28×H10013.0 s69.3 GB578 h

For L/2, the effective global batch was 1,024 images per optimizer step. Each GPU contributed 128 images across eight gradient-accumulation steps. Training from scratch took an estimated 24.1 days on eight H100 GPUs.

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

The paper introduces PT-Flow, a one-step generative model derived from a Schrödinger-bridge formulation. It combines closed-form bridge propagation, the Cole–Hopf transform, proximal-map ideas, and importance sampling to build a teacher-free, adversarial-free, distillation-free generator. The proposal network used for variance reduction is also deployed as the generator, which is the paper’s most interesting conceptual contribution. Experiments on class-conditional ImageNet at 256×256 report competitive single-forward-pass FID and throughput.

Critical strengths

<1>The proposal is generator construction is genuinely interesting and connects the bridge objective to a practical one-step architecture without a separate distillation stage. <2>The reported ImageNet result and throughput are competitive among from-scratch one-step models.

Critical weaknesses

<1>The variance result controls the log-weight, but the sample-budget argument needs a second-moment or tail condition on the actual importance weights; the appendix invokes such conditions without stating or verifying them. With only eight proposal samples, the normalized ESS cannot fall below one eighth, so the pseudocode’s recovery branch triggered below five percent cannot activate under the default setting. Since the variance reversal and alternating scheme are central claimed contributions, this mismatch substantially weakens the paper’s main theoretical claim.

<2> For simple Gaussian targets, the zero-viscosity guided proximal construction already differs from classical CFG density tilting at first order in the guidance weight, so the claimed agreement is not correct as stated. In addition, Table 3 appears to list iMeanFlow at 1.72 and W-Flow at 1.29 for single-step generation; if confirmed, the proposed 1.54 is competitive but not SOTA. These issues directly affect the conceptual interpretation of guidance and the paper’s empirical positioning, and if unresolved they support rejection.

Questions for authors

<1>The sample-budget argument requires a second-moment or tail condition on the actual importance weights, but Appendix H.2 only refers to additional tail conditions. Can you state these conditions explicitly and provide evidence that they hold for the trained potential at the operating point? If not, which theoretical claims should be read as heuristic rather than guaranteed?

<2>With the default K=8, the normalized ESS cannot fall below 1/8, so the pseudocode branch triggered when ESS/K is below 0.05 cannot activate. Is the deployed monitor based on a larger evaluation batch, a different weighting, or a different ESS definition? Please clarify the actual procedure, since this affects whether the alternating scheme is correctly specified.

<3>At the reported operating point, the experiments appear well outside the small-viscosity asymptotic window used to motivate the method. Which claims do you intend to hold at this operating point, and which are only motivational? This materially affects how much weight I assign to the theoretical contribution.

2 comments
AuthorsOpenReview ↗

Response to Reviewer BRbp

Q1. Empirical weight diagnostics and the second-moment condition. At the working point (ε=0.01, K=64), Table 7 reports a mean normalized ESS of 0.9895 for our proposal, against 0.0768 for the naive one. Applying 1/r−1 to these mean ESS fractions gives 0.0106 and 12.0, respectively (Fig. 2b). Our draft labeled these values χ²/K. To obtain the mean empirical χ², we must apply 1/r−1 to each input's ESS fraction before averaging. The measurements show that the weights were less concentrated under our proposal at the working point.

The condition is simple: each importance weight must have a finite second moment. The defensive mixture was already part of our proposal before the rebuttal. The weight is

w=G2ε(y−x)e−uθ(y)qdef(y∣x).w=\frac{G_{2\varepsilon}(y-x)e^{-u_\theta(y)}}{q^{\mathrm{def}}(y\mid x)}.

Our proposal is qdef=(1−α)q+αN(x,2εI)q^{\mathrm{def}}=(1-\alpha)q+\alpha\mathcal{N}(x,2\varepsilon I). The second part uses the same Gaussian G2ε(y−x)G_{2\varepsilon}(y-x) that appears on top of the weight. So the denominator is always at least αG2ε(y−x)\alpha G_{2\varepsilon}(y-x). That Gaussian cancels, and we are left with w≤e−uθ(y)/αw\le e^{-u_\theta(y)}/\alpha. For the second moment, this gives Eqdef[w2]≤1α∫G2ε(y−x)e−2uθ(y),dy\mathbb{E}_{q^{\mathrm{def}}}[w^2] \leq \frac{1}{\alpha} \int G_{2\varepsilon}(y-x)e^{-2u_{\theta}(y)}\\,dy.

If this integral is finite, the weight has a finite second moment. Assumption (A1) alone does not establish this stronger condition for every trained network. We will state the condition explicitly in Appendix H.2.

Theorem 3.9 proves an O(ε) bound on log-weight variance for the centered diagonal proposal under its curvature and remainder assumptions. The follow-up claim that one sample is enough as ε→0 (Eq. 42) additionally requires control of the weight moments as ε decreases. We use this limit result to motivate the proposal. Table 7 evaluates its behavior at our finite operating point.

Q2. ESS thresholds and recovery at K=8. The reviewer is correct that raw ESS/K has a floor of 1/8=0.125 at K=8. Algorithm 1 incorrectly labels this quantity as the controller statistic. The reported L/2 training run used the eight proposal weights from the training step and computed

s=ESS/K−1/K1−1/K=ESS−17.s=\frac{\mathrm{ESS}/K-1/K}{1-1/K}=\frac{\mathrm{ESS}-1}{7}.

We applied the thresholds to the smoothed value of s. A value of at least 0.30 indicated healthy operation, and a value below 0.05 triggered recovery. Because s=0 when ESS=1, the recovery threshold is reachable with eight draws. Table 7 used K=64 in a separate frozen-checkpoint evaluation. We will correct Algorithm 1 to specify the rescaling and smoothing.

Q3: Which claims hold at our setting. We agree, and we now say this clearly in the paper (Appendix K). At our setting, εd is about 41, which is outside the small-viscosity window. So we split the results into two groups.

  • True at our setting, for every ε>0 under the stated assumptions, with no small-viscosity assumption: the closed-form propagation (Eq. 3), the Tweedie mean (Thm. 3.3), the exact normalization (Thm. 3.5), and the loss that solves the bridge (Thm. 3.6). These are the main claims.

  • Small-ε results used to motivate the design: the approximation of the conditional mean by the prox (Eq. 4), the O(ε) bound on log-weight variance for the centered diagonal proposal under its curvature and remainder assumptions (Thm. 3.9), and the predicted naive blow-up under its approximations. At our setting, Table 7 compares normalized ESS on a frozen checkpoint: our estimator stays healthy (mean ESS/K about 0.99 at ε=0.01), while the naive proposal gives about 0.077. This comparison does not establish the asymptotic rate.

W2. On the FID. We agree, and we withdraw the state-of-the-art claim from the paper. Among one-step models that use no teacher and no adversary and train from scratch, PT-Flow-L/2 (FID 1.54) is competitive. W-Flow-XL/2 (1.29) has a lower FID and is a larger model (679M+49M). W-Flow-L/2 also has a lower FID (1.35). W-Flow shapes the output with a training-time field instead of solving the bridge in closed form. iMeanFlow (1.72) is XL size (610M). Our point is the closed-form one-step method and the variance fix. Our new B, M, L results show how FID improves with model size: 1.73 at 132.7M, 1.62 at 311.2M, and 1.54 at 462.9M.

On guidance. The reviewer is right, and this fixes a claim we wrote loosely. Our guided result and the usual CFG tilt match when w=0 (no guidance), but they can differ at first order in w, even in the limit ε→0, as the Gaussian example shows. The reason is that the second factor in our bridge re-adjusts itself instead of being tilted in the same way as the first. We therefore withdraw the claim that they agree to first order in w. They agree at zeroth order. This only changes how we compare to classical CFG. It does not change the sampler. Under its stated assumptions, Proposition 3.10 still gives an exact bridge onto the marginal defined by our guided potential.

Thank you for the detailed responses and for clarifying the ESS procedure, the scope of the theoretical claims, and the guidance and FID comparisons.

Reviewer_Xvb8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper "PT-FLOW: ONE-STEP SCHROEDINGER-BRIDGE GENERATION BY PROXIMAL TILTING" proposes a new one-step generative formulation based on the closed form solution to the Schrodinger-Bridge equation utilizing the Cole-Hopf transformation. One of the main contributions is a special estimator that is able to keep the variance bounded during the training. The authors perform experiments on ImageNet 256x256 and show that their approach perform strongly for a single setup. The authors also share an investigation with regards to the hyperparameters of the approach. In the Appendix a flowchart of the algorithm is included but the authors have so far not shared any code.

Critical strengths

The overall approach starting with utilizing the cole-hopf trafo to state a closed form solution to the SB equation is a good starting point and a novel perspective for one-step generative models. The authors compare to many relevant methods -although only for a single dataset- and the performance of the proposed method is good.

Critical weaknesses

Unfortunately I can currently not recommend the paper for acceptance due to the following reasons:

  • The writing of the paper can be improved. In the Reviewers' opinion the paper is in parts hard to follow. For instance the summary of the algorithm could be moved to the main part of the paper. Moreover, for their method there are some approximations involved which should be more clearly stated.
  • In the Reviewers' opinion the current evaluation of the new approach is insufficient: The authors only compute the results for their method for a single setting (i.e the 456 M parameter case). More settings should be included especially given that the section before showed that there is quiet some hyper-parameter tuning necessary.
  • Moreover the Reviewers expects experiments with regards at least one more dataset.
  • The picture of the car for the optimal setting is looking strange as it is obviously very distorted. The same is the case for some of the pictures in the appendix. Could the authors comment on this ?
  • Moreover, computational costs are insufficiently discussed in the paper.
  • As a potential is learned could the authors compare to recent energy-based approaches for generative modeling and how their approach compares ?
Questions for authors

See Weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer Xvb8

Thank you for highlighting the distinction between the closed-form bridge construction and the approximations used by the generator. We address your six points below.

W1. Algorithm and approximations. We agree that the algorithm should be easier to follow. We moved its summary into the main text and placed the main results before the ablations. During training, we learn a bridge potential and train a network to approximate its proximal map. At sampling time, the network maps noise to a sample in one forward pass.

At finite ε, the bridge propagation formula (Eq. 3), the conditional mean formula (Thm. 3.3), and the normalization of the bridge density (Thm. 3.5) are exact under their stated assumptions. Thm. 3.6 shows that minimizing the population objective solves the bridge problem. Using the proximal map in place of the conditional mean is a small-ε approximation. In this regime, Thm. 3.9 gives an O(ε) bound on the variance of the log weights for the centered diagonal proposal under its curvature and remainder assumptions.

Our ImageNet setting has εd ≈ 41, outside this regime. Table 7 compares proposals on the same frozen model. At ε = 0.01, the centered and defensive proposals give higher normalized effective sample size than the naive proposal, meaning that the weights are less concentrated. Verifying the theoretical rate or showing that training converges would require separate evidence.

W2. Model sizes. We evaluated three PT-Flow models at 256 × 256 resolution using one function evaluation (1 NFE) per generated sample.

ModelParameters1-NFE FID ↓
PT-Flow-B/2132.7M1.73
PT-Flow-M/2311.2M1.62
PT-Flow-L/2462.9M1.54

W3. Second dataset. On FFHQ-256, PT-Flow-B/2 achieves FID-50k of 2.12 at 1 NFE. We generated 50,000 images using sampler A with zero refinement steps. We computed FID using the 60,000 FFHQ training images as the real-image reference.

W4. Distorted samples. You are right that the car in the original figure has visibly distorted geometry. We examined 12 additional fixed-seed samples from the same ImageNet class and guidance setting (sports car, class 817, w = 1.25). Some of these cars also show distorted shapes or proportions. This shows that the failure can recur. Twelve samples are too few to estimate its frequency. We will acknowledge this limitation in the qualitative discussion.

W5. Training cost. All three ImageNet models used 160,000 optimizer steps.

ModelTraining GPUsSeconds/stepPeak memory/GPUTraining time
B/22 × H20010.860.2 GB480 h
M/24 × H20011.256.2 GB498 h
L/28 × H10013.069.3 GB578 h

For L/2, the effective global batch was 1,024 images per optimizer step. Each GPU contributed 128 images across eight gradient-accumulation steps. At inference, L/2 generates approximately 165 images/s on one H100, including VAE decoding (Table 6).

W6. Comparison with energy-based methods. Equilibrium Matching [EqM] learns the gradient of an implicit energy and generates samples by iterative optimization. PT-Flow also learns a potential and uses it during training to fit a separate network to the proximal map. Sampling then takes one network evaluation. A closer one-step comparison is Three-Body Scattering Modeling [TBSM], which trains its generator using an energy-distance objective and reports FID 1.63 with a latent DiT-XL on ImageNet-256. PT-Flow-L/2 reports 1.54 on that benchmark. The methodological distinction is that our training target comes from a Schrödinger bridge whose marginal is normalized under Theorem 3.5.

References

[EqM] Wang and Du. Equilibrium Matching: Generative Modeling with Implicit Energy-Based Models. 2025.

[TBSM] Sun et al. Three-Body Scattering for Generative Modeling. 2026.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Additional results on model size, FFHQ-256, and training cost

1. Model size on ImageNet-1k

We evaluated three PT-Flow models at 256 × 256 resolution using one function evaluation (1 NFE) per generated sample.

ModelParameters1-NFE FID ↓
PT-Flow-B/2132.7M1.73
PT-Flow-M/2311.2M1.62
PT-Flow-L/2462.9M1.54

2. FFHQ-256

PT-Flow-B/2 achieved FID-50k of 2.12 at 1 NFE. We generated 50,000 images at 256 × 256 resolution using sampler A with zero refinement steps. We computed FID using Torch-Fidelity-compatible Inception-V3 features and the 60,000 FFHQ training images as the real-image reference.

3. ImageNet training cost

We trained all three models from scratch using the same 160,000-optimizer-step schedule.

ModelTraining GPUsTime per stepPeak memory per GPUTraining time
PT-Flow-B/22 × H20010.8 s60.2 GB480 h (20.0 days)
PT-Flow-M/24 × H20011.2 s56.2 GB498 h (20.7 days)
PT-Flow-L/28 × H10013.0 s69.3 GB578 h (24.1 days)

For PT-Flow-L/2, the effective global batch was 1,024 images per optimizer step. Each of the eight GPUs contributed 128 images across eight gradient-accumulation steps.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.