acceptodds
Under review as a conference paper at ICLR 2027

Posterior Flow Matching

10 views

Abstract

Flow matching learns how to transport noise toward data, but its mean prediction does not directly report the endpoint uncertainty remaining at an intermediate state. This paper introduces Posterior Flow Matching (PFM), a general generative modeling framework that learns conditional velocity means and componentwise variances with a Gaussian posterior model. The same model evaluation supplies both a transport prediction and endpoint uncertainty, making posterior diagnostics available throughout generation. These learned moments support a mean-driven ordinary differential equation (ODE) and a posterior-driven stochastic differential equation (SDE), whose local noise scales adapt to the predicted uncertainty. Across image, point-cloud, molecular, and action generation, PFM improves generation quality and provides endpoint-posterior readouts throughout generation.

Primary area:
generative models
Keywords:
Generative Models, Flow Matching, Diffusion Models, Uncertainty Quantification

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ZB7TOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper matches the distribution itself of the Random Variable

$V= (X_1 - X_0 \mid X_t = x) $

to the field of mean vectors and variances (diagonalized). This differs from the usual Flow Matching that matches only the field to explicit samples from VV. The authors then utilize the learnt field for sampling of the usual ODE and also an instance of related SDE. Performance gain is probably from weighted ($gamma_{\theta, i}) flow matching objectives.

Critical strengths

Fitting a flow to vector field as a Random Variable is a worthy hindsight.

Critical weaknesses

Weaknesses

This paper is badly written with no regard to the reader's structural mathematical understanding. Most notably, it does not clearly distinguish what's random and what is not. Lowercase vv is once introcued as constant in equation (2), so considering its law is nonsensical (12) if we read this literally--- requiring the reader to reconstruct what is what from the context. But after a long while, I realized that the core of this paper is distributional matching of

(X1−X0∣Xt=x)(X_1 - X_0 \mid X_t = x)

to a model.

Such restructuring of consistency shall not be the reader's task. The document as a whole required the readers to restructure the logic.

Also, I understand the analogy of Section 3.3 as a separate story, but if one comes to apply a (V∣Xt=x)(V \mid X_t = x) from a different model to this, it is a big leap in concept that requires more justification. The system of SDE is governed with a completely different filtration, and while the result is interesting, it appears to be an "additional technical attempt" to be in the discussion—it added more to my confusion.

I support the idea (after I understood it) and support the result, but I don't recommend the paper of this construction for publication, and I feel that it needs major revision as well as additional theoretical work for connecting what seems to be all over the place.

Questions for authors

Flow matching and Diffusion models can both be understood in one common framework as Gradient Flow (See, https://arxiv.org/pdf/2204.04853(ICLR2023) or (https://arxiv.org/pdf/2407.04495 Physical Review X ), for few of examples, and they all differ only through what energy functional EE they are associated with:

ddtpt=−∇⋅(pt∇δEδpt)\frac{d}{dt} p_t = -\nabla \cdot (p_t \nabla \frac{\delta E}{\delta p_t} )

What is the Energy functional EE does this poseterior Flow correspond to?
Why is this paper's EE useful, and when is it expected to be good choice and when it is not? I feel that the explanation of it shall be better because it is probabilsitic is insufficient, since the preferred choice of probabilistic model as well as preferred choice of entropy differs across situations--- with myriads of tricks being found today, I believe that each researcher is responsible for answering such questions.

Likewise, when this is used for the related diffusion model (19), how does it change the original diffusion's energy functional?

In a related note, I feel that the performance gain is probably arising from weighted ($gamma_{\theta, i}) flow matching objectives. Please further elaborate in relation to flow-theory.

6 comments
AuthorsOpenReview ↗

1/4 — The velocity target and the sampler's filtration

Thank you for pointing out where the notation made the argument hard to follow. The revised paper distinguishes the reference variables, the moments PFM learns, and the sampler that uses them.

The vv in Eq. 2 is a conditional mean. The random velocity target is V=X1−X0V=X_1-X_0. Here X0∼p0X_0\sim p_0 and X1∼pdataX_1\sim p_{\rm data} are independent, and Xt=(1−t)X0+tX1X_t=(1-t)X_0+tX_1 (Eq. 1). At a fixed (x,t)(x,t), v(x,t)=E[V∣Xt=x]v(x,t)=\mathbb E[V\mid X_t=x] is a deterministic vector and vθ(x,t)v_\theta(x,t) estimates it. We rewrote Eq. 12 in Sec. 4.1 to show that the Gaussian approximates the conditional law of VV:

qθ(⋅∣x,t)=N(vθ(x,t),Σθ(x,t)).q_\theta(\cdot\mid x,t)=\mathcal N(v_\theta(x,t),\Sigma_\theta(x,t)).

At the population optimum, the Gaussian fits the conditional mean and componentwise variances when they are attainable (Proposition 4.1 and Appendix B.4). For a fixed inexact mean, the pointwise variance target is the conditional residual second moment, clipped to the stated bounds (Eqs. 43-44). The affine relation in Eq. 10 gives the endpoint moments in Eq. 15:

μθ=x+(1−t)vθ,Σθx=(1−t)2Σθ.\mu_\theta=x+(1-t)v_\theta,\qquad\Sigma_\theta^x=(1-t)^2\Sigma_\theta.

The SDE evaluates the learned fields at its own state YtY_t. Once training is finished, the drift and diffusion coefficients are fixed functions of state and time. The sampler's filtration includes its initial state and Brownian history. Since YtY_t is adapted to this filtration, the coefficients evaluated at YtY_t are adapted as well (revised Appendix A.5).

The original Sec. 3.3 and Appendix A.5 explain why the sampler can have the reference marginals. Let ptp_t be the density of XtX_t. Eq. 8 takes X0∼N(0,I)X_0\sim\mathcal N(0,I) to relate the velocity to the score. With the exact velocity and score, Eq. 19 uses aθ=λΣθa_\theta=\lambda\Sigma_\theta and bθ=v+(∇⋅aθ+aθ∇log⁡pt)/2b_\theta=v+(\nabla\cdot a_\theta+a_\theta\nabla\log p_t)/2. The product rule gives

−∇⋅(bθpt)+12∑i,j∂i∂j(aθ,ijpt)=−∇⋅(vpt).-\nabla\cdot(b_\theta p_t)+\tfrac12\sum_{i,j}\partial_i\partial_j(a_{\theta,ij}p_t)=-\nabla\cdot(vp_t).

With the same initial law, regularity, boundary conditions, and unique density evolution as in Appendix A.5, YtY_t has density ptp_t at each time. The calculation works for any admissible diffusion tensor with its corresponding drift correction. PFM uses the learned covariance to set that tensor.

The reference process and the sampler can have different joint laws across time. At a visited state Yt=yY_t=y, Eq. 15 estimates the mean and covariance of the reference endpoint posterior p(X1∣Xt=y)p(X_1\mid X_t=y) (original Sec. 4.3 and Appendix D.3). The sampler's own conditional endpoint law comes from its transition dynamics. The one-time marginal result concerns the exact continuous SDE. Appendix G.4 analyzes the additional randomness from estimating the derivative in the finite-step sampler.

AuthorsOpenReview ↗

2/4 — The reference energy and Eq. 19

PFM and FM use the same reference velocity, so they share the functional in your equation. Let ptp_t be the density of Xt=(1−t)X0+tX1X_t=(1-t)X_0+tX_1, where X0X_0 is standard Gaussian and independent of the data endpoint X1X_1. On 0<ε≤t≤T<10 < \varepsilon\le t\le T < 1, take positive smooth densities with finite displayed moments and entropies and the stated integration conditions. The velocity-score identity in submission Eq. 8, derived in Appendix A.4, gives

v(x,t)=xt+1−tt∇log⁡pt(x).v(x,t)=\frac{x}{t}+\frac{1-t}{t}\nabla\log p_t(x).

In the sign convention ∂tpt=−∇⋅(pt∇δE/δp)\partial_t p_t=-\nabla\cdot(p_t\nabla\delta E/\delta p), take E=GtE=\mathcal G_t:

Gt[p]=12t∫∣x∣2p(x),dx+1−tt∫p(x)log⁡p(x),dx.\mathcal G_t[p]=\frac1{2t}\int\\|x\\|^2p(x)\\,dx+\frac{1-t}{t}\int p(x)\log p(x)\\,dx.

The first variation is δGt/δp=∣x∣2/(2t)+(1−t)(log⁡p+1)/t\delta\mathcal G_t/\delta p=\\|x\\|^2/(2t)+(1-t)(\log p+1)/t. Evaluated at ptp_t, its spatial gradient is vv, so the equation gives the continuity equation for the reference path. The usual Wasserstein descent convention uses −Gt-\mathcal G_t. This functional represents the transport along the prescribed reference path. Eq. 13 gives the loss used to train PFM.

Koshizuka and Sato [1, Secs. 3.1–3.2] use a Lagrangian action between population marginals. Ikeda et al. [2, Secs. III–IV] relate diffusion accuracy to entropy production and Wasserstein geometry.

For score-based diffusion, the probability-flow ODE and the reverse SDE with tensor g(1−t)2Ig(1-t)^2I share the noising-path marginals pˉ1−t\bar p_{1-t} with the exact score and matching initialization (submission Eqs. 5-6 and Appendix A.6). Eq. 19 applies the ODE-SDE correspondence in Eq. 7 to the FM path ptp_t, with aθ=λΣθa_\theta=\lambda\Sigma_\theta. With the reference path and its velocity fixed, choosing a different tensor and applying its drift correction leaves the reference functional above unchanged. It changes how perturbations dissipate. Appendix A.5 gives the conditions for matching the one-time marginals, and Proposition 4.2 and Appendix C.3 give the relative-entropy calculation.

References

  • [1] Koshizuka and Sato. Neural Lagrangian Schrödinger Bridge: Diffusion Modeling for Population Dynamics. ICLR, 2023.

  • [2] Ikeda et al. Speed-accuracy relations for diffusion models: Wisdom from nonequilibrium thermodynamics and optimal transport. Physical Review X, 15, 031031, 2025.

AuthorsOpenReview ↗

3/4 — Why use the posterior covariance?

The learned covariance estimates how much the velocity targets vary in each coordinate at a given state and time. We choose the SDE noise so that it is isotropic after dividing by the posterior standard deviations (Eqs. 17-18 and Appendix C.1):

Σθ−1/2aθΣθ−1/2=λ(t)I,aθ=λ(t)Σθ.\Sigma_\theta^{-1/2}a_\theta\Sigma_\theta^{-1/2}=\lambda(t)I,\qquad a_\theta=\lambda(t)\Sigma_\theta.

The learned covariance sets the relative noise strengths across states and coordinates, and λ\lambda sets the overall strength. With the exact mean and derived score, the score and divergence corrections in Eq. 19 preserve the reference marginals under the conditions in Appendix A.5.

To see how this noise affects a density perturbation, let ptp_t be the reference density and p~t\tilde p_t evolve under the same SDE from a possibly different initial law. The drift for both uses the score of ptp_t. Appendix C.3 gives the equation for p~t\tilde p_t (Eq. 49):

∂tp~t+∇⋅(vp~t)=12∇⋅(aθp~t∇log⁡p~tpt).\partial_t\tilde p_t+\nabla\cdot(v\tilde p_t)=\tfrac12\nabla\cdot\left(a_\theta\tilde p_t\nabla\log\frac{\tilde p_t}{p_t}\right).

The derivative of KL(p~t∣pt)\mathrm{KL}(\tilde p_t\\|p_t) accounts for the motion of both densities. Their common transport terms cancel. For positive smooth densities, finite integrals, and vanishing boundary terms, Proposition 4.2 gives

ddtKL(p~t∣pt)=−12∫p~t(∇log⁡p~tpt)⊤aθ(∇log⁡p~tpt)dx≤0.\frac d{dt}\mathrm{KL}(\tilde p_t\\|p_t)=-\frac12\int\tilde p_t\left(\nabla\log\frac{\tilde p_t}{p_t}\right)^\top a_\theta\left(\nabla\log\frac{\tilde p_t}{p_t}\right)dx\le0.

Each squared coordinate derivative in the integral is weighted by the learned conditional velocity variance at that state, and λ(t)\lambda(t) scales the overall dissipation rate. With λ=0\lambda=0 throughout, both densities follow the same ODE and their relative entropy stays constant. Appendix H.6 illustrates the calculation with exact coefficients.

In submission Figure 7, we vary the overall noise strength and corresponding drift correction with the learned covariance field fixed. Moderate stochasticity improves generation over PFM ODE, while the largest tested strength reverses these gains. This supports choosing the overall strength for the task and step count without retraining. A covariance that is proportional to the identity and depends only on time gives an isotropic sampler. The ImageNet-256 comparisons below test local covariance at matched sampling time, including the cost of variance derivatives (revised Sec. 5.3 and Appendix H.7).

Table 1. ImageNet-256 at a fixed PFM checkpoint

Sampler250-step FIDEqual-time FID
PFM linear-schedule isotropic SDE23.6323.19
PFM covariance-fitted isotropic SDE23.3123.02
PFM SDE22.3422.34

Each coefficient is selected independently with the same validation budget. At 4.01-4.02 ms/sample, the samplers use 283/282/250 steps in the order shown.

The linear-schedule isotropic SDE uses a prescribed scalar schedule. The covariance-fitted isotropic SDE multiplies the same linear time factor by the average covariance trace per dimension and tunes its own coefficient. Each SDE uses the drift correction for its tensor, and the two isotropic tensors have zero spatial divergence. PFM SDE has the lowest FID at 250 steps and at matched sampling time.

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

The paper improved flow matching with a variance head trained jointly with the velocity mean via a diagonal-Gaussian likelihood. The defined an affine change of variables to convert these outputs into the mean and componentwise variance of the reference data-endpoint posterior. The mean corresponds to the usual ODE sampler and the variance defines a state-dependent diffusion tensor for an SDE. In addition, the score and divergence corrections preserve the reference marginals when the mean is exact. The experimental results improve on the ODE results and separate diagnostics evaluate the predicted variances against endpoint residuals.

Critical strengths

(1) The connection between regression and uncertainty is clean. This is a useful qualification when interpreting the uncertainty plots. (2) The experiments are comprehensive. They cover images, point clouds, molecules, and control, with ODE and SDE sampling compared using the same trained model. Figure 6 also checks whether the predicted variances agree with the observed prediction errors on held-out data. The molecular likelihood comparison is especially helpful: it keeps the mean prediction fixed, so any improvement comes from the covariance estimate.

Critical weaknesses

(1) Figure 7 shows that the choice of noise strength matters, but only sweeps the posterior-driven sampler. Table 9 therefore leaves open how much of the gain would remain against a separately tuned isotropic sampler or a time-only scalar variance fitted to the learned covariance. This limits the attribution of the gains to state- and coordinate-dependent uncertainty, even though the reported combined method performs well.

(2) The implemented SDE and its runtime are not specified precisely enough.

Questions for authors

(1) Does the advantage in Table 9 remain after independently tuning the isotropic noise coefficient on the same validation protocol? A time-only isotropic baseline with its scale fitted from the learned covariance would also help distinguish local noise allocation from the overall noise schedule.

(2) Is it stable that the PushT improvements across rollout randomness and training runs? It would be better to add the uncertainty intervals for the results reported.

5 comments
AuthorsOpenReview ↗

1/3 — Does the gain remain after independently tuning the isotropic SDEs?

Thank you for highlighting how PFM connects regression and uncertainty, the range of tasks we evaluate, and the covariance comparison at a fixed mean. We report independently tuned noise controls, full sampling times, and PushT results across training seeds and rollouts.

On ImageNet-256, PFM SDE gives lower FID than both independently tuned isotropic SDEs at 50 and 250 steps and at matched sampling time (revised Sec. 5.3 and Appendix H.7). All samplers in the new comparison use one PFM checkpoint.

The original Table 9 reports ImageNet-256 FIDs of 24.45 for PFM ODE, 24.31 for the isotropic SDE, and 22.28 for PFM SDE at a shared PFM checkpoint. The two SDEs use κ0=λ0\kappa_0=\lambda_0 (submission Appendix H.1), and submission Figure 7 varies the noise strength of PFM SDE. In the new comparison, we tune each SDE independently on the same 12-candidate grid, including zero. We use 10k validation draws per candidate and budget and evaluate the chosen settings on 50k test draws.

For the covariance-fitted isotropic SDE, c(t)=Eval[tr⁡Σθ(Xt,t)]/dc(t)=\mathbb E_{\rm val}[\operatorname{tr}\Sigma_\theta(X_t,t)]/d is the average trace per state dimension of the learned velocity covariance. We estimate it at 101 reference times, weight classes equally, and interpolate linearly. The tensor atime=κ0time(1−t)c(t)Ia_{\rm time}=\kappa_0^{\rm time}(1-t)c(t)I makes the noise strength vary with this average over time, with the same strength across states and coordinates at each time. The linear-schedule isotropic SDE uses aiso=κ0(1−t)Ia_{\rm iso}=\kappa_0(1-t)I, and PFM SDE uses aθ=λ0(1−t)Σθa_\theta=\lambda_0(1-t)\Sigma_\theta. Both isotropic tensors depend only on time, so their spatial divergence is zero. Their drift corrections therefore contain only the score term.

Table 1. ImageNet-256 at fixed steps and matched sampling time

PFM sampler50-step FID250-step FIDEqual-time stepsEqual-time FID
PFM ODE31.4824.4829223.94
PFM linear-schedule isotropic SDE29.6423.6328323.19
PFM covariance-fitted isotropic SDE29.2123.3128223.02
PFM SDE28.6722.3425022.34

The matched-time runs take 4.01-4.02 ms/sample on B200, including derivatives and endpoint readout.

References

Original submission. Posterior Flow Matching, Figure 7, Table 9, and Appendix H.1.

Revised manuscript. Posterior Flow Matching, Sec. 5.3, Appendix H.7.1, and Table 20.

AuthorsOpenReview ↗

2/3 — How do we implement and time PFM SDE?

The original submission gives the SDE in Eq. 19 and its coordinate drift in Appendix C.2 (Eq. 47). Appendix G.2 describes the variance derivatives, G.3 gives the Euler-Maruyama update, and G.5 gives the cutoff and output conventions. Submission Table 10 reports 3.42 ms/sample for PFM ODE and 3.98 ms/sample for PFM SDE on ImageNet-256 with B200.

With γθ=exp⁡ℓθ\gamma_\theta=\exp\ell_\theta, Σθ=diag⁡γθ\Sigma_\theta=\operatorname{diag}\gamma_\theta and aθ=λ(t)Σθa_\theta=\lambda(t)\Sigma_\theta, the sampler is

�0� (Y_t,t)dt+B_\theta(Y_t,t)dW_t,$$ where $B_\theta=\sqrt{\lambda}\operatorname{diag}\sqrt{\gamma_\theta}$ and $\widehat s_\theta=(tv_\theta-x)/(1-t)$ (Eqs. 16-18). For the new measurements, we use $\lambda(t)=\lambda_0(1-t)$. Each positive-noise step estimates the diagonal of the variance field's input Jacobian for the divergence correction using one vector-Jacobian product (VJP). We draw a fresh random vector whose entries independently take $+1$ or $-1$ with equal probability, independently of the Brownian increment. We multiply the VJP result by the same vector elementwise and reuse the forward graph. Zero-strength candidates skip variance derivatives. Appendix G.4 gives the effect of this estimate on the numerical step. We use uniform Euler or Euler-Maruyama to $T=0.99$ and return the endpoint mean from Eq. 15. We time the complete sampler on NVIDIA B200 with batch size 128, BF16 network execution, FP32 moments and coefficients, warmup, and synchronization. The measurement includes random increments, derivatives, and the final readout (revised Appendix H.7). **Table 1. ImageNet-256 sampling time and FID** | PFM sampler | 250-step sampling time | Equal-time steps | Equal-time FID | | --- | --- | --- | --- | | PFM ODE | 3.44 | 292 | 23.94 | | PFM SDE | 4.02 | 250 | 22.34 | Times are in ms/sample. The matched-time settings take 4.02 ms/sample each. Including readout, PFM ODE uses 293 forward passes and PFM SDE uses 251 forward passes plus 250 variance VJPs. ### References **Original submission**. Posterior Flow Matching, Eqs. 15 and 19, Table 10, and Appendices C.2 and G.2-G.5. **Revised manuscript**. Posterior Flow Matching, Appendix H.7.3 for the timing protocol, and Table 20 for the matched-time comparison.
AuthorsOpenReview ↗

3/3 — Are the PushT gains stable across training seeds and rollouts?

At 50 steps, the paired differences favor PFM SDE over PFM ODE and the covariance-fitted isotropic SDE. Both 95% confidence intervals lie above zero (revised Appendix H.5).

The original PushT evaluation reports a control score that averages the highest normalized overlap with the target reached in each test episode, using 50 test initial states (submission Table 2 and Appendix H.5). We now train five independent models per training method and run three rollout repetitions per model at each step budget from the same 50 test initial states. The three PFM samplers share the checkpoint, initial states, and rollout seeds within each training seed. We tune noise strength on separate validation states. The covariance-fitted isotropic SDE uses an isotropic average of the learned covariance at each time, estimated on validation states.

We average the scores over those states for each training seed and rollout repetition, then weight the training seeds equally. We bootstrap training seeds and then rollouts within each seed, keeping the test states fixed and the PFM samplers paired. We resample FM training seeds independently.

Table 1. PushT mean control scores with 95% confidence intervals (CI)

Sampler20 steps [95% CI]50 steps [95% CI]
FM ODE0.823 [0.805, 0.841]0.870 [0.853, 0.887]
PFM ODE0.828 [0.811, 0.845]0.877 [0.861, 0.893]
PFM covariance-fitted isotropic SDE0.836 [0.819, 0.853]0.889 [0.873, 0.905]
PFM SDE0.847 [0.831, 0.863]0.911 [0.897, 0.925]

At 50 steps, the mean paired differences are 0.034 [0.014, 0.054] against PFM ODE and 0.022 [0.006, 0.038] against the covariance-fitted isotropic SDE. At 20 steps they are 0.019 [0.002, 0.036] and 0.011 [-0.004, 0.026], respectively. The latter interval includes zero.

References

Original submission. Posterior Flow Matching, Table 2 and Appendix H.5.

Revised manuscript. Posterior Flow Matching, Appendix H.5 and Table 18.


Thank you again for your time and valuable feedback. We hope the above content can address your concerns. We have incorporated the additional results and clarifications into the revised manuscript. Please let us know if there are further questions.

Reviewer_4YEuOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper adds uncertainty prediction to flow matching. Besides the usual velocity mean, the network has one more head that predicts a componentwise variance of the velocity. Both are trained with one Gaussian NLL objective. With this, the model can tell how much uncertainty about the final sample is left at an intermediate state. The paper also uses the learned variance in a new SDE sampler. The local noise scale follows the predicted variance, and a drift correction keeps the marginals unchanged in theory. The method is tested on four domains.

Critical strengths
  1. The method is simple and easy to use. It only adds one output head and changes the loss. Sampling with the PFM ODE costs the same as with FM.
  2. The SDE is derived with theory. With the score and divergence corrections, it keeps the same marginals as the ODE when the model is exact.
  3. The same method is applied to four very different domains, and the PFM SDE improves over the PFM ODE in all of them.
Critical weaknesses
  1. Novelty is limited. Learning a mean and a variance with a Gaussian NLL is a known idea. EDM2 (Karras et al., 2024) already trains with a learned, time-only uncertainty weight of the same loss form. UA-Flow (Han et al., 2026) already learns heteroscedastic velocity uncertainty for FM. The main new part is using the learned variance to set the SDE noise.
  2. Missing baselines. There is no experimental comparison with UA-Flow or other uncertainty-aware generative models, either on sample quality or on the quality of the uncertainty.
  3. The FM baseline is only sampled with the ODE, and there is no FM + SDE baseline.
  4. The ImageNet setting is limited. The FID is 22–25, which is far from usual ImageNet-256 results. This is likely because there is no CFG. We don’t know if the gain still holds with CFG, with fewer sampling steps, or with larger models.
Questions for authors
  1. Can you compare with UA-Flow and other related uncertainty-aware methods?
  2. The PFM SDE takes 1.16–1.33× the sampling time of the PFM ODE. Can you compare them at the same compute, for example by giving the ODE more steps?
  3. Does the gain still hold on ImageNet with CFG and with fewer sampling steps?
6 comments
AuthorsOpenReview ↗

1/3 — How does PFM compare with UA-Flow and related methods?

Thank you for highlighting the simple training change, the justification for the corrected SDE, and the improvements across four domains. We compare what the methods learn, how well they generate samples and predict endpoints, and how PFM performs with fewer steps.

Several data endpoints can be compatible with the same state Xt=xX_t=x. PFM learns the mean and componentwise variance of the sampled velocity V=X1−X0V=X_1-X_0 at that state with one Gaussian negative log-likelihood (NLL), using Xt=(1−t)X0+tX1X_t=(1-t)X_0+tX_1 with independent endpoints. The affine relation X1=Xt+(1−t)VX_1=X_t+(1-t)V gives the reference endpoint-posterior moments in one evaluation (Eq. 15). The velocity covariance Σθ\Sigma_\theta sets the relative noise strengths across states and coordinates (Eqs. 17-18), and Eq. 19 gives the corresponding score and divergence drift corrections. Proposition 4.1 and Appendix B.4 give the population moments and the effect of variance bounds.

UA-Flow [1] targets errors in predicting the marginal velocity and estimates the correction from a minibatch. It propagates that uncertainty to the endpoint and uses variance gradients for guidance. In our matched comparison, UA-Flow uses FM pretraining followed by joint fine-tuning of the backbone and uncertainty head. PFM fits the mean and variance in a single training stage using the sampled reference pairs. The conditional variance of the sampled reference velocities can remain nonzero even when the predicted mean is exact (revised Sec. 2).

EDM2 [2] uses a related likelihood to learn a training weight that depends on noise level. Original Sec. 2 also discusses Flowing with Confidence [3], which propagates input-dependent network noise along trajectories.

In the equal-time generation comparisons, PFM SDE gives FID 3.05 versus 3.19 for UA-Flow with uncertainty-aware guidance on CIFAR-10 (Appendix H.2), and 22.34 versus 23.25 on ImageNet-256 (Sec. 5.1). Both comparisons use matched mean backbones and training-example budgets.

PFM gives lower calibrated diagonal-Gaussian endpoint NLL at all three CIFAR-10 reference times (Sec. 5.2 and Appendix H.2). Both methods use the same states at t=0.2,0.5,0.8t=0.2,0.5,0.8, constructed from independent Gaussian noise and data endpoints. PFM reads out μθ=x+(1−t)vθ\mu_\theta=x+(1-t)v_\theta and Σθx=(1−t)2Σθ\Sigma_\theta^x=(1-t)^2\Sigma_\theta in one evaluation. UA-Flow uses 101 forward passes and at most 50 covariance Jacobian-vector products to propagate its estimates (Appendix H.2). Each method keeps its own unclipped endpoint mean. We floor endpoint variance at 10−1210^{-12}, fit one positive multiplier per method and time on 5k calibration images, and freeze it for 5k test images.

Table 1. Calibrated endpoint prediction on CIFAR-10

ttRMSE ratioUA-Flow Δ\DeltaNLLPFM Δ\DeltaNLLPFM minus UA-Flow NLL [95% CI]
0.21.210.144-0.047-0.191 [-0.218, -0.163]
0.51.220.127-0.072-0.199 [-0.227, -0.171]
0.81.180.099-0.067-0.166 [-0.193, -0.138]

Root-mean-square error (RMSE) measures the error between each method's predicted endpoint mean and the known data endpoint at the same reference states. The table reports UA-Flow's RMSE divided by PFM's; values above one favor PFM. Calibrated Δ\DeltaNLL is per dimension relative to the calibrated residual-fitted isotropic covariance at the PFM mean. CI denotes confidence interval. We obtained 95% confidence intervals by resampling paired test images, with calibration fixed.

The ImageNet-256 endpoint test uses the same JiT-B/16 architecture, matched reference states, and each image's class label (revised Appendix H.2).

Table 2. ImageNet-256 endpoint prediction

ttRMSE ratio ↑\uparrowNLL difference ↓\downarrow
0.21.19-0.174
0.51.22-0.199
0.81.25-0.223

The RMSE ratio is UA-Flow/PFM; values above one favor PFM. The calibrated NLL difference is PFM minus UA-Flow per dimension; lower values favor PFM.

References

  • [1] Han et al. Flow Matching with Uncertainty Quantification and Guidance. arXiv preprint arXiv:2602.10326, 2026.
  • [2] Karras et al. Analyzing and Improving the Training Dynamics of Diffusion Models. CVPR, 2024.
  • [3] de Kruiff et al. Flowing with Confidence. arXiv preprint arXiv:2605.18472, 2026.
AuthorsOpenReview ↗

2/3 — How do the samplers compare at the same sampling time?

Giving PFM ODE more steps at the same sampling time still leaves PFM SDE with lower FID on CIFAR-10 and ImageNet-256. PFM SDE also gives lower FID than UA-Flow [1] with uncertainty-aware guidance (UG) on both datasets. On ImageNet-256, it gives lower FID than the isotropic SDE using a separately trained FM checkpoint. The original submission defines that FM sampler in Table 1 and Appendix A.6 (Eq. 32), with the numerical update in Appendix G.3 (Eq. 82). Revised Sec. 5.1 reports its measured result in Table 6.

The methods use DDPM++ and 800 epochs on CIFAR-10, and pixel-space JiT-B/16 and 600 epochs on ImageNet-256. The matched training-example budget counts both FM pretraining and fine-tuning for UA-Flow. We use uniform Euler for ODEs and Euler-Maruyama for SDEs, with T=0.99T=0.99 and endpoint-mean readout. UA-Flow + UG applies guidance every second Euler step, starting at step 2. For samplers with a noise or guidance strength, we compare 12 candidate values, including zero, under equal validation budgets. We set step counts by complete sampling time, including active derivatives and readout, and freeze strengths and step counts for the 50k-image evaluation. Within each dataset, we compute FID the same way for all methods, with classifier-free guidance and sample filtering disabled (Appendices H.1, H.2, and H.7).

Table 1. CIFAR-10 generation at equal sampling time

MethodStepsFIDSampling time (relative)
FM ODE643.581.27
UA-Flow ODE623.481.27
UA-Flow + UG543.191.27
PFM ODE633.261.27
PFM SDE503.051.27

Time is relative to the 50-step FM ODE.

Table 2. ImageNet-256 generation at equal sampling time

MethodStepsFIDSampling time (ms/sample)
FM ODE29124.914.01
FM isotropic SDE28223.844.03
UA-Flow ODE28824.134.02
UA-Flow + UG26623.254.02
PFM ODE29223.944.02
PFM SDE25022.344.02

The NVIDIA B200 times differ by less than 1%. PFM ODE at 292 steps and PFM SDE at 250 steps both take 4.02 ms/sample.

References

  • [1] Han et al. Flow Matching with Uncertainty Quantification and Guidance. arXiv preprint arXiv:2602.10326, 2026.
AuthorsOpenReview ↗

3/3 — Does the gain hold with CFG and fewer steps?

On ImageNet-256 with JiT-B/16, PFM SDE gives lower FID than PFM ODE at both 50 and 250 steps across the tested guidance settings (revised Sec. 5.1 and Appendix H.2). We hold the backbone size fixed to compare guidance and step count.

All methods use pixel-space JiT-B/16 (132.90M parameters for PFM, submission Table 7), 600 epochs, effective batch size 1,024, and label dropout of 0.1. The original comparison (submission Table 4) used the same backbone with zero label dropout. We select one exponential moving-average checkpoint for each of diffusion, FM, and PFM on the same validation setting at guidance scale 3 and 250 steps. For PFM, we select it with the ODE and use it for both samplers and all guidance scales. We tune the noise strength on separate validation draws. For each method, we compute FID in the same way, using 50k class-balanced images and the same initial seeds.

FM and PFM use classifier-free guidance (CFG) in velocity, vCFG=v∅+wCFG(vc−v∅)v_{\mathrm{CFG}}=v_{\varnothing}+w_{\mathrm{CFG}}(v_c-v_{\varnothing}), where vcv_c is the conditional velocity for class cc, v∅v_{\varnothing} is the unconditional velocity, and wCFGw_{\mathrm{CFG}} is the guidance scale. Diffusion uses CFG in its score. Following JiT [1], the chosen scale is active on 0.1<t<10.1 < t < 1 and equals 1 outside that interval. Setting wCFG=1w_{\mathrm{CFG}}=1 gives ordinary class-conditional generation using the same trained models.

Each mean network predicts clean images. FM and PFM train with velocity mean squared error and Gaussian negative log-likelihood, respectively. Diffusion converts the clean-image output to a noise prediction and score on its variance-preserving path and trains with noise-space mean squared error. We use uniform Euler or Euler-Maruyama to T=0.99T=0.99 and then read out the endpoint mean. Substituting vCFGv_{\mathrm{CFG}} in submission Eq. 8 gives the guided score estimate s^CFG=(tvCFG−x)/(1−t)\widehat s_{\mathrm{CFG}}=(t v_{\mathrm{CFG}}-x)/(1-t) (revised Appendix E.3). PFM SDE sets aθ=λ(t)Σθa_\theta=\lambda(t)\Sigma_\theta using the conditional velocity covariance for class cc and keeps both drift corrections in submission Eq. 19.

Table 1. ImageNet-256 FID with CFG and fewer steps

StepswCFGw_{\mathrm{CFG}}DiffusionFM ODEPFM ODEPFM SDE
50135.1233.2431.7728.81
5026.746.125.845.39
5035.214.584.364.09
250127.5525.6124.6722.53
25025.915.365.134.79
25034.433.963.783.57

The table fixes the step count in each row. PFM SDE also evaluates variance derivatives.

The ImageNet-256 comparison at a matched sampling-time budget tests ordinary CFG and UA-Flow's native uncertainty-aware classifier-free guidance (UCFG) [2] under the same validation budgets. PFM ODE and PFM SDE share a checkpoint (revised Appendix H.2).

Table 2. ImageNet-256 guided generation at a matched sampling-time budget

SamplerStepsFID ↓\downarrow
UA-Flow + CFG2704.2
UA-Flow + UCFG2604.0
PFM ODE + CFG2703.8
PFM SDE + CFG2503.6

All rows use a sampling-time budget of 7.1 ms/sample. CFG is classifier-free guidance. UCFG is UA-Flow's uncertainty-aware classifier-free guidance. Complete sampling time includes both conditioning branches, active derivatives, and endpoint readout. Lower FID is better.

References

  • [1] Li and He. Back to Basics: Let Denoising Generative Models Denoise. CVPR, 2026.
  • [2] Han et al. Flow Matching with Uncertainty Quantification and Guidance. arXiv preprint arXiv:2602.10326, 2026.

Thank you again for your time and valuable feedback. We hope the above content can address your concerns. We have incorporated the additional results and clarifications into the revised manuscript. Please let us know if there are further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.