acceptodds
Under review as a conference paper at ICLR 2027

Target-Guided Flow Matching

Abstract

Flow Matching has become an important generative modeling framework, which is trained by regressing the vector fields of fixed conditional probability paths. We prove that the probability path is jointly determined by the source distribution and the velocity field. This result motivates designing a source distribution that captures the statistical structure of the target distribution to improve the generative performance of Flow Matching. To this end, we propose Target-Guided Flow Matching (TGFlow), which utilizes the marginal second-order statistics (i.e., the variance of each dimension) of the target data to construct an initial source distribution. TGFlow preserves the analytical forms of the conditional velocity field and optimization objective of Flow Matching, requires no additional training costs, and shortens the average length of actual generation trajectories, bringing the generated distribution closer to the target distribution. Experiments on 2D synthetic datasets and high-dimensional image datasets such as CIFAR-10 and ImageNet demonstrate that TGFlow improves generation quality while reducing NFE. We further verify that TGFlow supports various samplers and can be extended to a broader range of source distributions.

Primary area:
generative models
Keywords:
Flow Matching;Generative Model;Machine Learning

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

This paper introduces Target-Guided Flow Matching (TGFlow) that modifies the source distribution from an isotropic Gaussian to honor the second-order statistics (e.g., covariance structure) of target distribution. It theoretically shows editing the source can affect the learning of the velocity field, and empirical evidences on CIFAR-10 and ImageNet-{32,64} show lower Frechet Inception Distance (FID) score and Neural Function Evaluation (NFE) than typical diffusion / flow matching methods.

Critical strengths
  • Since flow matching gives the freedom to choose an arbitrary source distribution, the line of thinking to design a source that's better align with the target is intuitive. The authors, in their writing, unveil the intuition very nicely.
  • The authors also identify the limitations of prior source-design works and propose a solution to them.
Critical weaknesses
  • Both theorem 1 and 2 are restating known flow-matching fundamentals from Lipman et al. (2024) and Lai et al. (2025).
  • Theorem 3 is trivial as it follows σmin2Λ≤σmin2I\sigma_{min}^2\Lambda \leq \sigma_{min}^2I with standard Gaussian KL. Also, it bounds for a untrained mental-model network, not really the KL between a learned/predicted target distribution and the actual target distribution.
  • Section 3.3 is not really discussing any insights of extending TGFlow to any arbitrary source distributions as the section name suggests; instead, it reads more like an introduction of data coupling via optimal transport.
  • The extension to non-Gaussian source distribution is an overclaim. There is no conceptual illustration of how TGFlow works beyond non-Gaussian source distributions (e.g., heavy-tail, exponential), and there is only one toy experiment where the source distribution is "moons" (with limited souruce estimation details), which obviously falls short to support the claim.
  • There is no other baseline methods of non-Gaussian source distributions in the experiments, not even any of the source-design methods the paper cites; besides, another reasonable baseline should be N∼(0,βIN\sim(0, \beta I) where β\beta is a rescaling factor to matched the TGFlow's estimated variance level (i.e., match in total energy) so that the readers can know the benefits come from the scale or the shape.

References

  • Lipman et al. (2024) Flow Matching Guide and Code.
  • Lai et al. (2025) The Principles of Diffusion Models
Questions for authors

Please refer to the critical weaknesses that I raised. I think the most important baseline shall be the matched variance-level rescaled isotropic Gaussian, while benefits over other source-design methods are secondary. The remaining concerns are conceptual. There is one more conceptual thing that I want to point out which is the practicality of obtaining second-order statistics in a reliable manner for high-dimensional data, whose difficulty might be not easier than generative modeling itself.

11 comments

Weakness 1: Both theorem 1 and 2 are restating known flow-matching fundamentals from Lipman et al. (2024) and Lai et al. (2025).

We thank the reviewer for the detailed comments. We agree that the original Theorems 1 and 2 build upon established theoretical foundations, including ODE flows, the continuity equation, and affine conditional flow matching. Their primary contribution lies in providing theoretical insights into our method, rather than introducing entirely new theoretical results. Specifically, their contributions are twofold: the original Theorem 1 clarifies the dependence of probability paths on the source distribution, whereas the original Theorem 2 establishes that TGFlow’s source distribution adjustment preserves the analytical forms of the existing conditional velocity field and training objective. To accurately reflect the relationship between these results and prior work, we will no longer present them as standalone theorems in the revised manuscript. Instead, we will present the original Theorem 1 as a Remark based on reference [1] and the original Theorem 2 as a Corollary based on existing reference [2].

[1] Lipman, Yaron, et al. "Flow Matching for Generative Modeling." ICLR 2023 (2023).

[2] Lai, Chieh-Hsin, et al. "The Principles of Diffusion Models." CoRR abs/2510.21890 (2025).

Weakness 2: Theorem 3 is trivial as it follows σmin⁡2Λ≤σmin⁡2I\sigma_{\min}^{2}\Lambda \leq \sigma_{\min}^{2}I with standard Gaussian KL. Also, it bounds for a untrained mental-model network, not really the KL between a learned/predicted target distribution and the actual target distribution.

We thank the reviewer for pointing out this issue. We agree that, under the Gaussian target distribution assumption, Theorem 3 follows from the standard Gaussian KL divergence formula and the covariance ordering relation σmin⁡2Λ≤σmin⁡2I\sigma_{\min}^{2}\Lambda \leq \sigma_{\min}^{2}I. Nevertheless, this result provides an analytical explanation of the benefit of constructing the source distribution using the target’s coordinate-wise variances: under these assumptions, the KL divergence between the terminal marginal distribution of TGFlow’s ideal probability path and the target distribution is no greater than the corresponding KL divergence obtained with the standard isotropic source. We also acknowledge that this result does not account for velocity-field learning error and therefore does not fully characterize the discrepancy between the actual generated distribution and the target distribution. To address this limitation, following the suggestion of Reviewer 1, we have added two theoretical results based on the stability analysis framework in reference [3]. These results show that, under the assumptions explicitly stated in the two new theorems, TGFlow’s Wasserstein-2 generation-error upper bound is no greater than the corresponding bounds obtained with either the standard isotropic source or the isotropic source with the same total variance as TGFlow’s source, and is strictly smaller under the specified conditions.

Assumption A1 (Moment conditions on the target distribution). The target distribution qq is a probability distribution on Rd\mathbb{R}^{d} with finite second moments and strictly positive marginal variances in all coordinates, that is,

EY∼q∣Y∣2<∞,di:=Var⁡Y∼q(Yi)>0,i=1,…,d.\mathbb{E}_{Y\sim q}\\| Y\\|^{2} < \infty,\quad\quad d_{i}:= \operatorname{Var}_{Y\sim q}(Y_{i}) > 0,\quad i = 1,\ldots,d.

Define

D=diag⁡(d1,…,dd),M=max⁡idi,Λ=DM,c=tr⁡(Λ)d.D = \operatorname{diag}(d_{1},\ldots,d_{d}),\quad\quad M = \max_{i}d_{i},\quad\quad\Lambda = \frac{D}{M},\quad\quad c = \frac{\operatorname{tr}(\Lambda)}{d}.

Here, DD collects the marginal variances and need not equal the full covariance matrix of q.q.

For each Σ∈Λ,I,cI,\Sigma \in \\{\Lambda,I,cI\\}, independently sample

X0Σ∼N(0,Σ),Y∼q,X_{0}^{\Sigma}\sim\mathcal{N}(0,\Sigma),\quad\quad Y\sim q,

and define the interpolation path

XtΣ=σtX0Σ+tY,t∈[0,1],X_{t}^{\Sigma} = \sigma_{t}X_{0}^{\Sigma} + tY,\quad\quad t \in \lbrack 0,1\rbrack,

where

σt=1−at,a=1−σmin⁡,0<σmin⁡<1.\sigma_{t} = 1 - at,\quad\quad a = 1 - \sigma_{\min},\quad\quad 0 < \sigma_{\min} < 1.

Let

ptΣ=Law⁡(XtΣ),p_{t}^{\Sigma} = \operatorname{Law}(X_{t}^{\Sigma}),

and define the ideal marginal velocity field and its smallest global spatial Lipschitz constant by

vtΣ(x)=E[Y−aX0Σ∣XtΣ=x],LtΣ=Lip⁡x(vtΣ).v_{t}^{\Sigma}(x) = \mathbb{E}\lbrack Y - aX_{0}^{\Sigma}|X_{t}^{\Sigma} = x\rbrack,\quad\quad L_{t}^{\Sigma} = \operatorname{Lip}_{x}(v_{t}^{\Sigma}).

Next, introduce the conditional covariance matrix and its Gaussian reference matrix:

CtΣ(x)=Cov⁡(Y∣XtΣ=x),GtΣ=(D−1+t2σt2Σ−1)−1.C_{t}^{\Sigma}(x) = \operatorname{Cov}(Y|X_{t}^{\Sigma} = x),\quad\quad G_{t}^{\Sigma} = \left( D^{- 1} + \frac{t^{2}}{\sigma_{t}^{2}}\Sigma^{- 1} \right)^{- 1}.

The above conditional moments are understood as the canonical versions induced by the Gaussian observation kernel.

Assumption A2 (Relative Deviation Bound for Conditional Covariances). There exists a common constant ζ≥0\zeta \geq 0 such that, for all x∈Rd,x \in \mathbb{R}^{d}, t∈[0,1],t \in \lbrack 0,1\rbrack, and Σ∈Λ,I,cI,\Sigma \in \\{\Lambda,I,cI\\},

\left\\| (G_{t}^{\Sigma})^{- 1/2}C_{t}^{\Sigma}(x)(G_{t}^{\Sigma})^{- 1/2} - I \right\\|_{op} \leq \zeta.\quad\quad\text{(R)}

To state the theorem, define the reference Lipschitz constant and its time integral by

ℓtΣ=max⁡i∣tdi−aσtΣiiσt2Σii+t2di∣,JΣ=∫01ℓtΣ,dt,\ell_{t}^{\Sigma} = \max_{i}\left| \frac{td_{i} - a\sigma_{t}\Sigma_{ii}}{\sigma_{t}^{2}\Sigma_{ii} + t^{2}d_{i}} \right|,\quad\quad J_{\Sigma} = \int_{0}^{1}{\ell_{t}^{\Sigma}\\, dt,}

and

κΣ=λmax⁡(Σ)λmin⁡(Σ),rΣ=max⁡idiΣii,\kappa_{\Sigma} = \frac{\lambda_{\max}(\Sigma)}{\lambda_{\min}(\Sigma)},\quad\quad r_{\Sigma} = \max_{i}\frac{d_{i}}{\Sigma_{ii}},

HΣ=κΣ2log⁡(1+rΣσmin⁡2).H_{\Sigma} = \frac{\sqrt{\kappa_{\Sigma}}}{2}\log\left( 1 + \frac{r_{\Sigma}}{\sigma_{\min}^{2}} \right).

Theorem 1 (Comparison of Time-Integrated Lipschitz Constants Across Source Distributions). Under the above setting and Assumptions A1–A2, the ideal marginal velocity fields satisfy

∣∫01LtΣ,dt−JΣ∣≤ζHΣ,Σ∈Λ,I,cI.\left| \int_{0}^{1}{L_{t}^{\Sigma}\\, dt - J_{\Sigma}} \right| \leq \zeta H_{\Sigma},\quad\quad\Sigma \in \\{\Lambda,I,cI\\}.

If the target's marginal variances are not all equal, then

ΔB:=JB−JΛ>0,B∈I,cI.\Delta_{B}:= J_{B} - J_{\Lambda} > 0,\quad\quad B \in \\{ I,cI\\}.

If additionally

ζ(HΛ+HB)<ΔB,B∈I,cI,(C)\zeta(H_{\Lambda} + H_{B}) < \Delta_{B},\quad\quad B \in \\{ I,cI\\},\quad\quad\text{(C)}

then

\int_{0}^{1}{L_{t}^{\Lambda}\\, dt < \min\left\\{ \int_{0}^{1}{L_{t}^{I}\\, dt,\\,\int_{0}^{1}{L_{t}^{cI}\\, dt}} \right\\}.}

Proof. Let

mtΣ(x)=E[Y∣XtΣ=x].m_{t}^{\Sigma}(x) = \mathbb{E}\lbrack Y|X_{t}^{\Sigma} = x\rbrack.

Since XtΣ=σtX0Σ+tYX_{t}^{\Sigma} = \sigma_{t}X_{0}^{\Sigma} + tY and σt+at=1,\sigma_{t} + at = 1,

vtΣ(x)=mtΣ(x)−axσt.v_{t}^{\Sigma}(x) = \frac{m_{t}^{\Sigma}(x) - ax}{\sigma_{t}}.

For t>0,t > 0, differentiating the conditional mean under the Gaussian observation model gives

∇xmtΣ(x)=tσt2CtΣ(x)Σ−1.\nabla_{x}m_{t}^{\Sigma}(x) = \frac{t}{\sigma_{t}^{2}}C_{t}^{\Sigma}(x)\Sigma^{- 1}.

Consequently,

∇xvtΣ(x)=tσt3CtΣ(x)Σ−1−aσtI.\nabla_{x}v_{t}^{\Sigma}(x) = \frac{t}{\sigma_{t}^{3}}C_{t}^{\Sigma}(x)\Sigma^{- 1} - \frac{a}{\sigma_{t}}I.

Define

AtΣ=(tD−aσtΣ)(σt2Σ+t2D)−1.A_{t}^{\Sigma} = (tD - a\sigma_{t}\Sigma)(\sigma_{t}^{2}\Sigma + t^{2}D)^{- 1}.

Because DD and Σ\Sigma are diagonal,

AtΣ=tσt3GtΣΣ−1−aσtI,∣AtΣ∣op=ℓtΣ.A_{t}^{\Sigma} = \frac{t}{\sigma_{t}^{3}}G_{t}^{\Sigma}\Sigma^{- 1} - \frac{a}{\sigma_{t}}I,\quad\quad\\| A_{t}^{\Sigma}\\|_{op} = \ell_{t}^{\Sigma}.

Thus,

∇xvtΣ(x)−AtΣ=tσt3(CtΣ(x)−GtΣ)Σ−1.\nabla_{x}v_{t}^{\Sigma}(x) - A_{t}^{\Sigma} = \frac{t}{\sigma_{t}^{3}}\left( C_{t}^{\Sigma}(x) - G_{t}^{\Sigma} \right)\Sigma^{- 1}.

Condition (R) implies

−ζGtΣ⪯CtΣ(x)−GtΣ⪯ζGtΣ.- \zeta G_{t}^{\Sigma} \preceq C_{t}^{\Sigma}(x) - G_{t}^{\Sigma} \preceq \zeta G_{t}^{\Sigma}.

Multiplying on both sides by Σ−1/2,\Sigma^{- 1/2}, we obtain

\left\\| \Sigma^{- 1/2}\left( C_{t}^{\Sigma}(x) - G_{t}^{\Sigma} \right)\Sigma^{- 1/2} \right\\|_{op} \leq \zeta\max_{i}\frac{(G_{t}^{\Sigma})_{ii}}{\Sigma_{ii}}.

It follows that

\left\\| \left( C_{t}^{\Sigma}(x) - G_{t}^{\Sigma} \right)\Sigma^{- 1} \right\\|_{op} \leq \zeta\sqrt{\kappa_{\Sigma}}\max_{i}\frac{(G_{t}^{\Sigma})_{ii}}{\Sigma_{ii}}.

Moreover,

max⁡i(GtΣ)iiΣii=σt2rΣσt2+t2rΣ.\max_{i}\frac{(G_{t}^{\Sigma})_{ii}}{\Sigma_{ii}} = \frac{\sigma_{t}^{2}r_{\Sigma}}{\sigma_{t}^{2} + t^{2}r_{\Sigma}}.

Therefore,

\left\\| \nabla_{x}v_{t}^{\Sigma}(x) - A_{t}^{\Sigma} \right\\|_{op} \leq \zeta\sqrt{\kappa_{\Sigma}}\frac{tr_{\Sigma}}{\sigma_{t}(\sigma_{t}^{2} + t^{2}r_{\Sigma})}.

For a continuously differentiable velocity field, the smallest global spatial Lipschitz constant equals the supremum of the operator norm of its Jacobian. Hence,

∣LtΣ−ℓtΣ∣≤ζκΣtrΣσt(σt2+t2rΣ).|L_{t}^{\Sigma} - \ell_{t}^{\Sigma}| \leq \zeta\sqrt{\kappa_{\Sigma}}\frac{tr_{\Sigma}}{\sigma_{t}(\sigma_{t}^{2} + t^{2}r_{\Sigma})}.

At t=0,t = 0, independence gives

v0Σ(x)=Eq[Y]−ax,L0Σ=ℓ0Σ=a.v_{0}^{\Sigma}(x) = \mathbb{E}_{q}\lbrack Y\rbrack - ax,\quad\quad L_{0}^{\Sigma} = \ell_{0}^{\Sigma} = a.

Using u=t/σt,u = t/\sigma_{t}, with du=dt/σt2,du = dt/\sigma_{t}^{2}, yields

\int_{0}^{1}{\frac{tr_{\Sigma}}{\sigma_{t}(\sigma_{t}^{2} + t^{2}r_{\Sigma})}\\, dt} = \int_{0}^{1/\sigma_{\min}}{\frac{ur_{\Sigma}}{1 + u^{2}r_{\Sigma}}\\, du} \\\\ = \frac{1}{2}\log\left( 1 + \frac{r_{\Sigma}}{\sigma_{\min}^{2}} \right). \end{array}$$ Continued in the next reply.

Hence,

∣∫01LtΣ,dt−JΣ∣≤∫01∣LtΣ−ℓtΣ∣,dt≤ζHΣ.\left| \int_{0}^{1}{L_{t}^{\Sigma}\\, dt - J_{\Sigma}} \right| \leq \int_{0}^{1}{|L_{t}^{\Sigma} - \ell_{t}^{\Sigma}|\\, dt \leq \zeta H_{\Sigma}.}

To establish positivity of the reference gaps, define

ft(r)=tr−aσtσt2+t2r.f_{t}(r) = \frac{tr - a\sigma_{t}}{\sigma_{t}^{2} + t^{2}r}.

For t>0,t > 0,

∂ft(r)∂r=tσt(σt2+t2r)2>0.\frac{\partial f_{t}(r)}{\partial r} = \frac{t\sigma_{t}}{(\sigma_{t}^{2} + t^{2}r)^{2}} > 0.

The reference quantities satisfy

ℓtΛ=∣ft(M)∣,ℓtI=max⁡i∣ft(di)∣,ℓtcI=max⁡i∣ft(di/c)∣.\ell_{t}^{\Lambda} = |f_{t}(M)|,\quad\quad\ell_{t}^{I} = \max_{i}|f_{t}(d_{i})|,\quad\quad\ell_{t}^{cI} = \max_{i}|f_{t}(d_{i}/c)|.

Since di=Md_{i} = M for at least one coordinate,

ℓtI≥ℓtΛ.\ell_{t}^{I} \geq \ell_{t}^{\Lambda}.

If the marginal variances are not all equal, let dmin⁡=min⁡idi<M.d_{\min} = \min_{i}d_{i} < M. Since f0(r)=−a<0,f_{0}(r) = - a < 0, we have

ft(dmin⁡)<ft(M)<0f_{t}(d_{\min}) < f_{t}(M) < 0

for all sufficiently small t>0.t > 0. Thus, ℓtI>ℓtΛ\ell_{t}^{I} > \ell_{t}^{\Lambda} on an interval of positive length, implying ΔI>0.\Delta_{I} > 0.

Also, writing d‾=d−1∑idi,\overline{d} = d^{- 1}\sum_{i}^{}{d_{i},} we have c=d‾/Mc = \overline{d}/M and

dmin⁡c<M<Mc.\frac{d_{\min}}{c} < M < \frac{M}{c}.

By strict monotonicity of ft,f_{t}, for every t>0,t > 0,

ft(dmin⁡/c)<ft(M)<ft(M/c).f_{t}(d_{\min}/c) < f_{t}(M) < f_{t}(M/c).

Thus, at least one of the two endpoint values has an absolute value strictly greater than ∣ft(M)∣.|f_{t}(M)|. Consequently,

ℓtcI>ℓtΛ,t>0,\ell_{t}^{cI} > \ell_{t}^{\Lambda},\quad\quad t > 0,

and hence

ΔcI=JcI−JΛ>0.\Delta_{cI} = J_{cI} - J_{\Lambda} > 0.

Finally, for each B∈I,cI,B \in \\{ I,cI\\}, condition (C) gives

\int_{0}^{1}{L_{t}^{\Lambda}\\, dt} \leq J_{\Lambda} + \zeta H_{\Lambda} \\\\ < J_{B} - \zeta H_{B} \\\\ \leq \int_{0}^{1}{L_{t}^{B}\\, dt.} \end{array}$$ Therefore, $$\int_{0}^{1}{L_{t}^{\Lambda}\\, dt < \min\left\\{ \int_{0}^{1}{L_{t}^{I}\\, dt,\\,\int_{0}^{1}{L_{t}^{cI}\\, dt}} \right\\}.}$$ This completes the proof. $\square$ Theorem 1 concerns ideal velocity fields, whereas the stability bound in \[1\] depends on a spatial Lipschitz upper bound for the learned velocity field. Condition (R) controls deviations of the target’s conditional second moments from the Gaussian reference, while condition (C) ensures that these deviations do not eliminate the comparison advantage of the TGFlow source in the reference setting. To connect this ideal-field analysis to learned models, we impose an additional regularity condition beyond Assumptions 1–3 in \[3\]: across different source distributions, the Lipschitz constant of the learned velocity field does not exceed that of the corresponding ideal velocity field plus a common nonnegative, integrable remainder. Under this condition and a common upper bound on the population velocity-field approximation error, the stability argument in \[3\], together with the terminal residual term, yields a Wasserstein-2 generation-error bound and allows us to compare the resulting upper bounds across source distributions.
Reviewer_EwU8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Targeted-Guided Flow Matching (TGFlow), which replaces the original Gaussian source in standard flow matching with a Gaussian whose diagonal covariance is constructed from the coordinate-wise variances from the target data, The paper reports improved generation quality and inference efficiency on CIFAR-10 and ImageNet datasets. It also analyzes the conditional flow construction and prove a terminal-distribution KL inequality for a Gaussian target under an idealized flow. In summary, this paper makes both theoretical and empricical contributions on the topic of flow matching.

Critical strengths

S1: TGF is a simple and useful source design question

TGFlow asks whether statistics already available from the training data can help improve the choice of source distribution without the need of learning another separate source model. The resulting construction is inexpensive and can be incorporated within the standard flow matching training framework with ease.

S2: Evidence across Datasets

The reported gains extend beyond toy examples to CIFAR-10 and ImageNet-32/64/128. The paper also investigates other related topics such as likelihood, sampling steps, diversity, and different ODE samplers.

Critical weaknesses

W1: Lack of comparison with recent work

The paper shows that TGFlow improves over a standard Gaussian source, but does not establish whether its target-guided diagonal covariance is competitive with other ways of choosing the source distribution. It is suggested that the authors include more recent methods, such as SourceFM (Lee et al., 2026a), MM-FM (Luo et al., 2026), CAR-Flow (Chen et al., 2025), and CSFM (Kim et al., 2026). The authors can explain how TGFlow differs from these methods and provide controlled experimental comparisons. Without these comparisons, the relative value of TGFlow’s source-design choice remains unclear.

W2: The experiments do not isolate the benefit of coordinate-wise target information. TGFlow changes both the overall scale and the coordinate-wise shape of the source covariance. As a result, the comparison with N(0,I)N(0,I) does not establish whether the improvement indeed comes from matching individual target variances. A source N(0,cI)N(0,cI) with c=tr⁡(Λ)/dc=\operatorname{tr}(\Lambda)/d, with per-channel and coordinate-shuffled covariance controls, would test this central claim. Theorem 3 does not resolve the empirical question: it concerns an ideal terminal distribution for a Gaussian target, while the image results depend on a learned velocity field. Image generation measurements of transport length would also substantiate the proposed mechanism.

References

[1] SourceFM: Lee, J., Kim, K., & Lee, J. (2026). “Is There a Better Source Distribution than Gaussian? Exploring Source Distributions for Image Flow Matching.” Transactions on Machine Learning Research.

[2] MM-FM: Luo, G., Cole, F., Zhang, S., Wan, Y., Lu, Y., & Sun, J. (2026). “Flow Matching for Multimodal Distributions.” CVPR 2026.

[3] CAR-Flow: Chen, C., Guo, P., Song, L., Lu, J., Qian, R., Fu, T.-J., Wang, X., Liu, W., Yang, Y., & Schwing, A. G. (2025). “CAR-Flow: Condition-Aware Reparameterization Aligns Source and Target for Better Flow Matching.” NeurIPS 2025.

[4] CSFM: Kim, J., Park, J., Jeon, S., & Kim, S. (2026). “Better Source, Better Flow: Learning Condition-Dependent Source Distribution for Flow Matching.” arxiv 2026.

Questions for authors

Q1: Does the main 2D comparison isolate the source distribution?

Appendix E.2 states that the baseline uses a Brownian-bridge conditional path with σ=1\sigma=1, but TGFlow uses the path from Section 3. Which Table 1 results use identical paths and couplings? Could you compare the standard and target-guided sources while holding both fixed?

Q2: Does the proposed non-Gaussian path start from its stated source?

In Section 3.3, Equation (21) gives T0(x)=σx+x0T_0(x)=\sigma x+x_0. For σ>0\sigma>0, this appears to start from a noise-convolved version of the distribution of x0x_0, rather than that distribution itself. What value of σ\sigma is used, and what is the actual initial marginal in the non-Gaussian experiments?

Q3: How should the training budgets in Table 6 be interpreted?

For CIFAR-10, 391K iterations at effective batch size 256 correspond to approximately 2,000 passes over 50K images, but the table reports 1,000 epochs. For ImageNet-128, 500K iterations at batch size 1,536 likewise appear inconsistent with 80 epochs. Which figures describe the runs used for the reported results, and were the FM and TGFlow models trained on the same number of images?

10 comments

Q1: Does the main 2D comparison isolate the source distribution? Appendix E.2 states that the baseline uses a Brownian-bridge conditional path with σ=1\sigma = 1, but TGFlow uses the path from Section 3. Which Table 1 results use identical paths and couplings? Could you compare the standard and target-guided sources while holding both fixed?

We thank the reviewer for this valuable suggestion. We agree that evaluating the effect of the source distribution under the same conditional path and coupling method would help provide a clearer understanding of its contribution. In the non-Gaussian experiments, TGFlow used a zero-noise linear interpolation path, whereas the baselines used Brownian bridge conditional paths. Therefore, the original comparison did not fully control for differences in conditional paths. To isolate the effect of source distribution adjustment, we have added controlled experiments in which both the baselines and TGFlow use zero-noise linear interpolation paths (σ=0\sigma = 0), with the coupling method kept the same and only the construction of the source distribution varied. The experimental results are presented below. We will update the experimental results table in the revised manuscript to include these new results.

Wasserstein-2 Distance Results (↓)

Method / OT path variantN→8gaussiansmoons→ 8gaussiansN→moonsN→scurve
FM0.409 ± 0.080—0.180 ± 0.0240.169 ± 0.027
I-CFM0.344 ± 0.1031.509 ± 0.2480.153 ± 0.0120.262 ± 0.131
RF0.401 ± 0.0561.343 ± 0.2460.237 ± 0.0540.293 ± 0.090
SB-CFM1.280 ± 0.070—0.164 ± 0.0240.137 ± 0.033
OT-CFM0.331 ± 0.0390.673 ± 0.0490.138 ± 0.0330.131 ± 0.025
DSB0.424 ± 0.0841.811 ± 0.2470.304 ± 0.0450.339 ± 0.076
DSB-IMF0.410 ± 0.1041.538 ± 0.3080.327 ± 0.0440.349 ± 0.085
DSBM-IPF0.443 ± 0.1231.624 ± 0.0690.301 ± 0.0560.372 ± 0.034
[SF]²M-I,0.336 ± 0.0391.252 ± 0.3340.168 ± 0.0250.173 ± 0.057
[SF]²M-Exact0.267 ± 0.0550.719 ± 0.0720.120 ± 0.0210.149 ± 0.039
TGFlow0.261 ± 0.0520.452 ± 0.0210.117 ± 0.0200.121 ± 0.014

Normalized Path Energy Results (NPE, ↓)

Method / OT path variantN→8gaussiansmoons→ 8gaussiansN→moonsN→scurve
FM0.225 ± 0.033—0.955 ± 0.1010.835 ± 0.046
I-CFM0.188 ± 0.0212.508 ± 0.1821.021 ± 0.0890.918 ± 0.088
RF0.035 ± 0.0210.200 ± 0.1130.054 ± 0.0340.065 ± 0.047
SB-CFM1.016 ± 0.014—0.138 ± 0.0730.073 ± 0.038
OT-CFM0.021 ± 0.0120.046 ± 0.0300.043 ± 0.0260.028 ± 0.021
DSB0.025 ± 0.0180.100 ± 0.0770.091 ± 0.0450.078 ± 0.053
DSB-IMF0.040 ± 0.0320.171 ± 0.1150.070 ± 0.0900.044 ± 0.016
DSBM-IPF0.039 ± 0.0250.191 ± 0.1280.098 ± 0.0990.184 ± 0.103
[SF]²M-I,0.182 ± 0.0162.645 ± 0.2011.024 ± 0.0830.930 ± 0.068
[SF]²M-Exact0.019 ± 0.0090.026 ± 0.0220.052 ± 0.0290.037 ± 0.028
TGFlow0.012 ± 0.0080.009 ± 0.0070.037 ± 0.0240.022 ± 0.010

As shown in the table, TGFlow achieves better performance across all four task settings. Notably, the improvement is particularly pronounced in the moons → 8gaussians setting, suggesting that target-guided source construction is especially beneficial for target distributions with multimodal, clustered structures.

Q2: Does the proposed non-Gaussian path start from its stated source? In Section 3.3, Equation (21) gives T0(x)=σx+x0T_0(x) = \sigma x + x_0. For σ>0\sigma > 0, this appears to start from a noise-convolved version of the distribution of x0x_0, rather than that distribution itself. What value of σ\sigma is used, and what is the actual initial marginal in the non-Gaussian experiments?

We thank the reviewer for this thoughtful comment. We agree that clarifying the endpoint distributions of the conditional paths would help improve the clarity of our presentation. Equation (21) gives a general path formulation that includes a smoothing parameter, but we actually used σ=0\sigma = 0 in the original non-Gaussian experiments. This setting was not explicitly stated in the original manuscript. Therefore, the path used in the experiments is

xt=(1−t)x0+tx1,x_{t} = (1 - t)x_{0} + tx_{1},

which satisfies xt=0=x0x_{t = 0} = x_{0}. The corresponding initial source distribution is exactly the non-Gaussian source distribution after the target-guided adjustment, without any additional Gaussian convolution. At inference time, we sample directly from this source distribution to initialize the ODE, consistently with the starting point of the training path. We will explicitly clarify this setting in the revised manuscript.

Q3: How should the training budgets in Table 6 be interpreted? For CIFAR-10, 391K iterations at effective batch size 256 correspond to approximately 2,000 passes over 50K images, but the table reports 1,000 epochs. For ImageNet-128, 500K iterations at batch size 1,536 likewise appear inconsistent with 80 epochs. Which figures describe the runs used for the reported results, and were the FM and TGFlow models trained on the same number of images?

We appreciate the reviewer’s careful attention to this aspect. We recognize the value of reporting training budgets in a clear and consistent manner to support reproducibility and meaningful comparisons. We acknowledge that the iteration counts in the original Table 6 were inconsistent with the numbers of training epochs and the effective batch sizes. This inconsistency arose from a typographical error in the reported iteration counts, and we apologize for the confusion it caused. The updated training budgets are presented below. In addition, for each dataset, FM and TGFlow were trained using the same cumulative number of images to ensure matched training budgets and a fair comparison. We will update the training budget table in the revised manuscript accordingly.

ParameterCIFAR-10ImageNet- 32ImageNet- 64ImageNet- 128
Effective batch size256102420481536
Training epochs100020010080
Iterations195.3K250.2K62.6K66.7K

W1: Lack of comparisons with recent work

We thank the reviewer for recommending these relevant works. We agree that further clarifying the methodological differences and discussing recent approaches to source distribution design would help better highlight the contributions of TGFlow. A key advantage of TGFlow is its simple design: the target-guided source distribution can be constructed without training. We only need to precompute the per-coordinate variances of the target data to determine a fixed anisotropic Gaussian source, without additionally learning source distribution parameters or introducing auxiliary networks, while retaining the standard flow matching training objective. Specifically,compared with SourceFM, TGFlow requires no additional filtering or resampling of initial noise and constructs the source distribution directly from the per-coordinate variances of the target data. Compared with MM-FM, TGFlow does not require fitting a Gaussian mixture distribution. Compared with CAR-Flow and CSFM, TGFlow does not require conditioning information or an additional network for learning the source distribution. We will incorporate these discussions into the Related Work section of the revised manuscript to provide a more comprehensive comparison with existing methods.

In addition, following the experimental settings in Appendix E.3 of our paper, we have conducted additional comparisons with the above methods on CIFAR-10. The results are presented below. The experimental results demonstrate that TGFlow’s training-free source construction strategy, based on target statistics, outperforms the other baseline methods in the evaluated settings. Due to time constraints, we commit to including the corresponding comparison results on ImageNet in a future revision.

MethodFIDNFE
SourceFM3.88±0.16110
CAR-Flow7.99±0.09110
CSFM7.45±0.27110
MM-FM3.01±0.18110
TGFlow2.39 ± 0.12110

W2: The experiments do not isolate the benefit of coordinate-wise target information. TGFlow changes both the overall scale and the coordinate-wise shape of the source covariance. As a result, the comparison with N(0,I)\mathcal{N}(0,I) does not establish whether the improvement indeed comes from matching individual target variances. A source N(0,cI)\mathcal{N}(0,cI) with c=tr⁡(Λ)/dc = \operatorname{tr}(\Lambda)/d, with per-channel and coordinate-shuffled covariance controls, would test this central claim. Theorem 3 does not resolve the empirical question: it concerns an ideal terminal distribution for a Gaussian target, while the image results depend on a learned velocity field. Image generation measurements of transport length would also substantiate the proposed mechanism.

Regarding the question: “TGFlow changes both the overall scale and the coordinate-wise shape of the source covariance. As a result, the comparison with N(0,I)\mathcal{N}(0,I) does not establish whether the improvement indeed comes from matching individual target variances. A source N(0,cI)\mathcal{N}(0,cI) with c=tr⁡(Λ)/dc = \operatorname{tr}(\Lambda)/d, with per-channel and coordinate-shuffled covariance controls, would test this central claim.”

We appreciate the reviewer’s insightful observations and constructive suggestions. We agree that further examining the specific contribution of coordinate-wise target variance information would provide a deeper understanding of the benefits of TGFlow. To this end, we have conducted additional experiments and will incorporate these new experimental results into the revised manuscript.

To distinguish whether TGFlow’s benefits arise from adjusting the overall noise scale or from coordinate-wise target variance information, we have added two ablation experiments on CIFAR-10. The first uses an isotropic Gaussian source distribution N(0,cI)\mathcal{N}(0,cI), where c=tr⁡(Λ)/dc = \operatorname{tr}(\Lambda)/d, so that it has the same total variance as TGFlow while assigning the same variance to every coordinate. In the second, we randomly permute the diagonal entries of Λ\Lambda once to construct

Λπ=diag⁡(λπ(1),…,λπ(d)),x0∼N(0,Λπ),\Lambda_{\pi} = \operatorname{diag}(\lambda_{\pi(1)},\\ \ldots,\lambda_{\pi(d)}),x_{0} \sim \\ \mathcal{N}(0,\Lambda_{\pi}),

and keep this permutation fixed throughout training and inference. This control preserves the original set of variance values and the total variance while disrupting the correspondence between these variances and the target coordinates. These two controls examine the effects of coordinate-wise scale differences and their correspondence with the target coordinates, respectively. The experimental results are presented below. At the same NFE , TGFlow achieves the lowest FID, outperforming both the scale-matched isotropic and coordinate-shuffled controls. These results suggest that TGFlow benefits from incorporating coordinate-wise target variances and preserving their correspondence with the target coordinates.

MethodFIDNFE
N(0,Λπ)\mathcal{N}(0,\Lambda_{\pi})7.56 ± 0.29110
N(0,cI)\mathcal{N}(0,cI)4.01± 0.23110
N(0,Λ)\mathcal{N}(0,\Lambda)2.39 ± 0.12110
Reviewer_176qOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the target-guided flow matching, which reconstruct the source distribution using the marginal second-order statistics of the target data distribution. As a preparation, the paper first proved that the proposed method preserves the analytical forms of flow matching. Then the paper utilizes the marginal second-order statistics of each dimension to construct an initial source distribution for the proposed method. Moreover, the paper proved that a better source distribution reduces the KL divergence with a gaussian target distribution example. Finally, the paper experiments on multiple datasets to show the effectiveness of the proposed method.

Critical strengths
  1. The proposed method is relatively simple and straightforward, can be adopted in various scenarios.

  2. The experiments and the comparison of generation diversity and across different samplers show the effect of the proposed method.

Critical weaknesses
  1. The technique used by the proposed method is relatively too heuristic without sufficient in-depth understanding. The theoretical advantage of using the marginal second-order statistics should be further verified by presenting the convergence improvement upon other methods under the metrics that has been used by previous flow matching works like Wasserstein-2 distance, and consider more complex distributions beyond the simple monomial gaussian distribution as the target distribution. It would be more informative to state possible theoretical improvements like reducing the Lipschitz constant of the velocity field (which is a key factor in Wasserstein-2 analysis of flow matching, see [1] for reference). More detailed evaluation could improve the credibility of the proposed method.

  2. The clarity and the presentation of the paper could be improved. In my opinion, section 2.2 is of common knowledge in flow matching, and its effect for supporting the idea of the proposed method is relatively weak. I believe this sections could be (partially) moved to appendix, making more space for presenting a pseudo code of the proposed method for clearer understanding. In addition, there are some mixed use in notations like in line 273, equation (15) σ\sigma should be σmin⁡\sigma_{\min} for coherence.

[1] Benton, J., Deligiannidis, G., & Doucet, A. (2023). Error bounds for flow matching methods. arXiv preprint arXiv:2305.16860.

Questions for authors
  1. In equation (10) of line 212, is it important to use the unbiased sample variance (employing Bessel’s correction with N−1N − 1 rather than NN)? If the target distribution is multimodal, is the marginal second-order statistics enough to capture the information of the target distribution?

  2. In experiments, what is the confidence interval/error bars in Figure 5 and Table 2 (like which you stated in Table 1)?

  3. How does constructing a modified initial source distribution affect the geometric behavior (e.g. Lipschitz continuity) of the velocity field? Is it feasible to establish an error analysis in Wasserstein-2 distance similar to Theorem 1 of [1]?

[1] Benton, J., Deligiannidis, G., & Doucet, A. (2023). Error bounds for flow matching methods. arXiv preprint arXiv:2305.16860.

Comments on the paper

1 comment

Dear AC and reviewers,

Thank you for your constructive feedback and engagement during the rebuttal. We have strengthened the theoretical analysis and added baseline comparisons, controlled experiments, and non-Gaussian source evaluations. We summarize these additions below, starting with theory. The corresponding content has been highlighted in the revised manuscript to facilitate your review.

  • Connecting source design with generation error. We added two theoretical results allowing non-Gaussian targets and dependence across coordinates under explicit assumptions. By analyzing time-integrated Lipschitz constants of ideal velocity fields and incorporating velocity-field approximation error and terminal residual noise, we provide sufficient conditions for TGFlow to admit a smaller W₂ generation-error upper bound than both standard and total-variance-matched isotropic sources. We also clarified the conditions and the relationship between our original foundational results and existing theory.

  • Clarifying the extension to non-Gaussian sources. For sampleable sources with finite, nonzero coordinate-wise variances, TGFlow standardizes each coordinate and rescales it using the target's normalized marginal variances. This procedure does not require Gaussianity and preserves the standard conditional Flow Matching objective. The non-Gaussian experiments use zero-noise linear interpolation, ensuring that the initial marginal is exactly the adjusted source.

  • Lowest FID among the compared methods after adding four baselines. We added SourceFM, MM-FM, CAR-Flow, and CSFM on CIFAR-10. At the same NFE of 110, TGFlow achieves an FID of 2.39 ± 0.12, outperforming the best additional baseline result of 3.01 ± 0.18. This supports constructing sources directly from target marginal statistics.

  • Using target variance information in each dimension still improves performance when total variance is matched. The total-variance-matched isotropic source and the source with shuffled coordinate-wise variances achieve FIDs of 4.01 ± 0.23 and 7.56 ± 0.29, respectively, versus TGFlow's 2.39 ± 0.12. These controls support the contribution of coordinate-wise variances and their correspondence with target coordinates.

  • Controlled experiments and trajectory measurements. Holding the conditional path formulation and coupling method fixed while varying only source construction, TGFlow achieves the lowest W₂ distance and normalized path energy across all four 2D tasks. In image generation, the average sampling trajectory length decreases from 1.51 ± 0.05 to 0.65 ± 0.02, supporting shorter transport trajectories through source adjustment.

  • New experiments validate TGFlow's applicability to non-Gaussian source distributions. We added Student-t₅-to-moons and exponential-to-moons experiments, covering heavy-tailed and asymmetric sources. TGFlow achieves the lowest W₂ distance and normalized path energy among the compared methods on both tasks, extending the evidence beyond the original moons-source experiment.

  • Improved reliability and implementation details. We supplemented Figure 5 and Table 2 with means and sample standard deviations from five independent random seeds, corrected iteration counts, and clarified that FM and TGFlow use the same cumulative number of training images. Marginal variances require one dataset pass, with O(Nd) time and O(d) additional storage, without estimating a full covariance matrix. We also added training and inference pseudocode and improved organization and notation.

TGFlow incorporates variance information from each dimension of the target data into source design without learning additional source parameters, while retaining the standard Flow Matching objective. The added theory clarifies its connection to generation-error upper bounds and the conditions under which the results hold. The experiments further support the contribution of coordinate-wise target information and broaden the evidence for non-Gaussian sources. We hope these additions address the reviewers' main concerns and clarify TGFlow's methodological value and practical effectiveness.

Thank you again for your time and careful consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.