acceptodds
Withdrawn from ICLR 2027

Smoother Flow Matching via Contrastive Trajectory Repulsion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

14 views

Abstract

Trajectory crossing remains a critical bottleneck in Flow Matching (FM), and previous works typically view these crossings from a theoretical optimization perspective causing velocity averaging. They attempt to address it indirectly by post-hoc distillation or endpoint coupling, without explicitly regulating the intermediate trajectories. In this paper, we introduce a new network learning perspective: crossing points inherently induce large local Lipschitz constants in the target velocity field, leading to two drawbacks. First, high Lipschitz constants correspond to high-frequency signals in the velocity field that neural networks struggle to fit due to spectral bias. Second, they also imply drastic velocity variations, leading to severe numerical integration errors in few-step inference. To alleviate this, we propose CoFlow, a framework that introduces the contrastive learning paradigm into FM to explicitly repel trajectories during training, thereby lowering the local Lipschitz constants of the velocity field. Specifically, we formulate CoFlow from a Stochastic Differential Equation (SDE) perspective by injecting a repulsive drift term. This drift actively guides the forward process of positive samples away from negative trajectorie, effectively reducing the local Lipschitz constant. Furthermore, we derive an equivalent stochastic interpolant formulation from this SDE, providing a simple and tractable design space to control the influence of negative samples. Extensive experiments on ImageNet 256x256 demonstrate that CoFlow significantly reduces FID compared to standard FM in few-step inference (e.g., 20 steps), with no added training overhead. The code can be accessed from: https://anonymous.4open.science/r/CoFlow-A0A0

Primary area:
generative models
Keywords:
Flow Matching, Generative Models, Contrastive 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_6e2ZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CoFlow modifies intermediate training paths by repelling them from negative trajectories while preserving their endpoints. The paper argues that this encourages smoother velocity fields and improves few-step generation. It also derives an affine relationship between the population-optimal DeltaFM and standard FM fields. ImageNet experiments show improvements across two SiT model sizes and combinations with REPA, OT-CFM, and velocity centering.

Critical strengths
  • The training modification is simple and adds little computation. The ImageNet improvements appear across model sizes and with or without REPA, making the empirical result interesting.
  • The DeltaFM analysis provides a useful additional result. Applying the derived transformation at inference achieves performance close to explicit DeltaFM training and improves CoFlow further.
Critical weaknesses
  • The motivation conflates conflicting conditional targets with a nonsmooth marginal velocity field. The former does not establish the latter.
  • Repelling sampled trajectories is not shown to reduce the Lipschitz constant of the learned marginal field. Endpoint preservation establishes valid transport endpoints, not improved smoothness.
  • The Euler error bound depends on spatial smoothness, temporal smoothness, and velocity magnitude. The paper does not establish that its trajectory modification improves their combined contribution.
  • Direct evidence for the proposed mechanism is limited to the toy experiment. ImageNet FID improvements alone cannot distinguish reduced integration error from changes in learning difficulty or approximation error.
  • The main image experiments use Euler–Maruyama sampling. A deterministic ODE comparison is missing despite ODE truncation error being central to the explanation.
Questions for authors
  • Can you provide direct measurements of learned-field smoothness on ImageNet, such as Jacobian norms?
  • Under deterministic ODE sampling at matched NFE, does CoFlow improve quality and reduce error relative to a converged solution?
Reviewer_XmtrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CoFlow, a path modification approach for flow matching (FM). For a positive and a reference pair sampled from the minibatch, it replaces the linear interpolant of standard FM with this new stochastic interpolant, x^t=xt+λt(1−t)(xt−xtr)\hat{x}_t=x_t+\lambda t(1-t)(x_t-x_t^r), and provide an SDE interpretation with an added repulsive drift. The network is trained with the derivative of this modified path. The motivation is to separate crossing trajectories could reduce the spatial Lipschitz constant of the learnt velocity field, also improve numerical integration at fewer step sampling. The paper shows lipschitz/integration error analysis with 2D toy experiment and performance gains on class-conditional ImageNet-256 generation compare to standard FM.

Critical strengths

The authors propose a practical approach, which is a simple interpolant modification without additional learned components. CoFlow reuses existing minibatch pairs as negative references and requires no additional backbone evaluation or auxiliary network. In Table 4, they justify their design choice by showing that using the relative position of a randomly sampled reference is effective. Furthermore, this is supported by performance gains over standard FM in fewer step cases. For instance, Table 1 shows meaningful improvements across two cases (w/ and w/o REPA). Figure 5 further suggests that the benefit is concentrated in the lower NFE regime.

Critical weaknesses

One of the main concern is the connection between the motivation and the method remains unclear. In the current manuscript, it is not sufficiently demonstrated that the proposed method addresses the motivation, i.e., decrease crossing and increase smoothness of learnt field. Fig 3 provides evidence on a 2D toy problem, but moving a sample away from the original position of one reference does not guarantee to establish fewer conflicts among trajectories. Authors should show more analyses to bridge the gap here. Fig 3(b) indeed reports lower Lipschitz constant of learnt fields, but this does not directly quantify trajectory crossing (as claimed in the first bulletpoint at Intro). Comparing same time crossings/near-conflicts between conditional training paths before and after modification would more directly test whether CoFlow achieves its intended effect. Also, authors should show evidence beyond the toy setting, as the prevalence of near-conflicts and their reduction by CoFlow are not measured on ImageNet. Trajectory straightness or curvature, for example, the normalized flow straightness analyses in CAF [A] and VRFNO [B], time-dependent Jacobian norms or local Lipschitz estimates would help connect the toy evidence to the main generation results.

Regarding the second bulletpoint, the connection to Doob’s h-transform is overstated. The attraction–repulsion analogy is understandable, but it does not establish a direct connection between Doob h-transform formulation and CoFlow with a clear distinction: the attractive drift arises from conditioning on a terminal endpoint, whereas CoFlow designs repulsion from an intermediate reference. I’m not sure this whole standalone section in the method part about Doob’s is necessary in the current manuscript, and believe it needs to be toned down and presented as motivation.

Finally, the last bulletpoint also has a huge issue, where it claims theoretical insight into DeltaFM, but it is already present in the original paper with deriving the affine relation between the optimal DeltaFM and FM fields. The added contribution here appears to be the practical extension to CoFlow with VC. The authors should distinguish this practical extension from the existing theoretical result and revise the contribution statement accordingly.

Taken together, all three stated contributions need stronger support or more precise framing to clarify.

Another problem is concerning the equivalnce between the repulsive SDE and proposed stochastic interpolant in Appendix C, which is the key theoretical framing of the whole paper. The Gaussian reparameterization in Eq41-42 is valid for the Brownian integral at a fixed time. However, the repulsive integral depends on the state history driven by the same Brownian noise. The manuscript does not establish how this dependence is preserved when passing to the simulation-free reparameterization in Eq8. Thus, matching the Gaussian term’s marginal distribution alone does not establish the claimed marginal equivalence of the complete constructions. This part need clarification.

Regarding the experiments, there are two concerns. First, the logic/analysis and the main sampling evaluation use different dynamics, e.g., eq4 and toy experiments in Fig3 says about deterministic ODE integration, whereas the main table is regarding SDE sampling. The authors should report matched-NFE ODE results for the main table using the same trained checkpoints and evaluation settings.

Second, the OT-CFM baseline needs clarification. The OT-CFM results in Table 3 look particularly weak, even compared with standard FM. Since standard minibatch OT can perform poorly in high-dim/many-class conditional generation, as discussed in C²OT [C], the authors should examine C²OT and its experimental setup and evaluate CoFlow in that setting.

And a comment regarding presentation: Fig2 should be improved to make it self-contained. Current version is not reflecting whole concept and full comparison of two methods. For example, in the CoFlow part, the blue trajectory is not explicitly connected to the displacement producing the modified path.


[A] Constant Acceleration Flow (NeurIPS 2024)

[B] Straighten Viscous Rectified Flow via Noise Optimization (ICCV 2025)

[C] The Curse of Conditions: Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation (ICCV 2025)

Questions for authors

Could the authors also report CoFlow results with CFG? Table 5 shows substantial gains from the CoFlow+VC, suggesting that learning modified paths can still benefit from additional guidance at sampling time. I am curious how CoFlow interacts with CFG and whether it provides complementary gain as well.

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

This paper proposes CoFlow, a training framework for Flow Matching (FM) that aims to mitigate trajectory crossings and smooth the learned velocity field for few-step generation. Drawing intuition from Doob's h-transform and contrastive learning, the authors modify the standard linear stochastic interpolant by adding a time-vanishing repulsive term driven by negative reference trajectories sampled from the training batch. In addition, they show that the optimal velocity field of prior contrastive flow matching (DeltaFM) under independent negatives reduces to an inference-time affine transformation, which they term Velocity Centering (VC) and combine with their method at test time. Experiments on 2D synthetic data and class-conditional ImageNet 256×256 with SiT backbones show that CoFlow improves generation quality in the few-step regime.

Critical strengths
  1. The motivation connecting trajectory crossings to high local Lipschitz constants, neural spectral bias, and ODE solver truncation error is intuitive and clearly presented, offering a clean perspective on why crossing paths hurt few-step flow matching.
  2. The resulting simulation-free interpolant modification (Algorithm 1) is very simple to implement and adds negligible training overhead, while achieving consistent few-step improvements over vanilla FM across SiT-S/2, SiT-B/2, and SiT-XL/2 backbones (both with and without REPA).
  3. The analysis connecting DeltaFM to an inference-time Velocity Centering (VC) operation is a neat practical observation that clarifies how unconditional velocity repulsion shifts the learned drift.
Critical weaknesses
  1. Conceptual gap between "crossing repulsion" and random negative pairing, as well as theoretical inconsistencies:

    • Random negative pairing vs. trajectory crossing repulsion: The core motivation (Figure 1 and Section 1) argues that when two trajectories cross, spatially adjacent points xA≈xBx_A \approx x_B have conflicting velocities, and repelling them lowers the local Lipschitz constant. However, in the default formulation (Eq. 12 and Algorithm 1), the reference trajectory (x0r,x1r)(x_0^r, x_1^r) is sampled uniformly at random from the mini-batch. In a high-dimensional latent space (d=4096d = 4096), a randomly selected trajectory xtrx_t^r is almost never near xtx_t. Thus, x^t=(1+λt(1−t))xt−λt(1−t)xtr\hat{x}_t = (1 + \lambda t(1-t))x_t - \lambda t(1-t)x_t^r is simply a linear combination of two independent interpolants rather than a local repulsion of colliding paths. This is echoed in Table 4, where selecting negatives via "Nearest" (79.40 FID) or "Lipschitz" (78.12 FID) yields only marginal differences compared to "Random" (79.55 FID). While the authors note in Section 4.3 that Nearest and Lipschitz are computed at the endpoint x0x_0 to avoid time discontinuities, this further means that intermediate crossings at t∈(0,1)t \in (0, 1) are never actually detected or selectively repelled.
    • Disconnect between ODE proofs and Euler-Maruyama (SDE) sampling: The unbiasedness proof in Appendix D (Eq. 49–56) relies on the deterministic continuity equation (probability flow ODE), and the Lipschitz truncation error bound in Eq. (4) and Appendix F (Eq. 66) is also derived for deterministic Euler ODE integration. However, all ImageNet experiments in Section 4 use a 20-step Euler-Maruyama (EM) SDE sampler. In standard SiT, EM sampling converts the learned velocity uθ(xt,t)u_\theta(x_t, t) into the marginal score ∇xtlog⁡pt(xt)\nabla_{x_t} \log p_t(x_t) using the Gaussian conditional property of xt=(1−t)x0+tx1x_t = (1-t)x_0 + t x_1. Under CoFlow (Eq. 12), x^t\hat{x}_t mixes two data samples (x0,x0r)(x_0, x_0^r) and two noise variables (x1,x1r)(x_1, x_1^r), so p(x^t∣x0)p(\hat{x}_t \mid x_0) is no longer a simple Gaussian with variance t2It^2 I and the standard score-velocity conversion no longer holds exactly for the modified marginal ρt(x^t)\rho_t(\hat{x}_t). It would be better to clarify how the score/diffusion term is computed during EM sampling and whether the unbiasedness guarantee still holds under the SDE sampler.
    • Imprecision in the SDE-to-SI derivation (Section 3.3 and Appendix C): First, in Appendix B (Eq. 26–28), the Itô integral ∫0tg(ξ)s(ξ)dwξ\int_0^t \frac{g(\xi)}{s(\xi)} dw_\xi only matches σ(t)x1\sigma(t) x_1 in marginal law at a fixed time tt; holding x1x_1 fixed across tt gives a deterministic line dxt=(x1−x0)dtdx_t = (x_1 - x_0)dt rather than a sample path of the SDE in Eq. (1). Second, in Eq. (6) and Appendix C (Eq. 33–37), the integrating factor s(t)s(t) is applied by canceling α(t)xt\alpha(t)x_t against dx^td\hat{x}_t, which conflates the unperturbed state xtx_t and the perturbed state x^t\hat{x}_t in the drift term. Since CoFlow ultimately defines the multi-sample stochastic interpolant x^t=I(t,z)\hat{x}_t = I(t, z) directly in Eq. (8) and Appendix D, presenting it directly within the stochastic interpolant framework would be much more rigorous.
    • Novelty and positioning relative to DeltaFM: As acknowledged in Appendix E (line 856), the closed-form optimal velocity derivation for DeltaFM adapts the derivation already given in the appendix of DeltaFM (Stoica et al., 2025). Moreover, in Table 2, standalone DeltaFM (32.32 FID, 15.73 sFID, 56.33 IS) substantially outperforms standalone CoFlow (37.76 FID, 20.23 sFID, 52.92 IS), which weakens the claim that trajectory-level repulsion is more effective than velocity-level contrastive alignment.
  2. Missing experimental controls, unverified high-dimensional claims, and result anomalies:

    • Missing FM + VC baseline in Table 2: In Table 2 (SiT-B/2 + REPA at NFE=20), the authors show that adding Velocity Centering (CoFlow + VC) improves CoFlow from 37.76 to 27.74 FID, beating DeltaFM (32.32 FID). However, FM + VC at NFE=20 is omitted from Table 2 (it is only evaluated at NFE=50 in Table 5, where FM + VC already matches/beats DeltaFM, and where CoFlow is omitted). Including FM + VC at NFE=20 in Table 2 (and CoFlow / CoFlow + VC in Table 5) is essential for a fair comparison.
    • No empirical verification of Lipschitz reduction on ImageNet: While Figure 3(b) measures the mean local Lipschitz constant on the 2D Two-Moons toy task, there is no quantitative check showing that CoFlow actually reduces the empirical local Lipschitz constant (or trajectory curvature) of the learned velocity network on ImageNet 256×256.
    • Unspecified class conditioning during negative sampling: All ImageNet experiments are class-conditional, yet Algorithm 1 and Section 3.4 omit class labels yy. When (x0r,x1r)(x_0^r, x_1^r) is sampled via randperm across the mini-batch, x0rx_0^r generally belongs to a different class yr≠yy^r \neq y, while the network uθ(x^t,t,y)u_\theta(\hat{x}_t, t, y) is conditioned only on the positive sample's class yy. It is unclear whether negative pairing is restricted to the same class or performed across classes, and how cross-class trajectory mixing affects class-conditional marginals and Classifier-Free Guidance (CFG).
    • Anomaly in OT-CFM baseline (Table 3) and lack of CFG / ODE solver evaluation: In Table 3 (SiT-B/2 without REPA, NFE=20), OT-CFM achieves 69.90 FID and OT-CFM + CoFlow achieves 63.92 FID, which are both significantly worse than standard FM (61.43 FID) and CoFlow (52.79 FID) under the exact same setting in Table 1. In addition, all main experiments use λ=0.2\lambda = 0.2 even though λ=0.5\lambda = 0.5 performs much better in Table 4 (75.06 vs. 79.55 FID), and no results are reported with standard ODE solvers (e.g., deterministic Euler / Heun) or with standard CFG (w>1.0w > 1.0).
  3. Writing, formatting, and presentation issues:

    • The manuscript contains several typos and grammatical errors that should be cleaned up, including "during infernce" (page 2, line 102), "its machanism" (page 3, line 125), "NEF= 20" / "NEF=50" instead of "NFE" (page 7, line 372; page 16, Table 5 caption), "C omparison" (Table 5 caption), "Lipschitze constants", "adjusts intermediate to encourage" (missing "trajectories"), and "can further improves" (page 9, Section 5, lines 477–482), "continuty equation" (page 15, line 785), and "Sepcifically" alongside a duplicated paragraph at the end of Appendix E (page 17, lines 909–915).
    • Multi-letter subscripts in math mode such as LCoFlowL_{CoFlow} (Eq. 10 and Appendix D) and L(ΔFM)L^{(\Delta FM)} (Appendix E) should be typeset with \text{} or \mathrm{} to avoid awkward spacing.
Questions for authors
  1. Could the authors clarify how the Euler-Maruyama (EM) SDE sampler is implemented for CoFlow in Section 4? Specifically, since x^t\hat{x}_t in Eq. (12) is a combination of two data points and two Gaussian noises, the standard SiT identity relating the velocity uθ(x^t,t)u_\theta(\hat{x}_t, t) to the marginal score ∇x^tlog⁡ρt(x^t)\nabla_{\hat{x}_t} \log \rho_t(\hat{x}_t) no longer holds exactly. How does CoFlow perform when evaluated with a deterministic ODE Euler solver (which directly matches the continuity equation in Appendix D and the Lipschitz error bound in Appendix F)?
  2. In Table 4, why does repelling from a randomly chosen trajectory in the batch ("Random", 79.55 FID) perform almost identically to repelling from the "Nearest" (79.40 FID) or highest-"Lipschitz" (78.12 FID) trajectory? Could the authors measure the empirical local Lipschitz constant (e.g., via finite-difference directional derivatives along sampling trajectories) on the trained ImageNet SiT models to verify whether CoFlow actually smooths the learned velocity field in high dimensions?
  3. To make the comparison with DeltaFM in Table 2 and Table 5 complete, could the authors report FM + VC at NFE=20 in Table 2, as well as CoFlow and CoFlow + VC at NFE=50 on SiT-B/2 and SiT-XL/2 in Table 5? Also, why does standalone DeltaFM outperform standalone CoFlow by more than 5 FID points in Table 2?
  4. In class-conditional ImageNet training, are the negative reference trajectories (x0r,x1r)(x_0^r, x_1^r) sampled within the same class yy or across arbitrary classes in the mini-batch? In addition, how does CoFlow perform when combined with standard Classifier-Free Guidance (CFG), and why does OT-CFM in Table 3 (69.90 FID) degrade performance so heavily compared to vanilla FM in Table 1 (61.43 FID)?

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

Related papers

Discussion (0)

Sign in to comment.