acceptodds
Under review as a conference paper at ICLR 2027

Dual Flow Matching: Learning Coupled Flows from Paired Views of a Shared Target

Abstract

Standard flow matching predicts velocity from a single noisy state, leaving the relationship between paths toward the same data target implicit. We introduce Dual Flow Matching (DualFM), which couples two independently corrupted views of a shared target within one backbone. The network observes both states and predicts both velocities in a single evaluation. Independent training clocks vary the information available in each view, while synchronized sampling evolves the two states jointly. Training remains simulation-free and requires no pretrained teacher. We prove that auxiliary conditioning weakly reduces irreducible regression risk and derive an exact common-task identity for comparing field-approximation errors. Our evaluation spans MNIST, CIFAR-10, ImageNet32, and ImageNet64, including class-conditional generation with SiT-XL/4 at the larger resolution. The UNet benchmarks show improved multi-step generation over I-CFM, with approximately 0.1% additional parameters and nearly unchanged large-batch sampling latency. Matched controls show gains beyond model capacity and observation information, while diagnostics show more accurate target estimates and learned use of instance-level correspondence. The gains also extend to minibatch optimal transport coupling. These results establish shared-target joint prediction as an effective way to improve flow matching while keeping the backbone size nearly unchanged.

Primary area:
generative models
Keywords:
flow matching, generative modeling, coupled dynamics, shared-target learning, image generation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_RU8NOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Dual Flow Matching (DualFM), which jointly predicts the velocities of two independently corrupted views of the same target using a single shared backbone. Training uses independent noise sources and independent clocks for the two views, while inference synchronizes the clocks and jointly evolves both states, returning one output branch.

The paper provides a statistical motivation based on reduced conditional regression risk when an auxiliary observation is available, a common/difference-coordinate analysis of the synchronized joint flow, and an exact common-task excess-risk identity for comparison with single-state flow matching. Empirically, the method is evaluated on MNIST, CIFAR-10, ImageNet32, and class-conditional ImageNet64, including UNet and SiT-XL/4 backbones. The authors also provide controls for parameter count, source variance, observation information, cross-view interaction, training clocks, target correspondence, and minibatch OT coupling, together with diagnostics of target reconstruction and branch agreement.

The results suggest that explicitly coupling two noisy views of the same instance can improve finite-network flow estimation and multi-step generation at very small additional backbone cost.

Critical strengths
  1. The empirical controls are unusually thorough. The authors do not merely compare DualFM with ordinary I-CFM. They separately control for parameter count, source variance, posterior observation information, number of processed states/targets, clock synchronization, and exact instance correspondence. The information-matched I-CFM control is particularly useful because it shows that the improvement cannot be explained solely by the additional posterior information contained in two noisy observations.

  2. The method provides consistent gains at very small backbone overhead. On CIFAR-10 and ImageNet32, DualFM improves mean FID over three training seeds across essentially all practically relevant sampling budgets. The gains also appear with SiT-XL/4 on ImageNet64 and when combined with minibatch OT coupling. The reported parameter and forward-compute increases are negligible, making the method practically attractive.

Critical weaknesses
  1. Independent-clock training does not theoretically determine the synchronized inference field. The method trains on a two-dimensional clock distribution but evaluates only the one-dimensional diagonal t=st=s. The paper's transport argument describes the canonical conditional-mean field restricted to this diagonal, but that value is not uniquely specified by minimizing an almost-everywhere two-clock regression objective. The authors provide empirical diagonal diagnostics, which mitigate the concern, but the theoretical justification of the actual train-to-sample procedure is incomplete.

  2. The distinction from the closest existing paired-view method would benefit from stronger empirical support. The paper identifies DUPA as closely related: it also independently corrupts the same image at multiple noise levels and supervises multiple velocity predictions. DualFM differs importantly by allowing full cross-view conditioning and by jointly evolving the two states at inference rather than using a feature-alignment loss. However, there is no direct empirical comparison. Since this is the closest conceptual prior work, such a comparison would materially clarify the practical novelty of DualFM.

A secondary limitation is scale. The larger SiT-XL/4 experiment is encouraging but uses ImageNet64 and a single training seed. Thus the evidence for scaling to modern high-resolution generative settings remains preliminary.

Questions for authors
  1. Training/inference mismatch: Since independent continuous clocks assign probability zero to the diagonal t=st=s, the training risk does not uniquely determine the field used during synchronized inference. Can the authors provide a stronger argument for why independent-clock training should recover the desired diagonal conditional field, beyond architectural continuity and the empirical diagnostics? For example, is there a regularity argument that controls diagonal error from nearby off-diagonal training error?

  2. Closest baseline: How does DualFM compare directly with DUPA under a matched backbone and training budget? I understand the conceptual distinction between feature alignment and full joint conditioning, but a controlled comparison would materially help assess whether joint prediction and joint inference provide a meaningful advantage over the closest existing paired-view approach.

  3. Scaling: The ImageNet64 SiT-XL/4 result uses one training seed. Do the authors have additional evidence that the improvement is stable at this scale, or results at higher-resolution ImageNet/text-to-image settings? This is not necessary for establishing the small-scale result, but it would affect how broadly I interpret the contribution.

3 comments
AuthorsOpenReview ↗

Response to Reviewer RU8N

We thank the reviewer for the constructive feedback. We address the training–inference question, the DUPA comparison, and scaling below.

1. Independent-clock training and synchronized inference (W1 / Q1).

At the ideal population optimum, independent-clock training identifies the synchronized field within the class of continuous predictors. Let Q=(Xt,Hs,t,s)Q=(X_t,H_s,t,s), UU denote the paired target velocities, and f∗(Q)=E[U∣Q]f^{\ast}(Q)=\mathbb{E}[U\mid Q] be the canonical conditional-mean field. Gaussian corruption and independent uniform clocks give the training distribution full support for 0<t,s<10 < t,s < 1. The Gaussian posterior mean, and hence f∗f^{\ast}, is continuous on this domain for our bounded image data.

Proof. Squared regression gives

=\mathbb{E}\left[\lVert f(Q)-f^{\ast}(Q)\rVert^{2}\right].$$ A continuous predictor attaining the Bayes-optimal population risk therefore agrees with $f^{\ast}$ almost everywhere. If it differed at a diagonal point, continuity would make the difference nonzero in an open neighborhood. Full support gives that neighborhood positive training probability, contradicting optimality. Thus $$f(x,h,\tau,\tau)=f^{\ast}(x,h,\tau,\tau), \qquad 0 < \tau < 1.$$ The diagonal restriction is therefore exactly the conditional-mean field used in our synchronized transport result. Empirically, our synchronized diagnostics show lower reconstruction error at all five evaluated times on both CIFAR-10 and ImageNet32. Independent-clock training also outperforms synchronized-clock training at all four tested CIFAR-10 sampling budgets (revised Table 5). **2. DUPA comparison and training compute (W2 / Q2).** We evaluated a UNet adaptation of DUPA on ImageNet32 using the same backbone architecture as DualFM, Euler sampling, and FID-50k from EMA checkpoints. With the same batch of clean targets, DUPA requires approximately twice the training FLOPs per update of DualFM. Thus, DUPA at 300k updates approximately matches the training compute of DualFM at 600k; DUPA at 600k uses approximately twice that budget. Compute here denotes training forward/backward FLOPs. The I-CFM and DualFM entries below use the paper's seed-42 runs. Training compute is normalized to DualFM at 600k updates. | Method / updates | Training compute | NFE 2 | 5 | 10 | 20 | 50 | 100 | |---|---:|---:|---:|---:|---:|---:|---:| | I-CFM / 600k | ~1x | 199.123 | 46.216 | 20.045 | 13.495 | 10.754 | 9.755 | | DualFM / 600k | 1x | 174.343 | 44.590 | 19.472 | 12.602 | 10.000 | 9.238 | | DUPA / 300k | ~1x | 195.128 | 45.217 | 19.691 | 12.971 | 10.304 | 9.487 | | DUPA / 600k | ~2x | 196.237 | 44.583 | 19.439 | 12.682 | 9.705 | 8.983 | At approximately matched training FLOPs, DualFM has lower FID than DUPA at all six evaluated sampling budgets. With approximately twice the training compute, DUPA at 600k achieves lower FID at NFE 5, 10, 50, and 100, while DualFM remains better at NFE 2 and 20. These are single-run comparisons. **3. Stability and higher-resolution generation (W3 / Q3).** We added a second ImageNet64 training run for both methods, with the same class-conditional SiT-XL/4 backbone and 1.2M updates. DualFM achieves lower FID in both runs at every sampling budget. Below are means ± sample standard deviations over the two runs (revised Table 2(b)). | ImageNet64, SiT-XL/4 | 10 NFE | 20 NFE | 50 NFE | 100 NFE | 1000 NFE | |---|---:|---:|---:|---:|---:| | I-CFM | 12.058 ± 0.103 | 8.312 ± 0.050 | 6.615 ± 0.033 | 6.042 ± 0.009 | 5.621 ± 0.089 | | DualFM | **11.550 ± 0.018** | **7.541 ± 0.165** | **6.279 ± 0.006** | **5.851 ± 0.018** | **5.414 ± 0.098** | We also evaluated class-conditional latent-space generation on ImageNet256 with SiT-XL/2 at 400k updates. DualFM obtains 16.7 FID versus 18.8 for the published SiT baseline. Both use Euler-250, 50k samples, and the ADM evaluation protocol (revised Table 2(c)). | ImageNet256, latent SiT-XL/2 | FID ↓ | |---|---:| | SiT baseline, Gao et al. (2025), Table 5 | 18.8 | | DualFM | **16.7** | These results extend the evidence to two training runs at the larger pixel-space scale and to high-resolution latent generation.

Thank you for the additional experiments. The continuity argument addresses my theoretical concern at the population optimum, and the matched DUPA comparison strengthens the evaluation, although it remains limited to a single run. The new ImageNet64, ImageNet256, and Heun results also provide useful evidence of generalization, despite the unmatched high-resolution baseline. The authors have addressed most of my concerns, and I am satisfied with the response. I will maintain my score.

Thank you for taking the time to review our responses and the additional experiments. We appreciate your constructive feedback, which has helped us improve the manuscript.

Reviewer_WFUkOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose DualFM, a simple and theoretically supported extension to the flow matching training framework that jointly estimates the velocities for two independently constructed views of the same target example. DualFM places no restriction on the sampled time within each view, and only marginally increases the compute cost compared to flow matching (on UNet). The authors show DualFM improves upon flow matching across image generation benchmarks of varying complexity (MNIST, CIFAR-10, ImageNet32 & ImageNet64), across UNet and SiT-XL/4 architectures, and across denoising steps.

Critical strengths
  • DualFM improves over standard flow matching at FID50K in every single setting shown, in ImageNet (though low resolution) and across two architectures. It further is robust to increasing denoising steps, including up to 1000 shown.
  • The method is well ablated in Table 4, with all core ablation settings tested, and all are worse than the proposed method.
  • DualFM marginally increases compute over standard flow matching in the measured UNet model across both CIFAR-10 and ImageNet32
  • DualFM introduces both new ways of training and generation through its use of multiple corrupted views. Training on additional views can be studied given the negligible increased compute, (from my understanding) Figure 4 suggests that training with different auxiliary sample bounds (e.g., closer to the clean target) compared to the root sample may improve performance, and whether denoising with non-matched \tau may improve generation (e.g., if 100 NFE is desired, denoise the auxiliary for K steps and then denoise the root for 100-K steps).
Critical weaknesses
  • What is the compute overhead under the SiT-XL/4 architecture? Only UNet is mentioned. It would aid in understanding whether the compute similarity is restricted to UNet alone, or generalizes.
  • Results with classifier free guidance are missing for the class-conditional experiments. Does DualFM still improve over flow matching in that regime? This result would further support / hinder the significance of the proposed contribution.
  • Popular established benchmarks such as ImageNet256, and text conditioned settings (e..g, MSCOCO256) are missing. Results on these tasks (e.g., at 400K iterations) would greatly aid in understanding the robustness of DualFM to higher-fedility and more complex settings. Moreover, ImageNet256 is a standard benchmark for image generation works published at similar venues (ICCV, ECCV, CVPR, NeurIPS).
  • While all other settings have three training seeds at 600K steps, the SiT-XL/4 setting is singular in reporting only one seed at 1.2M steps. How does SiT-XL/4 on ImageNet64 behave across different seeds? Similar to the above questions, this helps understand DualFM's robustness characteristics to larger models on more complex tasks.
Questions for authors

None.

3 comments
AuthorsOpenReview ↗

Response to Reviewer WFUk

We thank the reviewer for the constructive feedback. We added the requested experiments and marked the revisions in blue.

1. Compute overhead with SiT-XL/4 (W1).

We profiled both SiT-XL/4 models on ImageNet64. DualFM uses one shared backbone forward pass to predict both velocities.

SiT-XL/4, ImageNet64I-CFMDualFMIncrease
Trainable parameters (M)674.890676.6250.2571%
Profiled forward GFLOPs / sample237.271237.3310.0252%
Batch-256 sampling time (s)70.507570.65600.2106%

The small overhead also holds for the transformer backbone. We added these measurements to revised Table 4 and Appendix B.12.

2. High-resolution latent generation and CFG (W3 / W2).

We added class-conditional ImageNet256 experiments with latent-space SiT-XL/2 at 400k updates. Without CFG, DualFM improves on the published SiT result under the same evaluation protocol: Euler-250, 50k samples, and ADM evaluation (revised Table 2(c)). This extends the evaluation to high-resolution latent generation. Text-conditioned generation remains future work.

ImageNet256, latent SiT-XL/2, without CFGFID ↓
SiT baseline, Gao et al. (2025), Table 518.8
DualFM16.7

For CFG, we additionally evaluated DualFM on ImageNet64 with SiT-XL/4 and Euler sampling at 5 NFE, tuning the guidance scale. DualFM achieves 19.83 FID, compared with 21.05 reported for ordinary FM with tuned CFG in Berthelot et al. (2026), The Coupling Within: Flow Matching via Distilled Normalizing Flows, Table 5.

ImageNet64, SiT-XL/4, Euler-5, tuned CFGFID ↓
I-CFM, Berthelot et al. (2026), Table 521.05
DualFM19.83

The published FM baseline uses a t2t^2 sampling grid and its own training recipe, so we present this as a comparison with a published reference.

We also evaluated DualFM on ImageNet256 with CFG=2, obtaining 7.8 FID. We could not find a published SiT baseline matching the 400k training budget, Euler-250 sampler, and CFG=2 setting. Training a matched baseline was not feasible within the rebuttal's time and compute budget. A matched comparison is needed to establish relative gains under CFG in this ImageNet256 setting.

3. ImageNet64 results across training seeds (W4).

We added a second training seed for both methods, using the same class-conditional SiT-XL/4 backbone and 1.2M updates. DualFM achieves lower FID in both runs at every evaluated budget. Below are means ± sample standard deviations over the two training seeds (revised Table 2(b)).

ImageNet64, SiT-XL/410 NFE20 NFE50 NFE100 NFE1000 NFE
I-CFM12.058 ± 0.1038.312 ± 0.0506.615 ± 0.0336.042 ± 0.0095.621 ± 0.089
DualFM11.550 ± 0.0187.541 ± 0.1656.279 ± 0.0065.851 ± 0.0185.414 ± 0.098
Reviewer_WFUkOpenReview ↗

Thank you for your rebuttal

I thank the authors for their rebuttal. The empirical results with ImageNet256 are strong, as are their results on the other experiments I've asked for. They have adequately answered my questions, and I adjust my score accordingly.

Thank you for your positive reassessment and support. We are glad that the additional experiments addressed your questions. Your constructive suggestions helped us strengthen the empirical evaluation, and we sincerely appreciate your time and careful consideration!

Reviewer_2UpFOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Dual Flow Matching jointly predicts velocities for two independently corrupted views of the same target using one shared backbone. Training samples independent noise levels, while inference synchronously evolves both states from independent Gaussian noise and returns one generated image. The paper provides conditional-regression and common–difference analyses, together with generation experiments and matched controls. Results show improved FID over I-CFM with negligible measured UNet inference overhead.

Critical strengths
  1. A simple modification with low measured overhead. Both velocities are predicted in one shared-backbone evaluation. The reported parameter, FLOP, and sampling-latency measurements support the efficiency claim for the tested UNets.

  2. Targeted controls strengthen the empirical evidence. Parameter-, source-, and information-matched comparisons address plausible alternative explanations for the gains. Independent-clock and target-pairing ablations further support the role of shared-target joint learning.

Critical weaknesses
  1. Experimental coverage remains limited. Evaluation is restricted to low-resolution image generation, with the main mechanism controls concentrated on CIFAR-10. ImageNet64 uses one training seed, and there is no high-resolution or latent-space validation. Broader experiments are needed to establish whether the gains persist beyond these settings and uniform Euler sampling.

  2. The mechanism behind the additional dual-state benefit remains incompletely established. Auxiliary conditioning reduces irreducible regression risk, but does not directly guarantee improved generation. Moreover, the synchronous common–difference decomposition shows that target information can be compressed into a single common state. Although DualFM improves over the information-matched control, the remaining gain is modest, and the analysis does not establish why two freely predicted velocities provide a better learning parameterisation.

Questions for authors
  1. The auxiliary-conditioning argument extends naturally to more observations. Have you evaluated three- or four-flow variants under matched training budgets, including corresponding source- or information-matched single-flow controls? These results could clarify why two flows are preferred.

  2. Could you compare against a model that learns only the common field or shared target estimate and updates the difference coordinate analytically? This would help isolate the benefit of learning two unconstrained velocity outputs.

  3. Do the gains persist with a stronger numerical solver or in a higher-resolution, latent-space setting? Such evidence would materially strengthen the generality of the method.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 2UpF

We thank the reviewer for the constructive feedback. We added the requested experiments and marked the revisions in blue.

1. Broader evaluation and robustness (W1 / Q3).

ImageNet256. We evaluated DualFM with latent-space SiT-XL/2 at 400k updates. It achieves 16.7 FID versus 18.8 for the published SiT baseline. Both use Euler-250, 50k samples, and the ADM evaluation protocol (revised Table 2(c)).

ImageNet256, latent SiT-XL/2FID ↓
SiT baseline, Gao et al. (2025), Table 518.8
DualFM16.7

Heun sampling. DualFM also improves FID with the second-order Heun solver on ImageNet32 (revised Table 3). NFE counts actual field evaluations, including predictor and corrector calls.

ImageNet32, Heun50 NFE100 NFE
I-CFM10.3318.073
DualFM9.5437.862

Two ImageNet64 training runs. We added a second run for both methods with the same SiT-XL/4 backbone and 1.2M updates. DualFM improves FID in both runs at every budget. Below are means ± sample standard deviations (revised Table 2(b)).

ImageNet6410 NFE20 NFE50 NFE100 NFE1000 NFE
I-CFM12.058 ± 0.1038.312 ± 0.0506.615 ± 0.0336.042 ± 0.0095.621 ± 0.089
DualFM11.550 ± 0.0187.541 ± 0.1656.279 ± 0.0065.851 ± 0.0185.414 ± 0.098

2. Learning only the common field (W2 / Q2).

We trained this control on ImageNet32. It receives both views and clocks but predicts only the common velocity. During training, the branch velocities are reconstructed analytically; during sampling, the difference coordinate follows its exact dynamics. It uses the same UNet core, 600k updates, and target batch 128.

ImageNet32, Euler10 NFE50 NFE100 NFE
I-CFM20.04510.7549.755
Common field + analytic difference19.78110.4679.487
Information-matched I-CFM19.98410.3769.579
DualFM19.47210.0009.238

DualFM outperforms this control at all three reported NFE budgets. This supports the benefit of learning two freely predicted velocities, even when the difference dynamics are known (revised Table 6; Appendix B.11).

The common state preserves target information, but this does not make the finite-network learning problem identical. Equation (13) shows that DualFM supervises target estimation and agreement between the branch estimates. The matched controls provide empirical evidence for the remaining generation gain; the regression-risk argument supplies its statistical motivation.

3. Three and four flows (Q1).

We trained three-/four-flow variants and a separate information-matched single-state control for each K on ImageNet32. All use the same UNet core, 600k updates, and target batch 128 (revised Table 6).

ImageNet32, EulerK2 NFE5 NFE10 NFE20 NFE50 NFE100 NFE
Information-matched I-CFM2178.46245.95719.98412.80810.3769.579
DualFM2174.34344.59019.47212.60210.0009.238
Information-matched I-CFM3167.92343.28419.53912.76810.3479.482
DualFM3163.12141.59718.90112.2289.7669.020
Information-matched I-CFM4168.83343.89219.44012.32510.2429.556
DualFM4165.89542.65319.06111.9329.5278.890

Each joint model beats its corresponding information-matched control at all six reported NFE budgets. Three flows perform best at 2–10 NFE, and four at 20–100 NFE. Thus, adding a fourth flow does not improve every budget. More views may provide more information while increasing prediction difficulty within the fixed backbone. We chose two flows as the minimal configuration for cross-view prediction; additional flows can further improve quality. These controls use one training seed.

Reviewer_2UpFOpenReview ↗

Official Comment by 2UpF

I appreciate the response from the authors, which solved most my concerns. However I wonder the experiments are still limited with low resolution, where most diffusion generations can handle more than 1k resolution generations. Hence I keep my rating

Thank you for your follow-up and constructive discussion. We appreciate the time and effort you have devoted to evaluating our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.