acceptodds
Under review as a conference paper at ICLR 2027

Joint Flow Matching: Consistent Generative and Discriminative Inference from a Shared Flow

Abstract

We study Joint Flow Matching (JFM), a flow-based framework for learning a joint distribution over paired variables such as images and class labels. JFM transports the variables in opposite directions: an image is transported from noise to data while its label is transported from data to noise. This provides a single probabilistic model in which generation and classification are represented as complementary conditional inferences. We formulate the joint distribution induced by this construction and characterise the relationship between its forward and reverse conditionals. We investigate the resulting model for classification and conditional generation, including confidence calibration and consistency between generated samples and classifier predictions. Experiments on conditional datasets demonstrate competitive classification accuracy, well-calibrated confidence without post-hoc calibration, and classifier-consistent generation. We further explore the shared joint distribution as a basis for interpreting the relationship between discriminative and generative behaviour.

Primary area:
generative models
Keywords:
Flow Matching, Generative Models, Classification, Conditional Generation, Joint Distributions, Probabilistic Modeling, Calibration

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Mq5oOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Joint Flow Matching (JFM), where there is one flow over pairs (X,Y)(X, Y) that starts from (Xnoise,Ydata)(X_{\text{noise}}, Y_{\text{data}}) at t=0t = 0 and ends at (Xdata,Ynoise)(X_{\text{data}}, Y_{\text{noise}}) at t=1t = 1. Integrating forward generates XX given YY, and integrating backward classifies. The main claim (eqs. 5-7) is that both directions are the conditionals of one induced joint, so generation and classification are consistent by construction. The method is experimentally validated.

Critical strengths

The construction is simple and quite practical, just one U-net does both conditional generation and classification with reasonable accuracy. The derivation of the induced joint and its density is clear, though the construction itself is not new (see weaknesses).

Critical weaknesses

I don't find any significant difference in the generation and classification task, or in the training loss, between JFM and Symmetrical Flow Matching [1], mentioned in the introduction. The paper's claimed addition is the joint and the consistency argument, which, in fact, does not identify the induced joint with the data. Indeed, it only shows that generation and classification are the two conditionals of whatever joint the flow defines. This is Bayes' rule for that joint, so it holds even when the joint differs from the data distribution. There is no comparison in experiments as well and other minor points.

[1] Caetano, Francisco, Christiaan Viviers, and Fons Van der Sommen. "Symmetrical Flow Matching: Unified Image Generation, Segmentation, and Classification with Score-Based Generative Models." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40. No. 4. 2026.

Questions for authors

Does the optimum of the loss recover the true dependence between the two variables, or a weaker one with the same marginals? How large is the gap at the label noise used for the images?

The consistency argument assumes continuous densities. How does it apply to one-hot labels, where the reverse output is a real vector turned into a class by an argmax?

Reviewer_h5q8OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper proposes a flow matching model that attempts to unify generation and classification. The paper motivates the approach with symmetric flow matching (Caetano et al., 2026); a reverse-coupled flow can support unified generation and classification. Experiments on conditional datasets demonstrate competitive classification accuracy, well-calibrated confidence without post hoc calibration, and classifier-consistent generation.

Critical strengths

None

Critical weaknesses

The paper has inadequate theoretical novelty or empirical results to back up its claims:

  1. How is the theoretical construction different from SFM? It seems the authors changed a Flow Matching idea into a Continuous Normalizing Flow. However, they still use a flow-matching objective, so I don't see what differs between JFM and SFM.

  2. There is no quantitative analysis of JFM against any relevant baselines such as SFM.

Questions for authors

None

Reviewer_ts3jOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This work provides a theoretical analysis of flow-based frameworks for learning a joint distribution over paired variables such as images and class labels and investigates the consistency between the conditional distributions of both directions.

Critical strengths
  • This work provides a formalization for flow-based frameworks for learning a joint distribution over paired variables such as the one proposed by [1]. This is useful as it shows theoretically that the joint flow results in consistent conditionals without requiring additional regularisation or other constraints.

[1]: Caetano, Francisco, Christiaan Viviers, and Fons Van der Sommen. "Symmetrical Flow Matching: Unified Image Generation, Segmentation, and Classification with Score-Based Generative Models." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40. No. 4. 2026.

Critical weaknesses
  • The proof for the consistency of the conditionals in the joint flow and subsequent findings seems to be a direct product of the formulation, even though it is possible I have misunderstood part of the derivations.

  • There are no numerical comparisons for the experiments provided in the paper. How does it compare to previous flow-matching based classification methods given similar compute? Are there specific gains in training or inference time?

Questions for authors
  • For the spiral dataset: why is it expected that unconditional flow matching will be able to produce a correct label mapping? Since no explicit labels are provided, the model learns to transport the multimodal distribution to the spiral distribution. In the JFM formulation this label information is explicitly provided by the pairing. For this reason, a more appropriate comparison would be a conditional flow model, if this comparison can be provided.

  • Section 5.2 describes that JFM enables interpretability in the features and patterns that the generative model associates with a given class using the classification confidence. Can't something similar be achieved with a simple normalizing flow or diffusion model, by inspecting and comparing likelihoods of generated samples?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.