acceptodds
Under review as a conference paper at ICLR 2027

Generative Modeling with Flux Matching

Abstract

We introduce Flux Matching, a new paradigm for generative modeling that generalizes existing score-based models to a broader family of vector fields that need not be conservative. Rather than requiring the model to equal the data score, the Flux Matching objective imposes a weaker condition that admits infinitely many vector fields whose stationary distribution is the data. This flexibility enables a class of generative models that cannot be learned under score matching, in which inductive biases, structural priors, and properties of the dynamics can be directly imposed or optimized. We show that Flux Matching performs strongly on high-dimensional image datasets and, more importantly, that our added freedom unlocks a range of applications including faster sampling, interpretable and mechanistic models, and dynamics that encode directed dependencies between variables. More broadly, Flux Matching opens a new dimension in generative modeling by turning the vector field itself into a design choice rather than a fixed target.

Primary area:
generative models
Keywords:
generative models, score matching, flow matching, markov chain monte carlo, sampling

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_aPz4OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Flux Matching, a loss for learning nonconservative vector fields whose stochastic dynamics have the data distribution as their stationary distribution. In this fixed-distribution setting, experiments on toy Gaussian mixtures and RNA velocity explore controllable and interpretable dynamics.

The authors then extend the objective to noise-conditioned fields across perturbed data distributions, providing an alternative to denoising score matching for diffusion-style generation. This extension is evaluated through image generation and sampling acceleration on CIFAR-10 and CelebA, and through causal attention constraints for generating simulated trajectories.

Critical strengths

S1) The paper develops an approach to learning distribution-preserving dynamics beyond the score that is well-founded. The projected Fisher formulation provides a theoretical basis for allowing constraints to shape the learned vector field. This flexibility is a promising contribution, even though its practical benefits require stronger validation.

S2) The core idea is presented clearly, particularly the distinction between matching the score and matching the divergence of probability flux. This makes the proposed formulation and its relationship to score-based modeling easy to follow.

Critical weaknesses

W1) Theoretical. The motivation and positioning of the generalization need further development. Nonconservative dynamics preserving a target distribution are established in prior work [1]. Although the authors mention this literature and connections to transport-based generation in Appendices F and E, these relationships are central to assessing the contribution and should frame the main discussion. Appendix G leaves the identification of useful objectives and constraints to future work, while the current demonstrations do not convincingly establish benefits that justify the additional complexity and computational cost.

W2) Empirical. The image-scale sampling claims lack essential comparisons and convincing evaluation. Since the image experiments use noise-conditioned fields, Flow Matching [2] is an essential baseline for claims of efficient sampling. Energy Matching [3] additionally removes noise conditioning and targets the stationary distribution at image scale, including the same CIFAR-10 task. The reported gap, approximately FID 3–4 for these approaches versus 9 for Flux Matching, is substantial and requires explanation through direct comparison. Figure 6 relies on 1K-sample FID without uncertainty estimates and omits these relevant baselines, providing little support for the claimed sampling advantage.

References:

[1] Ma et al. A Complete Recipe for Stochastic Gradient MCMC. NeurIPS, 2015.

[2] Lipman et al. Flow Matching for Generative Modeling. ICLR, 2023.

[3] Balcerak et al. Energy Matching: Unifying Flow Matching and Energy-Based Models for Generative Modeling. NeurIPS, 2025.

Questions for authors

Q1) Can you isolate the RNA-velocity gains from the transcription-model change introduced in Eq. (41)?

Q2) Have you tested fixed-target image sampling without noise annealing?

Q3) Can you compare against/discuss Flow Matching [2] and Energy Matching [3] on CIFAR-10? The paper’s sampling-efficiency claims should be evaluated against transport-based approaches - this is the most important question for my assessment. Or alternatives if applicable (e.g. OT-CFM [4] and Equilibrium Matching [5] or more -- there are many papers that came after DSM that are relevant)

Q4) Figure 6 shows better FID with fewer predictor–corrector steps, but this could reflect better handling of large numerical steps or transitions between noise levels. What evidence attributes these gains to faster mixing toward a fixed stationary distribution?

[4] Tong et al. Improving and Generalizing Flow-Based Generative Models with Minibatch Optimal Transport. TMLR, 2024.

[5] Wang et al. Equilibrium Matching: Generative Modeling with Implicit Energy-Based Models. NeurIPS, 2026.

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

Score Matching fits a network fθ(x)f_{\theta}(x) to the score of a data distribution ∇log⁡pdata\nabla \log p_{data} via the Fisher Divergence, which makes ff a generative vector field. This work introduces Flux Matching with learns any ff of the form ∇log⁡pdata+v\nabla \log p_{data} + v with ∇⋅(pdatav)=0\nabla \cdot (p_{data} v) = 0, by matching only the divergence of the probability flux. This allows ff to absorb pdatap_{data}-preserving dynamics while ff remains a valid generative vector field. The additional freedom can be used to include application specific losses and architectural constraints in order to learn potentially interesting dynamics, not possible with Score Matching.

The authors introduce the projected Fisher divergence and derive an equivalent reformulation, the Flux Matching loss. Several approximations are chosen to arrive at a tractable estimator. The benefits of Flux Matching is demonstrated in several domains such as RNA velocity data, images, etc. and compared to Score Matching.

Critical strengths

This work deals with an important shortcoming of Score Matching, namely the strictness of learning one particular vector field while ignoring a larger class of solutions that are equally valid generative vector fields for the same data distribution. The paper is well motivated and makes clear which problem it tries to tackle.

The theoretical derivation of eq.(6) and Theorem 3.1 look correct. I believe this is a novel contribution and could be interesting to the general field of generative modelling as it opens the doors to new applications. The work does not aim at improving the generative capabilities but allows for a fundamental extension of the underlying training objective, which is a rare and valuable contribution.

The work formulates a theoretically well-defined but intractable objective of the “projected Fisher divergence” and derives a tractable approximation, which is used in practice. While it is difficult to determine the value of the chosen approximations, it is impressive that it results in a scalable recipe at a modest computational slowdown of 2−4×2-4 \times compared to vanilla Denoising Score Matching.

The experiments demonstrate that the implemented estimator of Flux matching can train a usable generative model at image-scale problems (sec. 4.3) and that the additional freedom it buys can be exploited by an architecture (sec. 4.5).

Critical weaknesses

Theory:

  • The paper does not demonstrate that the projected Fisher divergence beats the obvious, simple approach of minimizing E[rθ2]\mathbb E[r_{\theta}^2]. The entire motivation rests on the assumption that this naive approach is "sensitive to derivative-level artifacts" (l.161). While this may well be plausible, a simple experiment to test this assumption would provide valuable insights and encourage the reader to engage with the subsequent sections.
  • The projection operator Πflux\Pi_{flux} (l.163) is not adequately introduced even though it is crucial for understanding as it precisely removes the benign components that don’t need to be fitted to the ground-truth score. Please introduce it more clearly in the main text using a separate equation e.g. via the Helmholtz decomposition.
  • The authors attempt to explain the dense theoretical derivations in sections 3.2 and 3.3 (while moving the bulk of them to the appendix) and to summarize the key contributions; however, for an average reader lacking the necessary background knowledge, this is difficult to follow.

There is a large gap in motivating the necessary approximations that were taken to get from the Flux Matching loss eq.(6) to the estimate that is used in practice, algorithm 1:

  • For example, the Flux Matching loss requires access to the score of the data distribution, unknown during training and thus approximated via a kernel batch KDE. While this is common practice in recent works, I think, it can nonetheless pose a source of error, especially given that the score estimate appears non-linearly and in multiple terms (uu, rr, etc.). A proper analysis of batchsize is necessary to validate the estimate and why particularly σ2=10−3\sigma^2 = 10^{-3} was chosen for RNA velocity data. Other approximations that were used and partly motivated: Short horizon T=4σ2T = 4 \sigma^2, no ablation against the exact pathwise estimator of (∂xt/∂x0)T∇xtr(xt)(\partial x_t/\partial x_0)^T \nabla_{x_t} r(x_t) and only 4 MCMC steps.
  • To arrive at a trustworthy estimator, it is necessary to systematically establish the validity and accuracy of each approximation in isolation and in combination on suitable synthetic data with ground-truth. At the very least, a model that was trained should be evaluated by a more accurate estimate and the resulting error reported, but the paper does neither of that.

Lacking verification of the estimator on 2D toy problems:

  • Sec. (4.1) presents how Flux Matching differs from Score Matching on 2D toy-data and what this additional freedom in the vector field could look like. While this is insightful for intuition, it does not depict any actual training results. Fig. 2 shows three analytically constructed distribution preserving fields and fig. 4 evaluates their Score and Flux Matching losses, where the latter is trivially zero, while the practical estimate could be noisy or even biased.
  • It would be highly insightful to check that a trained model will converge to the ground truth. In 2D, training should not pose a computational problem and you could compare the exact loss with the estimate.
Questions for authors

I will raise my score if the rebuttal provides empirical evidence for the validity of the estimator as well as the approximations chosen:

  1. Models trained from random initialization on the 2d mixture of section 4.1 with algorithm 1. Report the exact projected Fisher divergence eq.(4), the implemented loss over training and the stationarity residual V(fθ)V(f_{\theta}) eq.(38). If possible, ablate over the approximations (batch-KDE vs exact score, TT, etc.) in isolation.
  2. For the trained image models please verify that the generative field is stationary via E[rθ2]\mathbb E[r_{\theta}^2] (evaluate with large batch KDE score), reported at a few σ\sigma values and compared with the Denoising Score Matching models.

Question:

  1. As a diffusion model, Flux matching gains freedom over score matching as the trajectories can curve differently than is dictated by the score of pσp_{\sigma}, which could be exploited for few-step ODE sampling. Why did you not attempt to straighten the trajectories, as in reflow, via an application specific loss?
Reviewer_bN9FOpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes to use general flux matching to learn generative models. In contrast, prior works focus on learning the score function, which is a special case of what's proposed in this paper. The paper suggests several benefits of matching flux (in different forms) compared to matching just the score function.

Critical strengths
  1. The paper is well-written, although it takes some effort for someone not fully invested in diffusion models or physics to understand the contribution of this paper.
  2. I believe that flux matching has potential if sufficient effort is made to understand and modify it suitably to improve its results, esp. in image generation.
Critical weaknesses

According to the paper, the primary motivation to consider flux matching when the current score matching already works well is to have controllability, interpretability, structure, etc. The results are also organized along these lines. However, it's hard to be convinced by the current evidence. What do these experiments mean for a practitioner? For example, can the authors show a real-world case where the degrees of freedom are varied and the score matching method is not sufficient? Similarly for Sec. 4.5

Questions for authors

Please see weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.