acceptodds
Withdrawn from ICLR 2027

Same Flow, Different Paths: Variance Reduction in Flow Matching

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In flow matching (FM), a velocity model is trained using a predefined path that connects data and noise samples (e.g., ). In this work, we study the choice of this path from an optimization perspective by analyzing the variance of stochastic gradients. We consider the class of paths that induce the same marginal distributions and marginal velocity field , and therefore the same FM objective. Our main finding is that the choice of path can fundamentally change the convergence rate of SGD, even when the FM objective remains exactly the same. (i) For a linear velocity model and one-dimensional Gaussian data, we derive a tight bound on the SGD iteration complexity up to logarithmic factors and find an analytically optimal path that minimizes this bound among linear paths inducing the same FM problem. (ii) We then extend the variance analysis to general FM problems and formulate path selection at a fixed as the variance-minimization problem PathOpt, constrained to . We show that this constraint is essential: reducing variance without it can lead to slower convergence. (iii) Since the constraint cannot generally be verified directly, we derive an equivalent formulation with constraints that can be estimated from samples, allowing paths to be found numerically. Our theoretical results are supported by experiments with Gaussian data, Gaussian mixture models, and real datasets.

Primary area:
generative models
Keywords:
flow matching, variance reduction, stochastic optimization, stochastic gradient descent, path optimization

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_9cWPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies following problem of optimizing the choice of interpolating curves in the context of flow matching. Given a probability density path ptp_t and a mimicking field vtv_t, one says that a family of interpolating curves gt(x0,x1)g_t(x_0,x_1) belongs G(pt,vt)G(p_t,v_t) if minimizing the flow-matching objective E[∣g˙t(x0,x1)−vt(θ,xt)∣2]E[|\dot{g}_t(x_0,x_1)-v_t(\theta,x_t)|^2] induces vtv_t. The question addressed here is how to choose optimally g∈Gt(pt,vt)g\in G_t(p_t,v_t) in such a way that, if we minimize the flow matching objective using stochastic gradient descent, then the number of iterations it takes to achieve a given precision is the smallest possible. One of the main results consists in a full solution to this problem in the case when $v(\theta,x)= \theta x $ and the data distribution $ p^{data} $ is a centered Guassian with arbitrary variance. For general distributions, the main contribution is to formulate the problem as an expectation minimization task with explicit objective. The constraint g∈Gt(pt,vt)g\in G_t(p_t,v_t) is described through a set of differential inequalities.

Critical strengths
  • The paper opens the discussion on an interesting problem that might be of practical relevance.
  • Explicit solution in a toy Gaussian example with explicit characterization. Moreover, the authors convincingly explain that choosing different gg can lead to substantial difference on the SGD performance.
  • The general formulation of the problem does not admit a closed form solution, but it admits a clean interpretation and can be tackled numerically
Critical weaknesses
  • The depth of the theoretical result is limited. In fact, only the Gaussian example is analyzed. Though I agree that the general case is intractable, one would like to see that the theoretical analysis extends, possibly in a weaker form, to other target data distributions such as log-concave laws for example. Appendix K contains a study of Gaussian mixtures, but the conclusion seems to be that there exists no optimal g in this case, if I understand correctly.

  • The description of GG in terms of differential inequalities is nice, but requires priori knowledge of some gˉ∈G\bar{g}\in G. Showing that such gˉ\bar{g} exists, or exhibiting one, appears to be a non trivial problem. This limits the applicability of Theorem 4.3.

Questions for authors
  • The solution to the quadratic problem defining the conditional velocity field in flow matching is given in terms of a conditional expectation. I would have expected conditional expectation to be a useful tool in the description of the admissible set GG. Can the authors comment on why this is not the case?

  • Can one expect problem (FEASIBLEPATHOPTθ) to admit an optimal solution? The Gaussian mixture example seems to suggest that this may fail even in simple settings.

Reviewer_aZLXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Within the flow matching (FM) framework, this paper studies how the choice of conditional paths affects stochastic gradient descent (SGD) convergence while preserving the same marginal distributions and marginal velocity field. It provides convergence analysis and an optimal path within the examined family for a one-dimensional Gaussian toy model with a linear velocity model. For general FM problems, it formulates variance-minimization criteria and sample-estimable constraints for checking gt∈G(pt,vt∗)g_t \in G(p_t, v_t^*). Experiments cover Gaussian and Gaussian mixture (GMM) toy models and real datasets. For real datasets, the authors use grid search with approximate constraints and variance-related scores to screen candidate paths, followed by FID evaluation.

Critical strengths

None

Critical weaknesses
  1. The motivation for the proposed variance reduction is unclear. The paper does not sufficiently explain what limitation of existing FM training motivates the method or why gradient variance is an important bottleneck. It remains unclear whether the primary goal is variance reduction itself, improved training efficiency, or better sample quality, and how the proposed method connects these objectives.

  2. Excessive background and technical detail obscure the research contribution. The opening sections read more like a technical report of definitions and derivations than a focused argument for a research contribution. Established FM/CFM relationships and optimization formulations receive substantial attention, while the limitations of prior methods, the research gap, and the distinction between background and new contributions remain insufficiently developed.

  3. Extensive theoretical analysis is built around a restricted one-dimensional toy model. A substantial portion of the paper analyzes Gaussian data with a linear velocity model, and the explicit convergence improvements remain confined to this setting. The multidimensional extension also relies on separability. The paper does not sufficiently explain what broader insights justify this emphasis or distinguish the initial Gaussian-path characterization from the subsequent complexity and optimal-path contributions.

  4. The practical roles of PathOpt and the kernel constraints are not sufficiently established. The PathOpt objective specifies what to minimize, while the kernel-based constraints provide a feasibility test. However, the paper does not clearly demonstrate how these components work together to produce reliable path selection in practice. In particular, the effect of the kernel constraints on candidate selection and the degree to which empirical tolerances preserve the reference FM objective are insufficiently documented.

  5. The main text lacks an overview and a complete pipeline. The method must be reconstructed from separate descriptions of the reference model, candidate construction, constraint checking, variance scoring, training, and final selection. Although algorithmic details appear in the appendices, the main text does not clearly present the complete FM algorithm or explain how its steps follow from the theory.

  6. The experiments emphasize toy models, while key validation in practical FM training is missing. Direct validation of the proposed mechanism is concentrated in Gaussian and GMM toy models. The real-data experiments do not adequately demonstrate how the method changes gradient variance and convergence in neural FM training, or how these changes relate to the reported FID improvements. Generated-image examples are also absent, limiting qualitative assessment of sample quality and diversity.

  7. General applicability is not established by the real-data experiments. Practical path selection remains restricted to a predefined four-parameter family and a finite grid search. The experiments do not examine whether the method remains effective under alternative path representations or search settings. The evidence therefore supports local improvements within a particular parameterization, rather than establishing a broadly applicable path-selection algorithm.

Questions for authors
  1. What does the chosen one-dimensional toy model reveal beyond standard Gaussian conditional paths and their velocity-field formulas?

  2. Is the method primarily intended to lower FID, improve training efficiency, or accelerate optimization convergence?

  3. What is the proposed FM algorithm, and can the authors provide an overview and a complete pipeline?

Reviewer_5XnHOpenReview ↗
Rating
1
Confidence
3
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies variance reduction in flow matching by choosing conditional paths. The authors show that different paths can produce the same marginal distributions and velocity field but different stochastic-gradient variances. For a linear model with Gaussian data, they derive tight SGD convergence bounds up to logarithmic factors and an analytically optimal path. They further propose PATHOPT and a feasible formulation with sample-estimable constraints. They conduct experiments on Gaussian data, GMMs, and five real image datasets.

Critical strengths
  1. Interesting theoretical perspective. The paper studies path selection from an optimization perspective. The variance decomposition clearly explains why different paths can yield different SGD convergence rates while preserving the same FM objective.

  2. Solid theoretical analysis in the simplified setting. The matching upper and lower bounds and the analytically optimal path are valuable results. The proposed feasible formulation also extends path optimization to more general FM problems.

Critical weaknesses
  1. Significant gap between theory and practice. The convergence analysis is limited to SGD with a linear velocity model and Gaussian data, while the real experiments use nonlinear neural networks trained with Adam/AdamW. Although the authors extend the variance analysis to general FM problems, there is no theoretical guarantee that minimizing PATHOPT improves convergence in these practical settings.

  2. Limited experimental evidence for practical efficiency on real data.

  • The improvements on real datasets are not consistent. In particular, Flowers-102 shows no improvement, and the gain on AFHQ-Cat is small and not statistically significant. Notably, these are the two higher-resolution, latent-space experiments in the paper (256×256 and 512×512), and are the closest to large-scale image generation among the evaluated settings.
  • Moreover, the paper mainly reports final FID rather than training convergence or wall-clock time. The additional cost of learning and selecting paths is not included, making the practical efficiency gains unclear.
Questions for authors

See Weaknesses

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

Related papers

Discussion (0)

Sign in to comment.