acceptodds
Under review as a conference paper at ICLR 2027

Stochastic Flow Map for Count Data

Abstract

High-dimensional count data are common in scientific applications, but most diffusion and flow models are designed for continuous or categorical data, and generation often requires many sequential model evaluations. We propose Count Flow Map, a generative model that learns finite-time transitions directly in count space for one- or few-step generation. Our model directly learns stochastic transitions over finite time intervals, using Poisson births and Binomial deaths to preserve nonnegative integer counts without a predefined maximum. These transition models are trained to match the underlying local birth–death dynamics and to maintain consistency across step sizes. We characterize the connection between local dynamics and finite-time transition consistency and derive a bound on the generation error. After validating Count Flow Map in several simulations, including a high-dimensional, high-count setting, we apply it to single-cell drug perturbation prediction and neural population forecasting, where it captures perturbation responses and supports forecasts of high-activity events with only one or a few model evaluations. Together, these experiments demonstrate that Count Flow Map enables high-quality generation directly in count space across inference budgets, from one-step to few-step generation, using a single trained model.

Primary area:
generative models
Keywords:
count data, deep generative modeling, flow matching, flow map, birth–death processes, few-step generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_NWxSOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a stochastic flow map for distilling flow matching models for count data. The authors introduced a consistency regularizer for count flow map training. The sampling process has zero discretization error with the one-step transition kernel (although there are estimation and approximation error for this kernel). The simulation results show that the proposed method with few-step generation is promising compared to Count-FM with large NFEs.

Critical strengths
  1. The paper proposed a distillation method for discrete flow models for count data by parameterizing transition probability;
  2. The experimental results including simulation and applications demonstrates the proposed distillation method for few-step generation has the competitive performance compared to the baseline Count-FM with large NFEs.
Critical weaknesses
  1. There are some quantities appearing for the first time without definitions, e.g., YiY_i in eq. 5 and ϵ\epsilon in eq. 11.
  2. The proposed method should also be compared with discrete diffusion models, such as discrete flow matching mentioned in the introduction, (in my opinion, if applicable).
Questions for authors
  1. What is the dimension of the transition kernel Ks,tK_{s,t}? Is the transition (kernel) matrix Ks,tK_{s,t} sparse (like the transition rate matrix QtQ_t)? If not, how do we handle the high-dimensional case?
  2. How should we understand "a Count Flow Map Ks,tK_{s,t} has diagonal generator QtQ_t" in Corollary 1?
  3. Are there any other distillation methods for discrete diffusion in existing works? Are they applicable to the application or simulation tasks in the paper?
  4. I am curious about whether the proposed method can beat the baseline Count-FM with other training-free high-accuracy (or high-order) samplers for few-step generation (e.g., with a budget of 8 NFEs). If applicable, can the authors report some results?
4 comments

We thank the reviewer for the detailed questions.

1. Undefined notation. In the updated manuscript, we clarify that YY in Eq. (5) is the sampled state at time tt, given Xs=xX_s=x. We also define ε=1−τ\varepsilon=1-\tau before the error bound.

2. Discrete diffusion comparisons. The original submission already evaluated D3PM and Discrete Flow Maps in both simulations. D3PM is a discrete diffusion baseline, whereas Discrete Flow Maps, a self-distilled extension of Discrete Flow Matching, learns categorical finite-time transitions. A more comprehensive comparison with other baselines, including discrete flow matching, has been reported in Count-FM [1].

3. Transition-kernel dimension and scalability. The generator QtQ_t is sparse, while the finite-time kernel Ks,tK_{s,t} is generally dense on the infinite count space. However, we never construct these matrices explicitly. Instead, for each input state and time interval, our model predicts the parameters of a Poisson–Binomial mixture with JJ components. We sample directly from this mixture, requiring only O(Jd)O(Jd) distributional parameters without enumerating possible states (Section 2.2 and Appendix A.2).

4. Diagonal generator. “Diagonal” refers to the time diagonal s=ts=t. Specifically, Qt=lim⁡h↓0(Kt,t+h−I)/hQ_t=\lim_{h\downarrow0}(K_{t,t+h}-I)/h, which recovers the infinitesimal local flow matching generator. We clarify this interpretation in Section 2.1 and strengthen the midpoint-identification argument in Appendix A.6.

5. Other discrete distillation methods. Discrete Flow Maps and Categorical Flow Maps are relevant finite-time methods for categorical states. Discrete Flow Maps is already evaluated in our simulations. These methods can represent counts using finite categories, but require a count-range restriction and do not inherently encode count ordering. Our method instead models unbounded counts directly.

6. Higher-accuracy Count-FM sampling. We evaluate both unit-jump sampling and binomial τ\tau-leaping. In neural forecasting, at 8 NFE, Count-FM achieves similar population CRPS with unit-jump sampling (0.022460.02246) and τ\tau-leaping (0.022420.02242). This suggests that the choice of solver has limited influence on Count-FM's forecast quality at this inference budget. In comparison, one-step Count Flow Map achieves population CRPS 0.022480.02248 in 2.33 ms, versus 5.21 ms and 6.34 ms for the two Count-FM samplers. But comparing against higher-order Count-FM solvers is interesting for future work.

References

[1] Ganchao Wei and John Pearson. Flow Matching for Count Data. NeurIPS, 2026. https://arxiv.org/abs/2605.07746

Thank you for your response and for clarifying the empirical performance. My main concern is whether the proposed transition kernel estimator can accurately estimate the true transition kernel in few-step generation. (Why don’t we just use flow maps for continuous flow matching?) Additionally, it seems difficult to see how the theory provides convergence or approximation guarantee for one-step generation. (What practical insights does the theory provide for few-step generation? Can the authors show when the proposed method works better than flow maps for continuous flow matching on count data?) I also see that the contribution beyond Count-FM seems to be limited (Reviewer f7dZ has a similar view on novelty).

We thank the reviewer for the follow-up and clarify 4 points.

1. Why not continuous Flow Maps?

Continuous flow models do not naturally preserve count structure. They operate in continuous space and generally require rounding or other discretization procedures to produce counts. Importantly, even small prediction errors near rounding thresholds can change the resulting counts (e.g., 0 versus 1), particularly affecting sparse data. In contrast, our method directly learns stochastic finite-time transitions on N0d\mathbb N_0^d.

2. Theoretical guarantees and practical insights for few-step generation.

Our theory provides three insights:

  1. Consistency: Accurate local rates alone do not guarantee accurate finite-time transitions. Under exact compositional consistency and local accuracy, the learned kernel recovers the finite-time transition of the learned generator, even in one step.
  2. Accuracy–efficiency trade-off: Emap≤2ηcompτ2/L\mathcal E_{\mathrm{map}}\le 2\eta_{\mathrm{comp}}\tau^2/L shows that improved consistency can reduce the number of sampling steps needed.
  3. Coupling: Endpoint truncation error scales with εEπ∥X1−X0∥1\varepsilon\mathbb E_\pi\Vert X_1-X_0\Vert_1, motivating OT coupling to reduce error without increasing sampling cost.

3. Empirical evidence.

Although we have not directly benchmarked continuous Flow Maps, existing results provide supporting evidence. First, Count-FM [1] outperforms continuous latent-space generative models (scDiffusion, scLDM, and scVI) on scRNA-seq generation in both W2W_2 and MMD, demonstrating the benefits of count-specific modeling. Second, our Count Flow Map achieves comparable or better quality than Count-FM with substantially fewer evaluations. It also outperforms Discrete Flow Maps in one-step generation. Together, these results suggest that Count Flow Map retains the advantages of count-specific modeling while substantially improving sampling efficiency.

4. Contributions beyond Count-FM.

While Count-FM learns infinitesimal transition rates, our method directly learns finite-time stochastic kernels through a count-specific parameterization and compositional self-distillation. This allows efficient one- or few-step generation without sequentially simulating the learned birth–death process.

[1] Ganchao Wei and John Pearson. Flow Matching for Count Data. NeurIPS, 2026. https://arxiv.org/abs/2605.07746

Thank you for the clarification. I will raise my score after reading your response. I still have some concerns about whether the proposed kernel is flexible enough to approximate the true transition in one step, even with unlimited data. The error bound decomposes the sources of error, but it is still unclear to me when we can expect few-step generation to be accurate.

Reviewer_wi6JOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Count Flow Map, a stochastic generative framework for efficient count-data generation. The method extends count flow matching from local birth–death dynamics to finite-time transitions using Poisson births and Binomial deaths. It combines diagonal rate matching with Chapman–Kolmogorov consistency to enable one- or few-step generation while preserving nonnegative integer counts. The authors also provide theoretical analyses of transition consistency and generation error. Experiments on synthetic count distributions, single-cell drug perturbation prediction, and neural population forecasting demonstrate competitive generation quality with substantially fewer sampling steps.

Critical strengths
  1. The proposed stochastic transition kernel provides a natural extension of count flow matching to finite-time generation while preserving the discrete and unbounded nature of count data.
  2. The method achieves strong low-NFE performance across synthetic and real-world datasets. The evaluation includes multiple baselines, generation-time comparisons, and two meaningful scientific applications.
Critical weaknesses
  1. The individual contributions of Chapman–Kolmogorov consistency, finite-time corrections, and mixture components are not sufficiently isolated.
  2. Count Flow Map uses substantially more parameters than Count-FM in the single-cell experiments, without matched-capacity or training-cost comparisons.
  3. The error bound contains an unmeasured finite-time approximation term, making its practical tightness difficult to assess.
  4. Neural forecasting shows considerable undercoverage, with only 64% empirical coverage at the nominal 90% level for one-step generation.
Questions for authors

Please see the weaknesses.

1 comment

We thank the reviewer for their positive feedback and useful suggestions.

1. Contributions of different pieces:

  • Compositional (Chapman–Kolmogorov) consistency: used for self-distillation in training (off-diagonal part), to encourage agreement between direct and composed transitions.
  • Finite-time corrections: allow birth and death parameters to depart from their first-order local-rate approximations over longer intervals, while preserving the local generator.
  • Mixture components: a cheap and tractable way to capture cross-coordinate dependence.

We clarify these roles in Sections 2.2–2.3 in the updated manuscript.

2. Model capacity and training cost. We add parameter-matched Count-FM results to Table 1 in the updated manuscript.

Count Flow Map has 14.52M parameters, compared with 5.36M for the original Count-FM and 14.50M for the matched version. Under independent coupling, one-step Count Flow Map achieves sliced W2=0.418W_2=0.418 versus 0.4490.449 for matched Count-FM. Under OT, the results are comparable (0.4010.401 versus 0.4050.405), but Count Flow Map requires only 1 NFE rather than 128. Therefore, even with matched parameter counts, Count Flow Map achieves comparable with substantially fewer sampling steps.

3. Finite-time approximation error.

In the updated manuscript, we derive an explicit finite-time error bound (Appendix A.7.3), under uniform local-accuracy assumptions, Emap(Π)≤2ηcompτ2/LE_{\mathrm{map}}(\Pi)\le 2\eta_{\mathrm{comp}}\tau^2/L, for LL equally spaced inference steps. Here, ηcomp\eta_{\mathrm{comp}} measures normalized worst-case midpoint inconsistency. The bound explains the roles of consistency and inference budget, but as pointed out by the reviewer, the numerical tightness remains unverified because ηcomp\eta_{\mathrm{comp}} is not estimated. Estimating this quantity is a potential direction for future work.

4. Neural forecast undercoverage. In the updated manuscript, we expand the calibration comparison in Appendix D.3. At nominal 90% coverage, Count Flow Map achieves 68.0% at 1 NFE and 72.1% at 16 NFE, compared with 72.7% for Count-FM at 16 NFE. Undercoverage therefore persists across both methods and remains a limitation.

Designing a conditional bridge with more stochasticity may help with this. Related bidirectional count perturbations have been explored in JUD [1], while the effects of path stochasticity have been studied in continuous flow matching [2]. We have added this possibility to the Discussion section.

References

[1] Yair Shenfeld, Ricardo Baptista, and Stefano Peluchetti. Jumping up and down: Denoiser diffusion models for discrete ordinal data. https://arxiv.org/abs/2610.02754

[2] Ganchao Wei and Li Ma. Stream-level Flow Matching with Gaussian Processes. ICML, 2025. https://openreview.net/forum?id=qg9p1I5lmp

Reviewer_f7dZOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces a generative model for count data. The goal is to improve the efficiency of generations by learning models for finite transitions. The resulting method is extensively evlauated in experiments, including two real biological data sets.

Critical strengths
  1. The papers experimental results are convincing and interesting. The results clearly demonstrate the improved inference-time efficiency of the proposed method.
  2. The writing is clear and the presentation of both the theoretical results and empirical results is strong.
Critical weaknesses

While the paper tackles a worthwhile goal and makes progress, the size of the research increment here appears to be somewhat moderate. I do not consider this a critical weakness, but am reporting this to give the other reviewers and AC a reflection of my view of the paper.

The theoretical results in the paper appear relatively standard adaptations of known ideas to the specific count setting. (They are still worthwhile.) If the authors disagree with this perhaps they can comment more on the novelty in the rebuttal period.

Questions for authors

See above regarding the novelty of the theoretical results.

2 comments

We thank the reviewer for their positive feedback.

We agree that the Kolmogorov equations and Chapman–Kolmogorov identity (for compositional consistency) are classical. Here, our contribution is to establish how a cheap and tractable Poisson–Binomial mixture can model finite-time count transitions with cross-coordinate dependence, while preserving the constructed local birth–death generator. We show that exact midpoint consistency and uniform local accuracy recover the finite-time transitions generated by the learned birth–death rates. We further derive an error bound for approximate consistency.

Thank you for the clarification regarding the theoretical novelty in the manuscript.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank all reviewers for their constructive feedback. We have revised the manuscript accordingly. The main updates are as follows:

  1. Method. We clarify the roles of finite-time corrections and mixture components, the meaning of the diagonal generator, and why sampling does not require constructing an infinite transition matrix.
  2. Theory. We clarify the distinction between classical Markov-process identities and our count-specific construction, and add a midpoint-identification result and an explicit finite-time error bound.
  3. Experiments.
    • Simulations and scRNA: We update the simulation results and add a parameter-matched Count-FM baseline to the single-cell comparison.
    • Neural population forecasting: We update the Count-FM sampling comparison and add comparative calibration results.

Dear Reviewers, The authors have now submitted their responses to the reviews. There is only a limited amount of time left for you to review their replies and provide any further comments.

Your feedback is crucial to ensure that the review process proceeds smoothly. We would greatly appreciate it if you could share your reactions before the discussion period closes, if you have not done so.

Thank you very much for your work. Best regards.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.