- 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
- The paper proposed a distillation method for discrete flow models for count data by parameterizing transition probability;
- 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
- There are some quantities appearing for the first time without definitions, e.g., in eq. 5 and in eq. 11.
- 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
- What is the dimension of the transition kernel ? Is the transition (kernel) matrix sparse (like the transition rate matrix )? If not, how do we handle the high-dimensional case?
- How should we understand "a Count Flow Map has diagonal generator " in Corollary 1?
- Are there any other distillation methods for discrete diffusion in existing works? Are they applicable to the application or simulation tasks in the paper?
- 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
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 .
2. Theoretical guarantees and practical insights for few-step generation.
Our theory provides three insights:
- 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.
- Accuracy–efficiency trade-off: shows that improved consistency can reduce the number of sampling steps needed.
- Coupling: Endpoint truncation error scales with , 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 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.
We thank the reviewer for the detailed questions.
1. Undefined notation. In the updated manuscript, we clarify that Y in Eq. (5) is the sampled state at time t, given Xs=x. We also define ε=1−τ 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 Qt is sparse, while the finite-time kernel Ks,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 J components. We sample directly from this mixture, requiring only 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=t. Specifically, Qt=limh↓0(Kt,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-leaping. In neural forecasting, at 8 NFE, Count-FM achieves similar population CRPS with unit-jump sampling (0.02246) and τ-leaping (0.02242). 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 0.02248 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