acceptodds
Under review as a conference paper at ICLR 2027

scMeanFlow: Fast, High-Quality, and Controllable scRNA-seq Synthesis via One-Step MeanFlow Matching

Abstract

Generative modeling of single-cell RNA sequencing (scRNA-seq) data shows great potential in alleviating sample scarcity and enhancing downstream analyses. However, existing deep generative models generally face critical limitations: they either operate on continuous approximations that ignore the inherently discrete and over-dispersed nature of raw counts, or rely on multi-step iterative sampling that severely restricts inference scalability. Furthermore, flexible condition injection mechanisms remain insufficiently explored. To address these challenges simultaneously, we propose scMeanFlow, an efficient two-stage generative framework. First, a Negative Binomial Autoencoder encodes discrete count vectors into a compact continuous latent space while explicitly preserving count-level statistical properties via a probabilistic decoder. Second, operating within this latent space, a MeanFlow network learns the interval-averaged velocity field for genuine one-step generation without ordinary differential equation (ODE) solvers, while Feature-wise Linear Modulation (FiLM) injects condition information through feature-wise affine modulation. Random condition dropout during training enables a single trained model to support unconditional, single-condition, and multi-condition synthesis. Extensive experiments on four real scRNA-seq datasets demonstrate that scMeanFlow achieves state-of-the-art generation quality. Notably, its inference speed is approximately 7x faster than the strongest flow-based baseline and over 45x faster than diffusion-based methods. Furthermore, scMeanFlow effectively improves rare cell type classification through data augmentation. Overall, scMeanFlow provides an efficient, count-aware, and controllable solution for single-cell data synthesis.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single cell, generation, meanflow

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

scMeanFlow uses an NBAE, MeanFlow and FiLM for efficient scRNA-seq generation. NB-AE is used for learning representations while preserving the discrete count generation ability. MeanFlow is used for one step generation. FiLM is used for condition information fusion. scMeanFlow showed competitive performance on generation quality and efficiency, compared with other baselines. The authors also demonstrate scMeanFlow's performance on multi-condition generation and rare-cell data augmentation tasks.

Critical strengths
  1. By combining several existing techniques, the model achieves over 45x speed compared with diffusion-based baselines.
  2. The author conducted comprehensive evaluation on several datasets, using statistical metrics, marker gene validation, and race-cell classification.
Critical weaknesses
  1. Limited novelty. NB-AE has been widely used in several works, including the scVI series and recent generation frameworks such as scDiffusion-X. Meanflow is used without substantial modification. FiLM is also an established method and shows relatively small improvements in the ablation study.

  2. Insufficient motivation. The primary advantage is generation speed, but it is hard to say that generation speed is a meaningful bottleneck. Even the slowest method generates 10000 cells within 110s.

  3. Most experiments focus on reproducing observed distributions, while the only downstream application is rare-cell classification, where the method is compared only against bootstrap resampling rather than competing generative models.

Questions for authors
  1. Can the author compare MeanFlow with standard Flow Matching using the same latent representation and backbone?
  2. How should the library size in Equation (10) be set during generation? How does this choice affect the generation results?
  3. Why does scDiffusion show a very poor marker gene expression correlation on PBMC 68k (Table 2) but outperform scMeanFlow on the same metric on Dentate Gyrus (Table 5)?
  4. How does the author unify discrete and continuous generated expression values from different methods?
Reviewer_4qC8OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes scMeanFlow, a two-stage scRNA-seq generator:

  1. A deterministic negative-binomial autoencoder maps raw counts into a continuous latent space.
  2. A conditional MeanFlow network maps Gaussian noise directly to that latent distribution in one network evaluation.
  3. FiLM and condition dropout support cell-type, site, joint-condition, and nominally unconditional generation.
  4. The NB decoder converts generated latents back into integer counts.

The principal claimed advantage is one-step sampling: 10,000 cells are reportedly generated in 0.72 seconds, while retaining competitive MMD, Wasserstein distance, and cell-type classification performance.

Critical strengths

Strengths

  • The motivation is practical and relevant: iterative diffusion or flow sampling can be unnecessarily expensive for large single-cell datasets.
  • Results in Table 1 are generally competitive. The improvements are strongest on PBMC3k KNNc and on MMD across the datasets.
  • Evaluating the full gene space rather than only highly variable genes is valuable.
Critical weaknesses

The architecture is primarily a combination of existing components:

  • Negative-binomial autoencoders for scRNA-seq are established.
  • FiLM conditioning and condition dropout are standard.
  • The interval-averaged velocity objective is taken from the published MeanFlow method.
  • More importantly, CFGen already combines latent flow matching, discrete likelihood decoders, multi-attribute conditioning, and rare-cell augmentation for single-cell generation. Consequently, the actual contribution is essentially: replace CFGen-style iterative latent flow matching with MeanFlow, and implement conditioning through FiLM. This may still constitute a useful application contribution, but it is not a new architecture or generative principle.
  1. The paper also lacks the most important ablation: ordinary flow matching and MeanFlow must be compared using the same autoencoder, network width, conditioning mechanism, training budget, and data. Comparisons against CFGen cannot isolate whether improvements arise from MeanFlow, architecture size, implementation, or training choices.

  2. Appendix C says that real and generated cells are “jointly projected” into a 50-dimensional PCA space. It is unclear whether:

  • a separate PCA is fitted for every method, or
  • one PCA is fitted using held-out real cells and generated cells from all methods.
  1. Equation 10 requires a library size ss for every generated cell, but the paper never explains how ss is obtained at inference.

  2. The multi-condition claim is supported only by UMAP visualizations for one observed site/cell-type combination. There are:

  • no quantitative multi-condition metrics;
  • no held-out site–cell-type combinations;
  • no evaluation of unseen combinations;
  • no comparison with CFGen’s multi-attribute generation.
  1. Appendix B describes:
  • MeanFlow hidden width: 64;
  • time and condition embeddings: 100 dimensions. Table 4 instead reports:
  • hidden width: 32;
  • embedding size: 20. These are material differences for both performance and runtime.
  1. The rare-cell experiment uses one split and only bootstrap resampling as the baseline. It needs repeated splits and comparisons with class weighting, conventional oversampling, and augmentation using other generators.
Questions for authors
  1. How is library size ss selected for a synthetic cell? Are any held-out library sizes used?
  2. Is PCA fitted separately for each generator, jointly across every generator, or only on real training data?
  3. Which real cells are used to train the KNN evaluator?
  4. How does MeanFlow compare with standard flow matching using exactly the same architecture and training configuration?
  5. Why are scVI and ACTIVA excluded from the runtime comparison?
  6. Which MeanFlow configuration is correct: width 64 and embedding 100, or width 32 and embedding 20?
  7. Can the model generate a site–cell-type combination entirely absent during training?
  8. How were marker genes selected, and was selection performed without test data?
Reviewer_owaXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes scMeanFlow a two stage generative framework for scRNA seq synthesis. In the first stage a Negative Binomial Autoencoder NB AE encodes discrete count vectors into a compact continuous latent space. It explicitly preserves count level statistical properties discreteness and over dispersion through an NB decoder. In the second stage a latent MeanFlow model learns the interval averaged velocity field. This supports genuine one step generation without ODE solvers or trajectory distillation. FiLM based conditioning with random condition dropout lets one trained model handle unconditional single condition and multi condition synthesis. Experiments show that this method can greatly improve inference speed and boost classification performance.

Critical strengths

1.It effectively resolves the speed quality trade off. By leveraging MeanFlow the model achieves genuine one step generation with state of the art fidelity and notable inference speedup. It can generate large numbers of cells rapidly.

2.FiLM with stochastic condition dropout lets one model handle unconditional single and multi condition generation without retraining. Ablations show it outperforms additive conditioning.

Critical weaknesses

1.scDiffusion marker correlation and kNNc of some baselines are unusually low. The authors are suggested to report baseline tuning details.

2.The latent space uses a deterministic non variational encoder. Yet generation starts from a Gaussian prior. Theoretical or empirical analysis of why this transport aligns well is needed for clarification.

3.The setting of library size s at generation time needs further description and clarification. MMD values reported as ±0.00 should be given with higher precision to distinguish methods.

Questions for authors

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.