acceptodds
Under review as a conference paper at ICLR 2027

Equilibrium Flow Maps: Compactified Generative Dynamics And Terminal Projection

10 views

Abstract

Flow matching has emerged as a powerful framework for generative modeling, but sampling typically requires numerically integrating a learned time-dependent vector field, requiring many network evaluations. Flow maps accelerate generation by learning the ODE solution operator directly, replacing many small numerical integration steps with a few large advances along the generative trajectory. Separately, recent methods such as Equilibrium Matching replace the time-dependent dynamics of flow matching with an autonomous vector field with data as equilibria, yielding dynamics with a natural infinite time horizon. These two directions suggest complementary goals: efficient generation with few network evaluations and principled iterative refinement as additional inference-time compute becomes available. To address these two goals simultaneously, we introduce Equilibrium Flow Maps (EqFM), which combine flow map amortization with autonomous equilibrium dynamics in a single model. EqFM compactifies the infinite-horizon autonomous dynamics onto a bounded domain, so the same network can advance a sample by any duration along the flow, from short refinement steps to a direct jump to equilibrium. We construct an autonomous field through infinite-time stochastic interpolants and derive complementary Lagrangian, Eulerian, and semigroup characterizations that lead to practical flow-map objectives for both direct training and distillation from an autonomous teacher. EqFM supports inference from one-step generation to iterative refinement without changing the underlying model. On ImageNet-256, EqFM improves from FID at 1 NFE to at 8 NFE, approaching its 500 NFE teacher at .

Primary area:
generative models
Keywords:
Generative Models, Equilibrium Models, Flow Matching, Flow Map, Diffusion

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_wiAiOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a one-step generator framework, which combines the flow map method and equilibrium matching theorem. The authors defines the infinite-horizon interpolant of autonomous generative transport and developed a stabilized CFG for the autonomous field. Experiment results on CIFAR-10 and ImageNet shows that the proposed method can achieve decent FID with very few inference steps.

Critical strengths

The motivation for the proposed method is natural. Autonomous flows depend only on duration, therefore flow maps carry over directly once equilibrium matching's built-in clock compactifies the infinite horizon.

Critical weaknesses

The current experiments are not enough. The ImageNet experiment only have one training run. And the training budget for MeanFlow and Equilibrium Flow Map are not controlled to be the same, as the authors use a third party checkpoint for MeanFlow. Some important baselines like improved MeanFlow are missing.

Questions for authors

Theorem 3.3 and training only handles \sigma \in [0,1), but the inference uses \sigma=1. Any special treatment or discussion for this?

On ImageNet 256, why not choose Improved MeanFlow instead of MeanFlow as a baseline? Improved MeanFlow achieved 1.72 FID with one step.

2 comments

We thank the reviewer for the careful reading and for recognizing the motivation for autonomous flow-map learning. We address the endpoint treatment, improved MeanFlow comparison, and provide new experiments compute-matched with the literature on CIFAR-10.

W1 Training and inference at σ = 1

Thank you for highlighting this point. Theorem 3.3 characterizes finite-duration maps for σ < 1, while the terminal map at σ = 1 is defined by the infinite-time limit established under our transport assumptions in Theorem 3.1. Φ_1 is therefore defined as the continuous extension of the map defined in Theorem 3.3.

In practice, we use a numerically stabilized objective that permits training at σ = 1, and include this terminal duration during training. Inference therefore evaluates an endpoint that is covered by training. The network predicts the terminal displacement directly, without converting σ = 1 into infinite autonomous time or numerically integrating to infinity. We will explicitly describe this endpoint treatment and distinguish the ideal theoretical objective from its stabilized implementation. Additional details have been added in Appendix C

Q1 Comparison with improved MeanFlow

We appreciate the suggestion to include improved MeanFlow 1 and have added its published results to Figure 2. We acknowledge its stronger one-step result of FID 1.72, compared with our 2.90.

Our implementation does not adopt the full set of architectural, objective, and training refinements used by improved MeanFlow, and our training budget is substantially smaller. Our EqFM training stage uses 400k updates, approximately 80 epochs, whereas the reported improved MeanFlow-XL model is trained for 800 epochs. EqFM also is instantiated from an autonomous-field checkpoint, which we report separately. Consequently, these results should be interpreted with the substantial differences in training recipe and budget made explicit. We emphasize however, our benefits over iMF in our monotonically improving results at larger inference compute budgets.

We selected MeanFlow 2 as the primary reference to keep the comparison focused on the autonomous flow-map construction. Improved MeanFlow remains an important broader baseline, and its results are now included on that basis.

Q2 Training budgets, inference budgets, and additional experiments

We agree that the ImageNet comparison is not fully matched in training compute. Figure 2 compares several methods at matched inference budgets, including EqFM and the evaluated MeanFlow checkpoint up to 128 NFE. This separates the question of sample quality at a given inference budget from the cost of training each model.

We also run improved MeanFlow 1 on CIFAR-10 under the same controlled training budget to establish a fair comparison with the results published in Boffi et al. We choose Boffi et al. 3 as a baseline to provide the most direct comparison, because EqFM extends their flow-map learning framework to autonomous, infinite-horizon dynamics.

Method1 NFE ↓2 NFE ↓4 NFE ↓8 NFE ↓16 NFE ↓
Lagrangian self-distillation8.104.373.343.333.57
Progressive self-distillation, midpoint12.818.435.965.074.64
Progressive self-distillation, uniform13.617.956.035.325.16
Improved MeanFlow, controlled budget, different architecture3.884.076.874.883.42
EqFM4.834.153.012.962.93

Our new CIFAR-10 experiments demonstrate our improvements over baselines with the same architecture and training compute budget, and we include an improved MeanFlow baseline trained with a similar budget, although this uses a slightly different architecture. Our results show our inference time scaling abilities given larger inference budgets despite having weaker few NFE performance than Improved MeanFlow.

We distinguish the new controlled CIFAR-10 comparisons from the broader ImageNet comparisons and clearly report the training and inference budgets for each.

References

[1] Geng et al., “Improved Mean Flows: On the Challenges of Fastforward Generative Models”, 2025.

[2] Geng et al., “Mean Flows for One-step Generative Modeling”, 2025.

[3] Boffi et al., “How to build a consistency model: Learning flow maps via self-distillation”, 2025.

Thank the authors for explaining the treatment for σ=1\sigma=1 and the additional experiments on CIFAR 10. I will raise my score by 1 point.

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

The authors consider learning an ODE without time conditioning (autonomous transport) that, in the infinite-time limit, transports the original noise density to the data density. Specifically, they consider a bijection rr between the time intervals τ∈[0,∞)\tau \in [0,\infty) and t∈[0,1)t \in [0,1) and define an interpolation in τ\tau-time as Iτ=x0+rτ(x1−x0)I_\tau = x_0 + r_\tau(x_1-x_0), which converges to x1x_1. They then use a model b^\hat{b} without time conditioning to predict I˙τ\dot{I}_\tau. They show that this model defines ODE dynamics that map points onto the data manifold in the limit. Furthermore, they use standard flow-map semigroup, Lagrangian, and Eulerian properties to learn a map for an arbitrary time step. The authors check the performance of the proposed method on ImageNet 256x256 and CIFAR-10.

Critical strengths

The authors propose a new theoretically justified method for autonomous transport on an infinite-time horizon. Further, they show that standard flow map learning can be applied to that method to learn a map for an arbitrary time step instead of long ODE integration with ODE solvers. Though the concurrent work BTM did similar research on the autonomous transport and flow maps for it and was already accepted at NeurIPS 2026 and published on arxiv around 1.5 months before the deadline for ICLR. So I see a risk that this is follow-up work on BTM and not concurrent work, which might limit theoretical significance.

Critical weaknesses

The relation between theory and practice is not sufficiently studied* In practice, the authors consider time cutoff and use interval [0, 1 - eps), which may change the optimal velocity field learned by the method and break the theoretical guarantees. Without such a cutoff, there might be numerical problems with the estimation of the true objective due to the 1/c(t) part. Furthermore, there are no details on which eps was used in the experiments. The analysis on how this cut-off changes the stability of the method and biases the learned field is needed.

The experiments do not clearly show improvements of EqFM. First, CIFAR-10 experiments are conducted under a very short 6 GPU-hours budget, without sufficiently motivating this choice. The reported FIDs are in the range of 8-16, which are quite high compared to the standard range of 1.5-3 in the standard one- or few-step methods. That might indicate that the method has not converged yet. The convergence analysis and longer runs are needed to understand the behavior of the method and its performance. Current FID results for CIFAR-10 are very poor.

The ImageNet comparison is also very limited. First BTM does not use CFG, while EqFM uses it. So the provided comparison is not fair. MeanFlow, while being. The authors also ignore numerous other established methods (e.g. [1-11]) for one-step and few-step generation for both CIFAR and ImageNet, making it hard to understand how the proposed method relates to other research in the field.

Examples of relevant baselines include:

[1] Tong, S., Ma, N., Xie, S., and Jaakkola, T. Flow Map Distillation Without Data. arXiv:2511.19428, 2025.

[2] Chen, H., Zhang, K., Tan, H., Guibas, L., Wetzstein, G., and Bi, S. π-Flow: Policy-Based Few-Step Generation via Imitation Distillation. arXiv:2510.14974, 2025.

[3] Zhou, L., Ermon, S., and Song, J. Inductive Moment Matching. ICML, 2025.

[4] Zhang, H., et al. AlphaFlow: Understanding and Improving MeanFlow Models. arXiv:2510.20771, 2025.

[5] Geng, Z., et al. Improved Mean Flows: On the Challenges of Fastforward Generative Models. arXiv:2512.02012, 2025.

[6] Song, Y., and Dhariwal, P. Improved Techniques for Training Consistency Models. ICLR, 2024.

[7] Lu, C., and Song, Y. Simplifying, Stabilizing and Scaling Continuous-Time Consistency Models. ICLR, 2025.

[8] Geng, Z., Deng, M., Bai, X., Kolter, J. Z., and He, K. Mean Flows for One-Step Generative Modeling. NeurIPS, 2025.

[9] Shlenskii, D., Gushchin, N., Novitskiy, L., Dylov, D. V., and Korotin, A. Midpoint Generative Models. arXiv:2605.29920, 2026.

[10] Song, Y., Dhariwal, P., Chen, M., and Sutskever, I. Consistency Models. ICML, 2023.

[11] Kim, D., et al. Consistency Trajectory Models: Learning Probability Flow ODE Trajectory of Diffusion. ICLR, 2024.

Questions for authors

Please address weaknesses

3 comments

We thank the reviewer for the careful reading and for recognizing the theoretical motivation and practical potential of autonomous flow-map learning. The reviewer raises important questions about the relation between our guarantees and implementation, the experimental budget, and the comparison with existing methods. We address these points below with clarifications and additional experiments.

W1 The relation between theory and practice near t=1

The autonomous field training used times included in [0,1- epsilon] with epsilon=1e-3, the diagonal term trains using times included in [0,1] but replaces the denominator c(σ) by (c(σ)+1e-2). These clarifications have been added to the paper.

We thank the reviewer for pointing out that numerical stabilization can bias the learned field. In Appendix C, we analyze both the time cutoff and bounded loss weighting, bound their effects on the population-optimal field away from the data manifold, and show that these bounds vanish as the regularization is removed. We also provide small-scale experiments to illustrate the practical tradeoff where weaker regularization reduces bias but can produce much larger gradients and make training less reliable.

W2 experiments supporting the benefits of EqFM

The reviewer is right that our original short training run alone cannot establish convergence or competitive final performance. A larger compute-matched experiment on CIFAR-10 has been added to support this. We match the training budget of Boffi et al. 1 and show improvement over the reference flow-map method at every reported NFE under these training settings. We choose Boffi et al. 1 as a baseline to provide the most direct comparison, because EqFM extends their flow-map learning framework to autonomous, infinite-horizon dynamics.

Method1 NFE ↓2 NFE ↓4 NFE ↓8 NFE ↓16 NFE ↓
Lagrangian self-distillation8.104.373.343.333.57
Progressive self-distillation, midpoint12.818.435.965.074.64
Progressive self-distillation, uniform13.617.956.035.325.16
Improved MeanFlow, controlled budget, different architecture3.884.076.874.883.42
EqFM4.834.153.012.962.93

We agree that this comparison does not establish state-of-the-art CIFAR-10 performance or demonstrate that training has fully converged. In the revised draft, we report the training budget and checkpoint selection explicitly and distinguish the controlled comparison from broader comparisons with methods trained under different recipes and budgets.

We have further run an extra experiment using the Improved MeanFlow recipe from Geng et al. 2 run on the same training budget as an extra comparison point and demonstrate our improved results as the inference budget increases. We emphasize to the reviewer, however, that the architectures are different to adhere to their published methodology.

Relation to BTM and the scope of our contribution

We agree that BTM 3 deserves clear attribution. There is substantial overlap in the study of autonomous transport, and BTM also discusses flow maps and iterative application. Thus, we carefully cite their theoretical contributions in our work. EqFM’s contribution is to formulate a bounded-duration family that includes both finite advances and the infinite-time equilibrium, derive origin and interval objectives for learning this family, and develop guidance rules for the resulting autonomous dynamics. We will make these distinctions more precise. To address concerns about concurrency, while drafting the original manuscript BTM was only recently publicly released, but as several months have since passed, we now cite them as a previous work.

W3 Guidance in the BTM comparison

We thank the reviewer for identifying the guidance mismatch. BTM’s 3 reported one-step map result is unguided, whereas our headline EqFM result uses guidance, so those numbers alone do not isolate the effect of the flow-map method. We have added unguided EqFM results on class-conditional ImageNet-256 on which EqFM still significantly outperforms BTM: FID 9.79 versus BTM’s 17.58 at 1 NFE, and 5.34 versus 16.53 at 5 NFE. These results appear in Section 5, “ImageNet-256 generation,” and the autonomous-model comparison table. We retain the guided EqFM results because guidance is also part of our contribution; its effect is examined in Section 5, “Classifier-free guidance,” and Appendix D.1. We interpret this benefit to our training of an "any advancement" flow map that has a richer training objective and enjoys more flexible sampling.

W4 Broader comparisons with one-step and few-step models

We appreciate the reviewer pointing out that the original comparison was too narrow. We have expanded the discussion and include the following published ImageNet-256 results alongside our updated EqFM results. The figure adds published results for Improved MeanFlow 2, AlphaFlow 4, IMM 5, and the non-REPA variants of FreeFlow 6 and π-Flow 7. We exclude results whose training uses REPA 8 or other pretrained perceptual features. These published points provide context, but their architectures, training budgets, and evaluation setups differ, so we do not present them as controlled training-budget comparisons. EqFM’s advantage in our experiments is its monotonic improvement as the same model receives more inference computation, from FID 2.90 at 1 NFE to 1.96 at 8 NFE and 1.87 at 64 NFE.

We selected MeanFlow 9 as the primary reference because it provides a standard time-dependent formulation against which to present our autonomous construction. We choose to not adopt additional objective-design and curriculum choices as presented in papers such as AlphaFlow 4. Incorporating those changes into EqFM alone would make it harder to attribute improvements to the autonomous construction. Nevertheless, this does not diminish the relevance of other baselines, which are now included in the comparison, and leave their objective design and curriculum choice applications to Equilibrium Flow Maps a relevant research direction.

References

[1] Boffi et al., “How to build a consistency model: Learning flow maps via self-distillation”, 2025.

[2] Geng et al., “Improved Mean Flows: On the Challenges of Fastforward Generative Models”, 2025.

[3] Lee et al., “Beckmann Transport Models: From Autonomous Flows to One-Step Maps”, 2026.

[4] Zhang et al., “AlphaFlow: Understanding and Improving MeanFlow Models”, 2025.

[5] Zhou et al., “Inductive Moment Matching”, 2025.

[6] Tong et al., “Flow Map Distillation Without Data”, 2025.

[7] Chen et al., “π-Flow: Policy-Based Few-Step Generation via Imitation Distillation”, 2026.

[8] Yu et al., “Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think”, 2025.

[9] Geng et al., “Mean Flows for One-step Generative Modeling”, 2025.

Thank you for the rebuttal!

Together with the provided clarification, the submission becomes stronger. Though I am still leaning to keep the current score because I share remaining concerns of the reviewer AmN9 regarding proper comparison, and given the fact that the original submission was too raw in the experimental section and required major revision during the rebuttal.

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

This paper proposes Equilibrium Flow Maps (EqFM), which learn to make large jumps along autonomous generative dynamics. The method maps the infinite time horizon to a bounded duration and learns both intermediate updates and a direct jump to equilibrium. The authors derive three training objectives and evaluate the method on CIFAR-10 and ImageNet-256. The results show that EqFM can substantially accelerate an autonomous teacher and improve sample quality with additional network calls.

Critical strengths
  1. The paper provides a coherent way to learn flow maps for autonomous dynamics. Time reparameterization makes it possible to represent the endpoint explicitly, and the tangent condition connects the field and map objectives naturally. I think this is a useful extension of existing flow-map methods.

  2. The multistep results are promising. On ImageNet-256, EqFM reaches FID 2.09 with eight calls, close to its teacher’s 2.00 with 500 calls. The compute-matched CIFAR-10 results also show improvements over MeanFlow at two, four, and eight calls. These results suggest that the learned maps can make effective use of additional inference computation.

Critical weaknesses
  1. I am still not fully convinced about when EqFM should be preferred over conventional flow maps. The perturbation results in Appendix D are interesting, and I think this motivation deserves more attention in the main text, with further experiments and analysis. In particular, it would help to compare autonomous and time-dependent flows both before and after flow-map distillation. This would clarify whether the observed advantage is already present in the underlying dynamics, or whether distillation preserves, strengthens, or weakens it.

  2. The large-scale comparison is useful, but the different training budgets make the absolute performance numbers difficult to compare fairly. I would like to see comparisons with more relevant baselines under matched training and inference budgets. It would also help to clarify where EqFM is most useful: does it sacrifice one-step quality for better results with several or many calls? The CIFAR-10 results suggest such a trend, but a broader comparison across one-step, few-step, and higher-call budgets with more baselines is needed to understand it.

Questions for authors
  1. One appealing feature of EqM is adaptive sampling, where the stopping point can depend on the current sample. Does EqFM retain this advantage? Could you adaptively choose the number of map evaluations for each sample, and how would this compare with the fixed-call schedules used here?

  2. Are there better sampling or optimization schemes that specifically exploit the autonomous dynamics? Can they also be used with the learned EqFM maps?

  3. If further large-scale experiments are not feasible during rebuttal, could you compare against a broader set of flow-map and one-/few-step baselines on CIFAR-10 under matched compute? Results across different call budgets would help address Weakness 2.

I would be willing to raise my score if the responses and additional evidence clarify these points.

4 comments

We thank the reviewer for the constructive assessment and for recognizing the value of autonomous flow-map learning and its multistep results. We agree that the paper should better explain when EqFM is useful and distinguish matched training comparisons from comparisons across different training setups. A larger compute-matched experiment on CIFAR-10 has been added to support this.

W1 When should EqFM be preferred over conventional flow maps?

We offer Equilibrium flow maps as an alternative to time-dependent flow maps with simpler parameterization and more stable performance scaling with inference time compute budget. The perturbation experiment in Appendix F.6 provides an important understanding of how our model is robust to accumulation of errors. Below, we examine its inference-time scaling and the role of explicit time conditioning in conventional flow maps.

Both conventional flow maps and EqFM support multistep generation, but time-dependent models must supply a model time at every step. Saied and Fleuret 1, §5 provide concrete evidence that this input can matter: they measure population time lag with a validated probe (their Figure 2), and changing only the model-time input reduces ImageNet-256 FID from 13.8 to 8.8 for SiT-XL/2 at 50 NFE (§5.2 and Appendix D.4). The same intervention improves an ADM U-Net from 22.7 to 16.7 (their Appendix Table 7).

We test a related, per-sample correction in Appendix E.7, Figure 5. A separately trained network estimates the current state’s noise level and supplies it as MeanFlow-XL/2’s time input 2. At 32 MeanFlow steps, the degradation is improved with the new time correction, but still worsens as more steps are added. EqFM, however, improves over both baselines monotonically improving past 32 NFE.

EqFM needs no prescribed model-time input. Its FID instead decreases as the number of steps increases, consistent with its ability to continue advancing a sample along the same autonomous dynamics.

While our new results rely on self-distillation to learn the flow map, it is possible to measure the performance of the underlying field using the boundary condition v₀(x) = b(x) to sample the autonomous time field. Similarly, Mean Flow does not use a teacher and similar process can be applied. We find that at 250×2 NFEs, using classifier-free guidance with coefficient 1.5, Mean Flow reaches an FID of 3.16 while EqFM reaches 2.92. For EqFM, we integrate ẋ(τ) = b(x(τ)) = v₀(x(τ)). For MeanFlow, we integrate its diagonal field, ẋ(t) = uθ(x(t), t, t).

These results suggest comparable underlying-field quality, with a modest advantage for EqFM. Appendix F.8 further investigates the role of time-input mismatch: correcting MeanFlow’s time input improves FID but does not eliminate its degradation as the inference budget increases, whereas EqFM continues to improve without requiring an absolute-time input.

W2/Q3 Training and inference budgets

We appreciate the request to distinguish training and inference budgets more clearly. We direct the reviewer to Figure 2 comparing several methods at matched inference budgets, including EqFM and MeanFlow 2, improved MeanFlow 3, and other works across budgets up to 128 NFE. This allows sample quality to be assessed at equal numbers of evaluations. The updated EqFM ImageNet comparisons are now from self-distillation results, giving a more equal comparison to other results though admitted not fully matched in training compute because initialization, training recipes, and budgets differ. We make this distinction explicit in our updated submission.

In the revised manuscript, we report compute-matched self-distillation CIFAR-10 experiments at 1, 2, 4, 8, and 16 NFE. Boffi et al. 4 provide the most direct baseline because EqFM extends their flow-map learning framework to autonomous, infinite-horizon dynamics.

Method1 NFE ↓2 NFE ↓4 NFE ↓8 NFE ↓16 NFE ↓
Lagrangian self-distillation8.104.373.343.333.57
Progressive self-distillation, midpoint12.818.435.965.074.64
Progressive self-distillation, uniform13.617.956.035.325.16
Improved MeanFlow, controlled budget, different architecture3.884.076.874.883.42
EqFM4.834.153.012.962.93

At the same training compute budget, our EqFM results show stronger performance compared to Boffi et al. and offer iMF as a separate baseline for a compute matched comparison.

The CIFAR-10 experiment suggests EqFM is competitive and can outperform classical flow map training. Our CIFAR-10 comparison against iMF demonstrates a slightly weaker but competitive one step/two step NFE result, but demonstrate more stably improving FID as inference budget increases. We emphasize to the reviewer, however, that the architecture is not held constant for this comparison to abide by their published methodology.

Q1 Does EqFM retain adaptive sampling?

Yes, EqFM permits sample-dependent stopping and duration selection. A sampler can repeatedly apply finite-duration maps and monitor when a stopping criterion is met.

In Appendix D.4 (Fixed-point sampling), we monitor when the image reaches a stationary point quantified by x_k - x_k-1. This reaches FID 1.99 at an average of 19.7 NFE on ImageNet-256, matching a fixed 20-NFE schedule. A looser stopping threshold reaches FID 2.19 at 14.3 NFE on average, compared with 2.23 at a fixed 14 NFE.

Q2 Sampling schemes that exploit autonomous dynamics

Autonomy gives us flexibility in how we allocate sampling steps. Each EqFM advance depends on the current state and a chosen duration, so jumps of different sizes can be composed without tracking an absolute time, yielding fixed point adaptive samplers like above.

We also tested tuning jump sizes on the same ImageNet-256 checkpoint. FID improves from 2.28 to 1.85 at 2 NFE and from 1.96 to 1.67 at 8 NFE, without retraining, demonstrating improved performance over the fixed point samplers above.

References

[1] Saied and Fleuret, “Velocity Scaling in Flow Matching”, 2026.

[2] Geng et al., “Mean Flows for One-step Generative Modeling”, 2025.

[3] Geng et al., “Improved Mean Flows: On the Challenges of Fastforward Generative Models”, 2025.

[4] Boffi et al., “How to build a consistency model: Learning flow maps via self-distillation”, 2025.

Thank you for the additional clarification, especially the additional experiment with the estimated noise level to correct MF timestep input. I’m leaning toward to increase my score to 3. Please revise the manuscript to highlight this motivation more clearly in the main text.

AuthorsOpenReview ↗

Manuscript revision

We have updated Section 5.2 of the revised manuscript to emphasize this motivation and experiment, as suggested. We appreciate your constructive feedback throughout the discussion.

Reviewer_AmN9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Equilibrium Flow Maps (EqFM), a method that combines time-independent flow matching with few-step flow-map generation. The general context for this work is that in standard flow matching, the velocity network depends on both the current state and the time step, and typical flow maps (such as MeanFlow) must condition on two time variables (start time and end time) to take large jumps. Recent equilibrium-based methods showed that explicit time conditioning can be removed because a noisy sample implicitly encodes its own noise level, allowing a time-independent velocity field to guide samples toward an equilibrium point on the data manifold where the velocity vanishes. However, integrating this time-independent field still requires many small steps at inference time.

EqFM combines these two general directions. To eliminate this multi-step integration bottleneck, EqFM trains a flow-map network (via self-distillation, similar to consistency models and MeanFlow) to directly predict the integrated displacement of a trajectory over a large jump rather than the instantaneous velocity. EqFM exploits time-independence so the network only needs a single step-size parameter in the unit interval indicating how far toward equilibrium to jump. This enables a trade off between quality and compute: a step size of zero allows to recover the instantaneous velocity field, a step size of one jumps directly from noise to the final equilibrium on the data manifold in a single network evaluation.

The authors also derive a loss weighting needed for the time-independent velocity field to slow down properly near the data manifold, and propose a tweak for Classifier-Free Guidance (CFG), which otherwise becomes unstable in time-free models at high guidance scales. The method is evaluated on CIFAR-10 from scratch and on ImageNet-256 by distilling a 500-step time-independent teacher.

Critical strengths

The paper makes two nice conceptual and practical contributions. First, they observe that a flow map for a time-independent velocity field only needs a single step-size parameter (how far to jump toward equilibrium) rather than two time parameters (start and end time). This provides a unified view connecting Equilibrium Matching, one-step transport models, and flow-map distillation. Second, the analysis of why Classifier-Free Guidance (CFG) fails at high guidance weights in time-independent models is insightful and practically useful. Because the network is not given the noise level explicitly, the conditional and unconditional branches can implicitly estimate different noise levels and push samples away from the data manifold at large guidance scales, which the authors fix with norm-matched or orthogonal guidance updates.

Critical weaknesses

I have a couple of concerns:

  1. Relation to Drifting. Generative Modeling via Drifting (Deng, Li, Li, Du, and He, Feb 2026). The paper cites several later 2026 preprints and builds directly on MeanFlow (Deng and He) and EqM / BTM (Du). However, it does not cite or discuss Generative Modeling via Drifting which is from the same time period and closely related to the current submission. Like EqFM, Drifting Models avoid explicit time conditioning and enable single step generation. In particular, those methods generate a drifting field that vanishes at equilibrium on the data distribution, folding the iterative trajectory into training via a stop-gradient target. On ImageNet-256, single-step Drifting trained from scratch outperforms not only single-step EqFM by a large margin, but even beats multi-step EqFM and EqFM's 500-step teacher. This, to me, seems very important since Drifting does not require a teacher model and thus presents a stronger baseline which is neither discussed nor compared against.

  2. Missing comparison with Equilibrium Matching (Wang and Du, 2025): The paper positions AFM and EqFM as correcting the missing clock weight in EqM and amortizing its sampling. However, EqM itself achieves a better FID on ImageNet-256 than both EqFM's 500-step AFM teacher and EqFM's best multi-step result. Yet it seems that EqM is omitted from all experimental tables and plots.

  3. Selective omissions in Table 3 (AlphaFlow, MeanFlow, and flow-map distillation): The authors note that they took their MeanFlow checkpoint from the AlphaFlow repository (Zhang et al., 2025). However, it seems that they omitted AlphaFlow itself from the experiments, even though AlphaFlow trained from scratch substantially outperforms teacher-distilled EqFM in the 1- and 2-step regimes. Moreover, while MeanFlow is plotted in Figure 2 (where it outperforms EqFM at 1 and 2 NFE), it is omitted from Table 3 alongside standard time-dependent flow-map distillation baselines (Boffi et al., 2025; Align Your Flow, 2025), leaving BTM as the only few-step baseline in the main table.

  4. Undocumented training terms in the ImageNet-256 teacher (Table 4). This table shows that training the ImageNet-256 autonomous field teacher required an active loss labeled "data anchor, 0.5" and a "Uniform-weighted sampler" that are never explained in the main text or Algorithm 1.

Questions for authors

1 Comparisons with Drifting (2026), Equilibrium Matching (2025), AlphaFlow (2025), and flow-map distillation:

  • Why was Generative Modeling via Drifting (Deng et al., Feb 2026)—which also uses an autonomous field vanishing at equilibrium and achieves much stronger 1-step ImageNet-256 generation from scratch—neither cited nor compared against?
  • Why was Equilibrium Matching (Wang and Du, 2025) omitted from the empirical comparisons, and why does EqM achieve a better ImageNet-256 FID than your AFM teacher if EqM is biased by omitting the inverse clock weight?
  • Why was AlphaFlow excluded from the ImageNet-256 comparisons (and MeanFlow omitted from Table 3 despite outperforming EqFM at 1–2 NFE in Figure 2), and how does EqFM compare against standard time-dependent flow-map distillation (Boffi et al., 2025; Align Your Flow, 2025) under the same teacher-student training budget?
  1. What is the exact definition of the "data anchor, 0.5" loss and "Uniform-weighted sampler" in Table 4, and is the anchor term required to prevent drift on the data manifold?
3 comments

We thank the reviewer for the careful reading, and are delighted to hear their appreciation of our conceptual and practical contributions bridging equilibrium models, one-step generative models, and flow-map distillation. We hope to address the reviewer's concerns below:

W1/Q1 Connections to drifting models

We agree that Drifting 1 and EqFM are related and have added a comparison to the revised submission. Both methods support one-step generation through equilibrium. A mathematical connection arises through Poisson constructions, which we detail in section A.3 of the appendix.

We emphasize that a key distinction between Drifting and EqFM lies in the meaning of equilibrium and the resulting generation dynamics. In Drifting, equilibrium is defined at the distribution level: the drifting field attracts generated samples toward real data while repelling them from other generated samples, vanishing when the expected forces cancel as the generated distribution matches the data distribution. In contrast, EqFM defines equilibrium at the individual sample level, where an autonomous velocity field drives each sample toward a data equilibrium at which the velocity vanishes. The AFM current has a source–sink balance between prior and data, but that balance alone does not identify it with Drifting’s generator-dependent update field.

We agree that Drifting is an important baseline and have added its results to our comparison. However, a direct comparison of ImageNet performance is difficult because Drifting uses pretrained feature representations and trains its generator for 1,280 epochs, whereas our EqFM training stage uses 400k steps, approximately 80 epochs, following autonomous field initialization. These substantial differences in training budget and supervision affect how the FID results should be interpreted. We acknowledge Drifting’s stronger one-step performance and will make these differences explicit alongside the comparison.

W2/Q2 Comparisons to EqM

We agree that EqM 2 is an essential baseline. The revised table reports its published ImageNet-256 FID of 1.90 at 500 NFE. The reviewer's concerns surrounding EqM's stronger performance on FID at 500 steps are legitimate, and we point the reviewer to the new self-distilled EqFM results, which beat EqM at 2 NFE, reaching an FID of 1.85. We also include a checkerboard experiment that isolates the effect of our clock-weighted objective: with the architecture, clock, and sampler held fixed, it produces lower transport error than the uniform-time EqM objective. These results help show the practical benefit of our approach, even though we recognize the ImageNet field results are not compute-matched. EqM was trained for 1,400 epochs, whereas the autonomous field in our original submission was trained for 80 epochs. We have clarified this difference so the field comparison is interpreted with the appropriate qualification.

W3/Q3 Comparisons to AlphaFlow

The reviewer is right to point out that AlphaFlow 3 is a relevant comparison, and we appreciate the opportunity to clarify our choice of baseline. Our new EqFM self-distillation results beat AlphaFlow α-Flow-XL/2 results at all inference budgets, and beat their α-Flow-XL/2+, fine-tuned for 60 epochs at batch size 1024, results at 2 NFE. We emphasize to the reviewer that the training budgets and architectures across EqFM and α-Flow are not held constant. We select MeanFlow 4 as the primary baseline because it provides a standard time-dependent formulation against which to present our autonomous flow-map construction. AlphaFlow improves MeanFlow through objective design and a training curriculum that stabilizes optimization. We did not adopt these additional techniques, aiming to keep the training recipe and presentation focused on EqFM’s construction. Incorporating them into EqFM while leaving the MeanFlow baseline unchanged would introduce further differences and make it harder to attribute improvements to our method. Nevertheless, AlphaFlow’s published results remain relevant, and we have added them to the comparison. The improved training techniques as presented in AlphaFlow remain a relevant research direction for Equilibrium Flow Maps.

W4/Q4 Clarifying Terms in Appendix

The reviewer is right that these terms were insufficiently explained. Both refer to the initial autonomous-field training.

For a clean data point x drawn from the data distribution, let bθ(x) be the predicted velocity and d the number of dimensions. The data anchor loss is

L_anchor(θ) = E_{x ∼ p_data}[‖bθ(x)‖₂² / d]

We train with L_total = L_AFM + 0.5 L_anchor. This penalizes nonzero velocity at data points, encouraging them to be equilibria of the learned dynamics.

The uniform-weighted sampler was introduced in the main text as the unbiased objective function AFM. It samples interpolation times uniformly up to a cutoff and weights each regression loss by the inverse clock speed. For our exponential clock, this weight is proportional to 1 / (1 − t) and is normalized to have mean one. We have clarified those aspects in the paper.

References

[1] Deng et al., “Generative Modeling via Drifting”, 2026.

[2] Wang and Du, “Equilibrium Matching: Generative Modeling with Implicit Energy-Based Models”, 2025.

[3] Zhang et al., “AlphaFlow: Understanding and Improving MeanFlow Models”, 2025.

[4] Geng et al., “Mean Flows for One-step Generative Modeling”, 2025.

Reviewer_AmN9OpenReview ↗

Grateful for the comments and additional results. Paper still needs work, in my opinion.

I would like to thank the authors for the additional results, discussions and clarifications. I am grateful in particular for acknowledging the relation to Drifting.

I have also read over the other reviews.

Overall, I feel that the paper makes several interesting and novel contributions. However, the quality of the presentation and especially the relation to prior works and comparison to most closely related baselines still needs to be clarified. Since this is a fundamental issue (also raised by reviewer SkPY for example) I think the paper needs to go over at least a major revision.

In relation to Drifting one thing I would like to highlight is that the authors argue that EqFM uses "400k steps, approximately 80 epochs, following autonomous field initialization." However, in the paper they state that their autonomous field is initialized from the public SiT-XL/2 checkpoint [32]. According to that paper, this checkpoint was itself trained for 7M steps on ImageNet-256.

Comparisons with other baselines such as EqM and AlphaFlow can also be improved. For example, if I understand correctly α-Flow-XL/2+ (which is plotted in Figure 2 of the paper) gets 2.58 (1 NFE) and 2.15 (2 NFE). In contrast, standard EqFM gets (2.90 at 1 NFE, 2.28 at 2 NFE). lt only achieves (1.85 at 2 NFE) via the FID-tuned per-jump schedule.

Due to the broad and serious nature of the concerns I will maintain my weak reject score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.