acceptodds
Under review as a conference paper at ICLR 2027

Recursive Flow Matching

Abstract

Generative models have emerged as a powerful paradigm for solving physics systems and modeling complex spatiotemporal dynamics. However, achieving high physical accuracy without incurring high computational cost remains a fundamental challenge, as existing approaches face a critical speed-fidelity trade-off. In this work, we introduce **Recursive Flow Matching** (*RecFM*), a generative framework for forecasting complex spatiotemporal dynamics. RecFM enforces self-consistency to align trajectories across discretization scales, reducing discretization errors and improving performance across metrics for physics-based tasks. To our knowledge, this is the first method to achieve high-fidelity one- and few-step (*2-4 step*) dynamic generation for scientific systems, matching state-of-the-art multi-step solvers without relying on physics-based residuals. Across challenging scientific benchmarks, RecFM achieves up to a 20× speedup over leading diffusion-based emulators while improving predictive accuracy. Furthermore, RecFM reduces mean squared error by over 15% compared to vanilla flow matching, offering a scalable and efficient solution for real-time scientific emulation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Flow Matching, Self Consistency, Dynamic Forecasting

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_Em6HOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a new few-step method for PDE solving, which enforces self-consistency across trajectory. Across three challenging scientific benchmarks, it achieves excellent predictive accuracy while reduces inference time.

Critical strengths
  1. The paper is well written and easy to follow.
  2. The experiments are thorough, including various evaluations, runtime comparisons, and ablation studies.
Critical weaknesses
  1. I have concerns about the novelty, as it appears to be primarily an application of consistency models to PDE solving.
  2. The paper also seems to overstate its contributions. To my knowledge, it is not the first work to propose a few-step or one-step generative approach to solving PDEs, and comparisons with these prior methods should be included.
Questions for authors
  1. Line 22 "scalable", what does “scalable” mean in this context?
  2. How does it perform compared to other few-step generative methods? For example,

[1] Physics-informed distillation of diffusion models for pde-constrained generation.

[2] Rectified flows for fast multiscale fluid flow modeling.

  1. I have some doubts about the need for few-step generative methods, since operator-based methods, such as neural operators, already provide fast, one-step predictions. How does your method compare with neural operators in inference time, and what advantages does it offer over them?

I would consider raising my score if my concerns are addressed.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Em6H

We thank the reviewer for the clear and constructive feedback, and address each point below.

"I have concerns about the novelty, as it appears to be primarily an application of consistency models to PDE solving."

We appreciate the concern, but RecFM uses a different mechanism from consistency models (General Response §1). Consistency models learn a map from xtx_t to the trajectory endpoint and enforce the semigroup property. This requires evaluating the network at different points along the trajectory, usually against a teacher or EMA target. RecFM learns a velocity field, supervises it on time-rescaled copies of each trajectory, and ties them together with a chain-rule constraint at a single point, without a teacher, EMA target or JVP. The paper also compares against two consistency-style methods on the same backbone, Shortcut (Table 1) and MeanFlow (Appendix J, originally Appendix I during initial submission), and RecFM outperforms both in the one-step setting.

"The paper also seems to overstate its contributions. To my knowledge, it is not the first work to propose a few-step or one-step generative approach to solving PDEs, and comparisons with these prior methods should be included."

We thank the reviewer for the pointers, and agree that few-step generative PDE methods exist. Our claim refers to a narrower setting: one- and few-step autoregressive spatiotemporal rollouts without a teacher model or PDE residuals. [1] requires a teacher and PDE residuals, [2] learns a single input-to-output map with several ODE steps, and PBFM (Baldan et al., 2026) relies on PDE residuals. We have added these works to related work, and comparisons are given below and in General Response §3.

"Line 22 "scalable", what does "scalable" mean in this context?"

We meant two things. Each forecast window costs only 1 to 2 network evaluations, so rollout cost grows linearly with the horizon. And the method needs no architectural changes, so it carries over to large backbones (a 116M-parameter HV-DiT here, and a 675M-parameter model on ImageNet in Appendix K (originally Appendix J)). These are the two senses we intended.

"How does it perform compared to other few-step generative methods? For example, [1] Physics-informed distillation of diffusion models for pde-constrained generation. [2] Rectified flows for fast multiscale fluid flow modeling."

[2] (Armegioiu et al.) trains a conditional rectified flow, a linear interpolant from noise to target conditioned on the input, and samples it in a few steps. This is the same objective as our Vanilla FM baseline. As we could not find a public implementation of [2], we also implemented the stronger two-stage reflow variant on our backbone for all three datasets (now in Table 1; see General Response §3). One-step RecFM achieves lower MSE on every dataset: 0.162 vs. 0.225 (SST), 0.0064 vs. 0.0072 (Navier-Stokes) and 4.2e-5 vs. 4.6e-5 (Helmholtz). Reflow also uses twice the training compute.

[1] (PIDDM) distills a pretrained diffusion teacher into a one-step student using an explicit PDE-residual loss, and generates full solution fields in one shot (e.g., Darcy, Poisson, Burgers) rather than autoregressive rollouts. Because it requires both a teacher and a closed-form residual, it cannot be applied to SST. The closest method we could evaluate is PBFM, which also trains with PDE residuals (Appendix G, originally Appendix F). RecFM is more accurate than PBFM on both PDE datasets while using 1 to 2 sampling steps instead of 20. We have added a discussion of PIDDM to related work.

"I have some doubts about the need for few-step generative methods, since operator-based methods, such as neural operators, already provide fast, one-step predictions. How does your method compare with neural operators in inference time, and what advantages does it offer over them?"

Thank you for raising this. We added FNO, PINO and DeepONet; the table below is also a new appendix section (Appendix D) in the revised PDF.

MethodSST MSESST Time [s]NS MSEHelmholtz MSE
FNO0.2060.41600.03372.9e-5
PINO--0.01542.8e-5
DeepONet0.2130.40750.03523.9e-5
RecFM (1-step)0.1620.43100.00644.2e-5
RecFM (2-step)0.1610.73530.00682.7e-5

Inference time is comparable, and RecFM is clearly more accurate on SST and Navier-Stokes. On Helmholtz, FNO and PINO are slightly more accurate than one-step RecFM, and two-step RecFM is the most accurate overall. Beyond accuracy, RecFM produces calibrated ensembles (CRPS, SSR) where neural operators give a single point estimate. It also needs no governing equation, which is why PINO has no SST entry.

We hope our responses have addressed the reviewer’s concerns. If there are any remaining questions, we would be happy to provide further clarification during the discussion period. We would sincerely appreciate reconsideration of the overall rating.

Thank you for your rebuttal, which addressed part of my concerns. I decided to raise my score to 3. I suggest the author removing claims such as "the first work to propose a few-step or one-step generative method".

Thank you for your feedback and for reconsidering your score. We will revise the manuscript accordingly.

Reviewer_sfs4OpenReview ↗
Rating
4(+2)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a new few step generation framework, that enforces self-consistency to align trajectories, for forcasting complex spatiotemporal dynamics. Experimengs are conducted to show the effectiveness.

Critical strengths

The proposed approach starts from physics intuition and establishes on scale consistency.

Critical weaknesses

Several points need to clarify

  1. How scale consisteny relate to self-consistency decsribed by the semigroup condition

  2. When \alpha goes to 0, does scale consistent reduce to mean flow [1]?

  3. How scale consistency relate to SoFlow [2] and TCF [3]? The trajectory term and consistency terms are very similar.

  4. .In the experiments, you have the statements:

In particular, it attains up to a 20× speedup over the diffusion-based baseline VideoPDE while also improving predictive accuracy and calibration.

Moreover, compared to vanilla flow-matching methods, which typically require ∼5 inference steps, RecFM produces high-quality results with only 1-2 steps, achieving over 15% lower MSE and substantially better SSR scores.

These two statements seem to be contraditory.

  1. Experiments should include Mean Flow and SoFlow as baselines.

References:

1 Mean Flows for One-step Generative Modeling, https://openreview.net/forum?id=uWj4s7rMnR

  1. SoFlow: Solution Flow Models for One-Step Generative Modeling, https://openreview.net/forum?id=cjb03GNqYw

  2. Trajectory-Consistent Flows: Enabling Fast Sampling for Flow Matching Models, https://openreview.net/forum?id=VFYGBTSNWK

Questions for authors

I don't understand how Figure (2) is related to Figure 1(b)

4 comments
AuthorsOpenReview ↗

Response to Reviewer sfs4

We thank the reviewer for the precise questions. General Response §1 covers what they have in common, and short answers follow below.

"How scale consisteny relate to self-consistency decsribed by the semigroup condition"

The two conditions constrain different objects. The semigroup condition Xu,t(Xs,u(x))=Xs,t(x)X_{u,t}(X_{s,u}(x)) = X_{s,t}(x) concerns the flow map and composes transports through intermediate points. Our condition, v(xt,t/α,α)=αv(xt,t,1)v(x_t, t/\alpha, \alpha) = \alpha v(x_t, t, 1), relates instantaneous velocities at a single point: it is the chain rule for t=ατt = \alpha\tau. It neither implies the semigroup property nor requires it, since a flow-matching field already defines a valid flow map.

"When α\alpha goes to 0, does scale consistent reduce to mean flow?"

No. As α→0\alpha \to 0, the secondary trajectory covers only the segment t∈[0,α]t \in [0, \alpha] of the primary path, which shrinks to the data endpoint. Its velocity αv(x,ατ,1)\alpha v(x, \alpha\tau, 1) goes to zero, both sides of the relation are O(α)O(\alpha), and pα→p0p_\alpha \to p_0. MeanFlow learns the average velocity over a finite interval [r,t][r, t] through the MeanFlow identity, which requires a JVP, and its one-step output evaluates a flow map. The two coincide only in the trivial limit (r→tr \to t for MeanFlow, α=1\alpha = 1 for RecFM), where both reduce to standard flow matching.

"How scale consistency relate to SoFlow and TCF? The trajectory term and consistency terms are very similar."

The reviewer is right that all three methods pair a flow-matching term with a consistency term; many few-step methods share this template. They differ in what the consistency term relates:

Network outputConsistency term comparesOne-step target
SoFlowsolution map fθ(xt,t,s)f_\theta(x_t, t, s)fθ(xt,t,s)f_\theta(x_t, t, s) vs. stop-grad fθf_\theta at the displaced point xt+v(l−t)x_t + v(l - t)a sample
TCFTaylor trajectory map xτ+(t−τ)vθ+12(t−τ)2uθx_\tau + (t-\tau) v_\theta + \frac{1}{2}(t-\tau)^2 u_\thetafθ(xτ,t2)f_\theta(x_\tau, t_2) vs. stop-grad fθ(fθ(xτ,t1),t2)f_\theta(f_\theta(x_\tau, t_1), t_2) (semigroup)a sample
RecFMvelocity vθ(x,τ,α)v_\theta(x, \tau, \alpha)vθ(xt,t/α,α)v_\theta(x_t, t/\alpha, \alpha) vs. αvθ(xt,t,1)\alpha v_\theta(x_t, t, 1), same xtx_tconditional mean

The trajectory terms also differ: TCF trains its velocity head to ignore the target time tt, while RecFM asks for a scaled output α(x1−x0)\alpha(x_1 - x_0) under a joint change of (τ,α)(\tau, \alpha). We have added both papers to related work.

"In particular, it attains up to a 20× speedup over the diffusion-based baseline VideoPDE ... compared to vanilla flow-matching methods ... achieving over 15% lower MSE ... These two statements seem to be contradictory."

The two statements compare against different baselines. The first compares speed against VideoPDE, a many-step diffusion model. The second compares accuracy against vanilla FM on our backbone, which needs about 5 steps where RecFM uses 1 to 2. The 20× figure is in fact conservative. Table 1 (SST) gives 19.75 s vs. 0.43 s / 0.74 s, about 46× / 27×, and Table 6 (Navier-Stokes; index has been updated from 5 during initial submission) gives 72.6 s vs. 1.59 s / 3.13 s, about 46× / 23×.

"Experiments should include Mean Flow and SoFlow as baselines."

Thank you for the suggestion. MeanFlow is already part of our evaluation. Appendix J (originally Appendix I during initial submission) compares it with RecFM on Navier-Stokes, using the same backbone and evaluation protocol. One-step RecFM is better on every metric: CRPS 0.031 vs. 0.042, MSE 0.0064 vs. 0.0093, and SSR 0.959 vs. 0.843.

MeanFlow needed considerable tuning before it trained reliably on autoregressive dynamics, which matches recent findings. The original MeanFlow target depends on the network itself, and its loss "has a much higher variance and is non-decreasing" (Improved Mean Flows, Geng et al., 2025). A recent MeanFlow forecaster for turbulent flows adds window-level consistency, a stop-gradient reformulation and a time-increment loss to make it work autoregressively (Yang & Xue, 2026).

We have also reimplemented SoFlow on Navier-Stokes with the same backbone and evaluation protocol: one-step RecFM is again better on every metric (CRPS 0.031 vs. 0.055, MSE 0.0064 vs. 0.0143, SSR 0.959 vs. 0.877). Both comparisons are in Appendix J of the revised PDF.

"I don't understand how Figure (2) is related to Figure 1(b)"

Figure 2 is the physical analogy. Each bounce multiplies the pendulum's velocity by α\alpha, which produces trajectories with velocities αi−1v∗\alpha^{i-1} v^\ast. Figure 1(b) applies the same rule in data space. Trajectory ii moves with velocity α(i)(x1−x0)\alpha^{(i)}(x_1 - x_0), and aligning time as τ=t/α(i)\tau = t/\alpha^{(i)} makes all trajectories pass through the same xtx_t, where they are compared.

We hope these responses address your concerns. If the reviewer finds these clarifications and additional experiments helpful, we would sincerely appreciate reconsideration of the overall rating.

AuthorsOpenReview ↗

Follow-up before the discussion period ends

Dear Reviewer,

Thank you again for your time. As the discussion period is about to end, we kindly request that you let us know whether our clarifications addressed your concerns. In brief:

  • Relation to the semigroup condition, MeanFlow, SoFlow and TCF: RecFM's consistency term relates velocities at the same point through the chain rule, while these methods constrain the flow map or the average velocity. Our reply compares them side by side in a table.
  • The two speedup statements: they compare against different baselines (speed against VideoPDE, accuracy against vanilla FM), so they do not conflict.
  • MeanFlow and SoFlow baselines: both are in Appendix J with the same backbone and protocol. One-step RecFM is better than both on every metric (MSE 0.0064 vs. 0.0093 for MeanFlow and 0.0143 for SoFlow).

Please let us know if any concerns or questions remain; we would be happy to address them.

Thank you very much for your explanation. I'll raise the rating.

Thank you for taking the time to read our responses and for raising your score. We appreciate your constructive feedback, which has helped us improve the paper. We will incorporate the clarifications from our discussion into the later revision.

Reviewer_NFa2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Recursive Flow Matching trains a shared velocity network across multiple time scales, with an additional consistency loss linking their predictions. Sampling uses the primary scale with one or a few Euler steps. The paper reports improved forecasting on sea surface temperature, Navier–Stokes, and Helmholtz benchmarks, alongside additional generation and ablation studies.

Critical strengths
  1. The training scheme is straightforward and achieves promising few-step forecasting results without requiring physics residuals in the loss.
  2. The evaluation includes predictive accuracy, probabilistic scores, physical metrics, runtime, and several ablations. This provides useful evidence that the training changes have practical benefits.
Critical weaknesses
  1. The theoretical explanation contains a central contradiction. Corollary B.4 claims that scale consistency reduces trajectory curvature and makes a single Euler step distributionally correct. Remark B.3 and Appendix H explicitly state that the same consistency relation can hold for any primary velocity field, regardless of curvature. Consistency of the continuous flow therefore does not establish the claimed one-step guarantee.
  2. The interpretation of one-step probabilistic forecasting remains unclear. Appendix H states that the population-optimal one-step output is the conditional mean. The paper reports CRPS and ensemble spread, but does not adequately explain how meaningful sampling diversity arises under this analysis. This matters for the claimed uncertainty-modeling benefit.
Questions for authors
  1. Can the authors reconcile Corollary B.4 with Remark B.3 and Appendix H, and identify which few-step guarantees actually follow from the objective?
  2. What produces the reported ensemble spread for a fixed condition in the one-step setting? How does one-step RecFM compare with direct regression using the same backbone and training budget?
2 comments
AuthorsOpenReview ↗

Response to Reviewer NFa2

We thank the reviewer for the careful reading of Appendices B and H (now I). Both points are well taken, and we address them below.

"Corollary B.4 claims that scale consistency reduces trajectory curvature and makes a single Euler step distributionally correct. Remark B.3 and Appendix H explicitly state that the same consistency relation can hold for any primary velocity field, regardless of curvature." (W1, Q1)

We thank the reviewer for catching this. We have corrected the corollary so that it states what follows from the objective, which also answers Q1:

  • (i) At the optimum, one Euler step of size hh on the secondary trajectory equals one Euler step of size αh\alpha h on the primary trajectory, x−hvθ∗(x,τ,α)=x−αhvθ∗(x,ατ,1)x - h v_{\theta^\ast}(x, \tau, \alpha) = x - \alpha h v_{\theta^\ast}(x, \alpha\tau, 1). KK-step sampling of every secondary trajectory therefore coincides with KK-step sampling of the primary trajectory on [0,α][0, \alpha], at any discretization, and as K→∞K \to \infty both recover the exact flow between pαp_\alpha and p0p_0 (Thm. B.2).
  • (ii) The secondary field has acceleration α2a(x,ατ)\alpha^2 a(x, \alpha\tau), so Lemma 2.1 gives its KK-step error explicitly. At α=1\alpha = 1 this equals the flow-matching bound: at the population level, Lcons\mathcal{L}_{\text{cons}} adds no curvature constraint (Remark B.3), and the few-step gain is a property of the learned field.

Table 10 measures that property with the quantities in Lemma 2.1. On all three datasets, RecFM has lower mean and 95th-percentile ∥a∥\Vert a \Vert and a smaller KK-step gap to a 256-step reference. The method and experiments are unchanged.

"Appendix H states that the population-optimal one-step output is the conditional mean. The paper reports CRPS and ensemble spread, but does not adequately explain how meaningful sampling diversity arises under this analysis." (W2, Q2)

At the optimum, one Euler step from t=1t = 1 returns x1−v∗(x1,1)=E[x0∣c]x_1 - v^\ast(x_1, 1) = \mathbb{E}[x_0 \mid c] for every x1x_1. For Navier-Stokes and Helmholtz this is the appropriate target, since the conditioning frames essentially determine the next window, and CRPS, as a proper score, then rewards accuracy. The spread of a trained one-step model reflects the learned field's remaining dependence on x1x_1.

For a stochastic system such as SST, sample diversity comes from multi-step sampling: from the second step on, the velocity is evaluated at a state that depends on x1x_1. Table 1 reflects this on SST: going from one to two steps raises the SSR from 0.984 to 1.004 at essentially the same MSE (0.162 vs. 0.161), so the added spread is not coming from larger errors. Accordingly, the uncertainty-modeling benefit we claim is empirical: the ensemble spread is calibrated against the actual error (SSR close to 1 in Table 1), and on stochastic data it grows with multi-step sampling. It is not a property of the one-step population optimum.

On the comparison with direct regression, we trained the same HV-DiT backbone with an MSE objective for the same number of epochs as RecFM. One-step RecFM is more accurate at similar runtime: MSE 0.162 vs. 0.269 (SST), 0.0064 vs. 0.0181 (Navier-Stokes) and 4.2e-5 vs. 1.5e-3 (Helmholtz), and CRPS 0.217 vs. 0.303, 0.031 vs. 0.061 and 0.0034 vs. 0.038. Both target E[x0∣c]\mathbb{E}[x_0 \mid c] at the population level, so the gap comes from training, not from the target. The full comparison is in General Response §3.

We hope our responses have addressed the reviewer's concerns. If there are any remaining questions, we would be happy to provide further clarification during the discussion period. If the reviewer finds these clarifications and additional experiments helpful, we would sincerely appreciate reconsideration of the overall rating.

AuthorsOpenReview ↗

Follow-up before the discussion period ends

Dear Reviewer,

Thank you again for your time. As the discussion period is about to end, we kindly request that you let us know whether our clarifications addressed your concerns. In brief:

  • Corollary B.4 and Remark B.3: we clarified the statement of Corollary B.4 in Appendix B. It gives the exact Euler-step equivalence between secondary and primary sampling and the resulting error bound, and the few-step gain is a property of the learned field, which Table 10 measures directly: the learned field is straighter than vanilla FM's on all three datasets.
  • Ensemble spread in one-step sampling: our reply explains where the spread comes from, and on SST the spread grows with a second step at essentially the same MSE (SSR 0.984 to 1.004).
  • Direct regression with the same backbone and budget: now in Table 1. One-step RecFM is more accurate on all three datasets, for example MSE 0.0064 vs. 0.0181 on Navier-Stokes and 4.2e-5 vs. 1.5e-3 on Helmholtz.

Please let us know if any concerns or questions remain; we would be happy to address them.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank all reviewers for their thoughtful and detailed feedback. Three topics came up in more than one review, so we address them here once and refer to this comment in the individual replies. Please also note that table and appendix indices have been updated in the new PDF file.

1. What the RecFM consistency term is

Fix a scale α\alpha. The secondary trajectory is the primary trajectory on t∈[0,α]t \in [0, \alpha], replayed over τ∈[0,1]\tau \in [0, 1] with t=ατt = \alpha\tau (Theorem B.2). The chain rule then fixes its velocity, dxdτ=αv(x,ατ,1)\frac{dx}{d\tau} = \alpha v(x, \alpha\tau, 1), and our consistency loss

Lcons=∥vθ(xt,t/α,α)−αvθ(xt,t,1)∥2\mathcal{L}_{\text{cons}} = \left\Vert v_\theta(x_t, t/\alpha, \alpha) - \alpha v_\theta(x_t, t, 1) \right\Vert^2

enforces exactly this relation at a single point xtx_t. It compares instantaneous velocities. It never composes steps or queries the network at a step-forward state, and it uses no teacher, EMA or stop-gradient target, and no JVP.

Consistency models, Shortcut, MeanFlow, SoFlow and TCF instead constrain the flow map (or the average velocity) along a trajectory. That is also why they behave differently in one step: those methods target a sample of p0p_0, while RecFM targets the conditional mean (see our reply to NFa2).

2. What the theory guarantees

We thank Reviewer NFa2 for the discussion of Corollary B.4, and we have updated it (see our reply to NFa2). The theory establishes three things. At α=1\alpha = 1, RecFM has the same population optimum as flow matching (Prop. B.1), so the target is unchanged. Every secondary trajectory is a valid transport between the right marginals (Thm. B.2). And Lcons\mathcal{L}_{\text{cons}} is a reparameterization constraint (Remark B.3).

The few-step gains come from what a finite network learns under this objective. Each xtx_t receives supervision on the same direction x1−x0x_1 - x_0 under many (τ,α)(\tau, \alpha) conditionings, and shared parameters tie the α=1\alpha = 1 output to all of them. Table 10 in the updated PDF file (original Table 9 during initial submission) measures the result directly. On all three datasets, the learned α=1\alpha = 1 field is straighter than vanilla FM's: lower straightness error SS, lower mean and 95th-percentile acceleration, and smaller 1- and 2-step gaps to a 256-step reference.

3. Updated Table 1

We added two baselines to Table 1, both on our HV-DiT backbone: Rectified Flow (two-stage reflow) and direct regression with the same backbone and training budget. The updated table is below and is also in the revised PDF. Neural operators (Em6H) and MeanFlow/SoFlow (sfs4) are in the appendix; see those replies.

MethodSST CRPSSST MSESST SSRSST Time [s]NS CRPSNS MSENS SSRHelm. CRPSHelm. MSEHelm. SSR
Perturbation*0.2810.1800.4110.42410.0900.0280.4480.2180.1110.004
Dropout*0.2670.1640.4060.42410.0780.0270.7150.0990.0490.631
DDPM*0.2460.1770.6740.30540.1800.1050.5730.1560.1530.563
MCVD*0.2160.1610.92679.1670.1540.0700.5240.1370.1280.867
DYffusion*0.2240.1731.0334.67220.0670.0220.8770.1440.1061.121
VideoPDE0.2160.1620.74619.7530.0330.00680.2050.0265.6e-44.334
Vanilla FM0.2600.2320.9141.52020.0360.00760.9110.0306.5e-41.485
HV-DiT regression (new)0.3030.269-0.41210.0610.0181-0.0381.5e-3-
ReFlow, 1-step (new)0.2560.2251.0980.43730.0330.00721.3670.00404.6e-51.296
ReFlow, 2-step (new)0.2550.2251.1060.73910.0330.00711.3020.00344.3e-51.559
Shortcut0.2490.2970.7980.44520.0340.00790.7170.01441.6e-40.467
RecFM (1-step)0.2170.1620.9840.43100.0310.00640.9590.00344.2e-51.090
RecFM (2-step)0.2160.1611.0040.73530.0320.00680.9320.00272.7e-51.440

*Reproduced from DYffusion. Regression is deterministic, so its CRPS reduces to the MAE and its SSR is 0 by construction (no ensemble spread); the revised PDF states this in Sec. 5.2.3. Rectified Flow trains stage 1 for as many epochs as RecFM and stage 2 on the stage-1 couplings for the same number again, so it gets at least twice RecFM's training compute.

Dear Reviewers,

Please review the authors' rebuttal and participate in the discussion to address any remaining questions before the period closes on Nov 18 AoE.

Thank you!

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.