acceptodds
Under review as a conference paper at ICLR 2027

CAT-Flow: Curvature-Adaptive sTeps for Flow Matching

17 views

Abstract

Flow matching has solidified its position as a leading framework for generative modeling, powering state-of-the-art systems such as FLUX and Stable Diffusion 3.5. However, the iterative nature of its ODE-based sampling process creates a fundamental efficiency bottleneck: the quality of generated samples is highly sensitive to the choice of step-sizes, and current models typically require 20 to 30 steps to reach satisfactory quality. To this end, we propose two lightweight, training-free, and efficient algorithms, CAT-OV and CAT-OT, which adapt step-sizes at inference time. This is made possible by establishing a novel and provable connection between flow matching sampling and gradient flow. Without requiring additional neural function evaluations, CAT-OT estimates curvature over time via a finite-difference approximation of the time-derivative of the vector field, while CAT-OV estimates curvature as a variance of past velocities via a gradient vector field. Under suitable conditions, both methods have truncation error bounds of constant order. Empirically, across four text-to-image flow matching models, CAT-OV and CAT-OT match the image quality of existing step-size heuristics with up to 40% fewer generation steps and negligible added wall-clock time, with the largest gains in metric scores in the low-step regime. We also validate our methods on audio generation, showing efficiency gains that generalize beyond image generation, and demonstrate compatibility with higher-order solvers. Beyond adapting step-sizes, our gradient flow connection opens a broader avenue for transferring optimization techniques to flow matching inference, with proof-of-concept demonstrations including momentum and gradient normalization extensions.

Primary area:
generative models
Keywords:
generative models, flow matching, machine learning, efficient generation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_MDp5OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

CAT-OT uses finite differences of previously evaluated velocities, while CAT-OV uses exponential moving averages of the first and second moments of the scaled velocity. Neither requires additional neural function evaluations (NFEs). The methods are motivated by a connection between conditional straight-line trajectories and gradient flow.

Critical strengths

The proposed methods are simple, lightweight, and require no additional training. The experiment shows performace gain.

Critical weaknesses
  1. CAT-OV does not consistently improve CLIP/AES on Krea, most Schnell differences are small, and CAT-OT can underperform on Krea. In appendix, The O(λ2) needs to be clearly identified as a local-error statement: the displayed global bound scales linearly with λ for fixed γ.

  2. the proof uses the unclipped step rule and does not directly cover the implemented fixed minimum step size. More importantly, the strictly positive lower-bound assumption in Eq. (13) cannot hold as written. Since Ak contains all positive convex combinations, a weight can approach one, making the variance arbitrarily small even for fixed, distinct, finite observations. The infimum is therefore zero. Identical support points, also permitted by the definition, give zero variance directly. The analysis needs a different condition, potentially restricted to the actual EMA trajectory.

Questions for authors
  1. How should Eq. (12) account for unnormalized EMA weights, and how can the infimum in Eq. (13) be strictly positive? What corrected guarantees apply to the clipped algorithms?

  2. What conditions support Eq. (8) for the marginal sampling field? Does removing (1−t) preserve the benefits, and how does the CAT-OV statistic relate to actual local error or acceleration?

2 comments

We are grateful to the reviewer for their positive feedback and recommendation on our work. We now respond to the reviewer’s detailed questions below.

  • How should Eq. (12) account for unnormalized EMA weights,

Eq. (12) handles the unnormalized EMA weights by treating the initial value v0=g0=0v_0 = g_0 = 0 as a virtual sample, which carries the remaining weight βk\beta^k, so that all weights sum to one. In addition, the norm of the EMA variance along the actual sampling trajectory is at least above 15 in our experiments with FLUX-1-dev and SD-3.5-large. A zero variance would indicate a flat path, which then would mean one could, in theory, take a large step to produce samples since there is no truncation error.

  • and how can the infimum in Eq. (13) be strictly positive?

Eq. (13) is an assumption, not a derived fact, and we need it only for the states actually visited by the sampler. We find this assumption reasonable for two reasons. First, pretrained models do not have purely straight flows and using CAT-OV with larger step-sizes inexorably introduces more variance along the discretized path (Appendix N, Figure 7). Second, as mentioned earlier, we empirically found that the sampled trajectories for FLUX-1-dev and SD-3.5-large never present a norm of the EMA variance below a positive value of 15.

  • and how does the CAT-OV statistic relate to actual local error or acceleration?

In short, more change of velocity in recent history leads to a larger CAT-OV statistic, which then leads to smaller step-sizes: this leads to smaller local error but a larger amount of total steps during sampling, i.e., a slower sampling. Below we provide a more detailed explanation.

The CAT-OV statistic is a proxy for the change of velocity X¨t=(d/dt)u(Xt,t;θ)\ddot{X}_t=(d/dt) u(X_t, t; \theta) across the discretization path, not a direct measurement of it. For a perfectly straight flow the velocity is constant and so X¨=0\ddot{X} = 0, thus, any change of velocity along the path indicates curvature. Since the CAT-OV statistic measures the discounted spread of the past values of (1−t)u(1 − t)u, it is large when the velocity has recently been changing, which is what a large X¨t\ddot{X}_t produces (u(Xt2,t2;θ)−u(Xt1,t1;θ)=∫t1t2X¨(τ)dτu(X_{t_2},{t_2};\theta) − u(X_{t_1},t_1;\theta) = \int_{t_1}^{t_2}\ddot{X}(\tau)d\tau for times t2<t1t_2<t_1). Now, the Euler local error over a step-size Δt\Delta_t is bounded by Δt2Mk/2\Delta^2_t M_k/2, where MkM_k bounds ∥X¨t∥\|\ddot{X}_t\| (Appendix E). Therefore, a large CAT-OV statistic gives both a smaller step-size Δt\Delta_t (Eq. (6)) and a smaller error bound; and a small statistic allows a larger step. The CAT-OV statistic controls this bound under the assumption in Eq. (13).

  • What corrected guarantees apply to the clipped algorithms?

We empirically observed that the clipping is seldom activated across the sampling path (i.e., for most of the sampling path we have that Δt∈(Δmin,1−t)\Delta t\in (\Delta_{min},1-t) in Algorithms 1 and 2), with the only exceptions observed on the final step of the adaptive algorithm, so that the total integration time does not go beyond 11.

  • What conditions support Eq. (8) for the marginal sampling field?

The learned vector field is the conditional vector field E[X1−X0∣Xt]E[X_1-X_0|X_t]. If we assume concentration of measure as a condition, e.g., by assuming that given any Xt=xX_t=x there is only one learned path X1−X0X_1-X_0 so that x=tX1+(1−t)X0x=tX_1+(1-t)X_0, then the conditional distribution is well-represented by this representative path (there is one large mode around this representative path). This would effectively allow the conditional sampling field to be well-approximated by the marginal sampling field corresponding to Xt=xX_t=x at time tt. This is an extension of our summarized explanation in lines 226-229, and we will include it in the paper.

  • Does removing (1−t) preserve the benefits,

Some of our experiments while writing the paper showed that removing (1−t)(1-t) leads to a small loss of performance. We did not include them in the paper since removing (1−t)(1-t) does not correspond to our theoretical motivation from equation (8), and so to the formulation of our methods.

AuthorsOpenReview ↗

Additional response

We would like to respond to the following observation raised as a “critical weakness”:

CAT-OV does not consistently improve CLIP/AES on Krea, most Schnell differences are small, and CAT-OT can underperform on Krea.

We would like to clarify that, as stated in lines 342-343 from our paper, our main focus is on efficiency performance rather than performance on absolute score values. We understand the quoted observation is only about performance on score values, and so we will focus our following discussion on that.

First, we highlight that CAT-OV shows remarkable outperformance consistency under the CLIP/AES metrics on FLUX-1-dev, SD-3.5-large and FLUX-1-Schnell (with smaller differences on the last one, as pointed out by the reviewer), which showcases the overall outperformance of CAT-OV. Moreover, there is still noticeable improvement on FLUX-1-Krea, though less consistent as pointed out by the reviewer. We also point out that a great strength of our work is that the outperformance of CAT-OV is consistent under the HPSv3 metric across all models.

Having shown the good performance of CAT-OV across models and metrics, we now explain some caveats when it comes to using CLIP/AES for image quality assessment.

CLIP embeddings majorly capture high-level semantics instead of visual features [1,2]. Thus, both CLIP and AES metrics, which are based on CLIP embeddings, are not strictly evaluating image quality. For this reason, we included the HPSv3 metric and gave more emphasis to it. Moreover, [3,4] argue that the semantic content of a generated image is established very early in the sampling trajectory, which means that semantic improvement will be small unless external changes (e.g., guidance) are made during sampling. Consequently, our methods, which only focus on sampling efficiency and image quality improvement as reflected by HPSv3, are not guaranteed to generate large improvements on CLIP and AES metrics. Thus, we do not see this as a “weakness”, but rather as a consequence of CLIP/AES not strictly measuring image quality improvement. Nonetheless, as we mentioned, CAT-OV still has noticeable gains in the CLIP/AES metrics anyways.

Regarding the observation about FLUX-1-Schnell, this is a model specially tuned to be a “few-step model” (produce images of higher quality in 1-4 steps). Thus, it is natural that our methods, which seek to reduce the number of steps needed for general flow matching models to generate high quality images, will have a smaller quality improvement on a model which is already optimized to be “few-step”. Remarkably, CAT-OV still shows better performance under this model anyway.

The underperformance of CAT-OT for FLUX-1-Krea may be due to the sensitivity to the classifier-free guidance scale, as explained in Appendix J.2. However and most importantly, CAT-OV does not have such limitations.


References:

Reviewer_JKt4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes CAT-OV and CAT-OT, two lightweight, training-free adaptive step-size algorithms for flow matching sampling. The key insight is a provable connection between the flow matching ODE and gradient flow, which justifies using curvature estimates to allocate step sizes, smaller steps in highly curved regions and larger steps in flatter regions. Both methods require no additional neural function evaluations and achieve comparable image quality with up to 40% fewer steps across FLUX, SD3.5, and other flow matching models, with negligible wall-clock overhead.

Critical strengths
  1. Well-motivated and principled. The proven connection between flow matching and gradient flow (Proposition 1) provides a clean theoretical basis for importing adaptive step-size techniques from optimization, making the overall approach well-grounded rather than purely heuristic.
  2. Highly practical with consistent gains. Both methods are training-free, require no extra NFEs, and add negligible overhead, yet consistently achieve comparable quality with up to 40% fewer steps across four models and two modalities (image and audio).
Critical weaknesses

None

Questions for authors
  1. The evaluation relies exclusively on non-reference metrics (CLIP, AES, HPSv3). While the absence of a single ground-truth reference is understandable given that different step-size strategies produce different images, could the authors justify this choice explicitly and discuss whether reference-based metrics (e.g., LPIPS against a high-step baseline) could provide complementary insight into sampling fidelity?
  2. The method is validated on four text-to-image flow matching models and one audio model. Given the claim of generalization beyond image generation, could the authors provide experiments on video flow matching models (e.g., Wan) to more convincingly demonstrate the breadth of applicability, or explicitly clarify the scope of model architectures and modalities where the method is expected to be effective? I would raise my score if my questions are addressed
3 comments

We are grateful to the reviewer for their feedback, and we are glad that no “critical weaknesses” were identified. We appreciate the reviewer’s openness to raising their score: we are now addressing the reviewer's questions.

    1. The evaluation relies exclusively on non-reference metrics (CLIP, AES, HPSv3). While the absence of a single ground-truth reference is understandable given that different step-size strategies produce different images, could the authors justify this choice explicitly and discuss whether reference-based metrics (e.g., LPIPS against a high-step baseline) could provide complementary insight into sampling fidelity?

We recall that our paper already presented a result with a reference-based metric in Appendix H: we have FID against a high-step baseline. Now, in addition, we have performed experiments using two other reference-based metrics: LPIPS and DreamSim, against a high-step baseline; we added DreamSim since it is a more recent metric. With both metrics, the lower the score the better. Here are the results for three flow matching models:

LPIPSDreamSim
NFEs121620121620
FLUX-1-dev
CAT-OV0.34010.28190.24760.09590.07290.0623
CAT-OT0.28040.22890.20290.06680.05570.04940
Dynamic0.43730.36990.31310.14310.10830.0874
Fixed0.44510.38110.33270.15050.1180.095
FLUX-Krea
CAT-OV0.28530.24190.21940.07740.06210.0558
CAT-OT0.30800.23210.21020.07190.05750.0548
Dynamic0.41770.35190.29270.14440.11230.0849
Fixed0.42060.35690.30040.14640.11970.0911
SD-3.5-large
CAT-OV0.49390.44570.42090.14690.11920.1076
CAT-OT0.47820.44740.42430.11340.10230.0964
Dynamic0.6630.59310.54350.25740.20010.1733
Fixed0.67750.60370.53720.27560.21070.1675

As the table shows, our methods outperform both baselines at every step count, by a wide margin, with reductions of roughly 20–55% in LPIPS and DreamSim at 12 steps. The efficiency gain is also clear: with 16 steps, our methods outperform the baselines at 20 steps in all cases, and can achieve this in most cases (9 out of 12) with only 12 steps.

    1. The method is validated on four text-to-image flow matching models and one audio model. Given the claim of generalization beyond image generation, could the authors provide experiments on video flow matching models (e.g., Wan) to more convincingly demonstrate the breadth of applicability, or explicitly clarify the scope of model architectures and modalities where the method is expected to be effective? I would raise my score if my questions are addressed.

Thanks for the suggestion. We have now evaluated generated videos using the official VBench [1] evaluation implementation and the video flow matching model Wan-2.2-5B. Similar to our evaluation for image generation, we randomly select a fixed subset of 100 unique prompts, assigning 6 to 7 prompts to each of the 16 evaluation dimensions of VBench. Each video is evaluated on its assigned dimension while retaining the official auxiliary annotations. We report the aggregate score computed using VBench’s official normalization and weighting.

NFEs8121620
Wan-2.2-5B
CAT-OV0.72860.74640.74930.7494
CAT-OT0.73860.75030.75550.7565
Fixed0.65910.66930.68450.7182

The results above show that our methods are also more efficient for video generation. With only 8 NFEs, both CAT-OV and CAT-OT already exceed the baseline’s score with 20 NFEs, a 60% reduction in step count. Our methods also outperform the baseline at every NFE budget.

{}

Conclusion: Addressing the reviewer’s two questions has allowed us to strengthen the evaluation of our work, highlighting additional gains by our proposed methods. We truly appreciate the reviewer for that.

{}


References:

[1] Z. Huang et al., "VBench: Comprehensive Benchmark Suite for Video Generative Models," 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA, 2024.

Good paper. Wish you good luck!

We are grateful to the reviewer for the positive and valuable feedback of our paper.

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

The paper proposes two training-free adaptive step-size methods for flow matching. Experiments are conducted to show the effectiveness of the proposed methods.

Critical strengths

The proposed methods are based on a novel and provable connection between flow matching sampling and gradient flow, and reduce the NFEs needed by up to 40% with negligible computational overhead.

Critical weaknesses

The proposed methods are analogous to Adam and RMSProp in SGD. Please explain why Adam and RMSProp can't be directly applied in flow matching.

Questions for authors

No

1 comment

We are grateful to the reviewer for their positive feedback and recommendation on our work.

RMSProp and Adam are per-element adaptive methods: they rescale each coordinate of the gradient by its own learning rate. Applied to flow matching sampling, with the vector field replacing the gradient, this would assign a different step-size to each element of the vector field, which makes them multirate ODE discretizers.

We did not take this route since essentially all flow matching samplers we are aware of, adaptive or not (Euler, adaptive Heun, and the works cited in our paper), use a single scalar step-size. Our work maps gradients to vector fields, so a global learning rate corresponds to that single step-size. This keeps our methods comparable to what is used in practice. To our knowledge, multirate methods are rarely used for flow matching sampling, and their effect on image quality is largely unexplored.

We do use RMSProp’s idea in CAT-OV: the running averages in Eq. (7) follow RMSProp (Appendix B), but we aggregate them into one scalar step-size through the norm. We also tested the other ingredient of Adam, namely, momentum, in Appendix M.1. A full study of per-element methods for flow matching sampling is a valuable direction, and we will mention it in Section 7.

We are happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.