Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

Flow Diverse and Efficient: Learning Momentum Flow Matching via Stochastic Velocity Field Sampling

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Momentum flow matching claims that training on stochastic velocity fields produces trajectories that are both diverse and efficient.

desk verdict Momentum field is a real idea but the training objective as written cannot learn segment-specific velocities, so the central claim isn't established. read the letter →

arxiv 2506.08796 v1 pith:MDUGY7NX submitted 2025-06-10 cs.CV

Zhiyuan Ma , Ruixun Liu , Sixian Liu , Jianjun Li , Bowen Zhou This is my paper
classification cs.CV
keywords flowmatchingrectifieddiffusionmodelsmomentumfieldgenerativemodelingoptimaltransportsamplingefficiencysamplediversity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that rectified flow's straight-line transport, while fast, sacrifices diversity because every trajectory follows a deterministic constant velocity. It proposes a new family of flow models, Momentum Flow (Discretized-RF), that replaces the single straight path with a sequence of short, piecewise-linear sub-paths. On each sub-path the velocity is updated by combining the previous velocity with fresh Gaussian noise, so the forward trajectory becomes increasingly stochastic as it approaches the noise distribution. The reverse process trains a network to match the velocity on each sub-path, preserving rectified flow's fast sampling while recovering some of the diversity of diffusion models. The central empirical claim is that Momentum Flow matches or beats Rectified Flow's FID and recall on CIFAR-10, CelebA-HQ, and ImageNet while using half the sampling steps.

What carries the argument

The central object is the momentum field: a velocity that combines the previous velocity component and a random velocity component at each step, written $v_t = \sqrt{\gamma}\, v_{t-1} + \sqrt{1-\gamma}\,\beta\,\epsilon_t$. This field defines a forward trajectory through anchor points $z_t = z_{t-1} + v_{t-1}$, and its one-step marginal $q(z_t|z_0)$ is a Gaussian whose mean and variance are derived in closed form (Eq. 8, Appendix A.2). The reverse process treats each adjacent pair $(z_t, z_{t-1})$ as a mini rectified-flow problem, inserting intermediate points via linear interpolation and training the network to predict the constant velocity $z_t - z_{t-1}$ on each segment. This piecewise-linear structure is what preserves fast sampling, while the stochastic velocity update is what expands the sampled trajectory space.

What would settle it

One concrete check would be to empirically evaluate whether the reverse composition actually transports the noise distribution to the data distribution: sample $z_T$ from the forward marginal in Eq. 9, run the deterministic ODE through all sub-paths, and measure whether the resulting samples match the empirical data distribution, not just achieve a low FID. A second direct test would be to compute the intermediate marginals $\pi(z_t)$ forward and compare them with the Gaussian form predicted by Eq. 8; if the marginals deviate substantially, the per-sub-path training may be solving a different transport problem than the paper assumes.

Watch

Extended reading notes

Core claim

Momentum Flow claims that learning a momentum field, defined by the recursion $v_t = \sqrt{\gamma}\, v_{t-1} + \sqrt{1-\gamma}\,\beta\,\epsilon_t$ with $\epsilon_t \sim \mathcal{N}(0,I)$, produces trajectories that are both efficient and diverse. The forward process starts from the rectified-flow velocity $v_0 = \beta(\epsilon_0 - x_0)$ and gradually perturbs it, so the trajectory is nearly straight near the data distribution and increasingly stochastic near the noise distribution. The reverse process divides the trajectory into sub-paths between adjacent anchor points and trains a network $u_\theta$ to match the velocity $z_t - z_{t-1}$ on each sub-path, using the objective $\mathcal{L}_{\text{MFM}}(\theta) = \sum_{t=1}^T \mathbb{E}_{m \sim U[0,1]}\left\lVert u_\theta(m z_t + (1-m) z_{t-1}, m) - (z_t - z_{t-1})\right\rVert^2$ (Eq. 13). The paper reports that with this formulation, Momentum Flow retains the sampling efficiency of Rectified Flow while achieving lower FID and higher recall, particularly on CelebA-HQ, and that the number of discretized anchor points $N$ and the momentum decay coefficient $\gamma$ control the trade-off between diversity and fidelity.

Load-bearing premise

The load-bearing premise is that training each short sub-path separately, from the marginal distribution at one anchor point to the marginal at the next, and then stitching the learnt straight segments together, produces a correct reverse trajectory whose final output is a valid sample from the data distribution.

Editorial extensions

If this is right

  • If the central claim is correct, flow-based generators can enjoy diffusion-level diversity without paying the full sampling cost: the same network, trained with the momentum-field objective, can use fewer function evaluations than a standard rectified flow while achieving comparable or better FID and recall.
  • The trade-off parameters $N$ (number of anchor points) and $\gamma$ (momentum decay) become practical knobs for controlling diversity versus fidelity; the paper's experiments indicate that $\gamma$ should be closer to 1 when $N$ is larger, and that the best setting depends on the dataset and budget.
  • The method implies that injecting noise on the velocity field is a viable alternative to injecting noise on $x$ itself, which may be useful for extending flow matching to settings where $x$ is discrete or constrained.
  • Because the reverse process is composed of independently trained sub-path straight lines, the approach can be combined with existing rectified-flow acceleration and distillation techniques that operate on straight segments.
  • The reported ability to generate with NFE as low as 10 while staying competitive suggests that momentum flow could be a drop-in replacement for rectified flow in latent diffusion pipelines, reducing sampling latency without retraining the backbone architecture.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An implicit consequence the paper does not develop is that the momentum-field recursion is mathematically identical in form to the DDPM noise schedule, so the same machinery used for DDPM posterior sampling (Eq. 10) could be adapted to derive a stochastic reverse process for Momentum Flow; the paper instead chooses the deterministic piecewise-linear reverse and does not compare these two options.
  • A testable extension would be to measure whether intermediate marginals $\pi(z_t)$ actually match the Gaussian form predicted by Eq. 8; if they do, the model could be used as a multi-scale feature extraction front-end, not just a generator.
  • The paper's diversity gains are most pronounced on CelebA-HQ and ImageNet, which suggests the method may be especially valuable on high-resolution or highly structured data; applying it to text-to-image models such as latent rectified flows would test whether the same benefit transfers to conditioned generation.
  • The claim would be strengthened by checking whether the deterministic reverse ODE preserves the stochastic diversity of the forward process; the paper uses recall as a proxy but does not directly verify that the reverse composition of sub-path marginals equals the data distribution.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes Discretized-RF (also called momentum flow matching), a flow-based generative model that discretizes the rectified-flow straight path into T sub-paths. In the forward process, the velocity obeys a momentum update v_t = sqrt(gamma) v_{t-1} + sqrt(1-gamma) beta epsilon_t, so that noise is injected on the velocity rather than directly on the sample x. The reverse process is intended to be composed of per-segment rectified flow steps, each solving a linear interpolation between adjacent marginal distributions pi(z_t) and pi(z_{t-1}). The main claims are that this trajectory family improves sample diversity and multi-scale noise modeling relative to rectified flow, and that empirically momentum flow 'consistently matches or even outperforms' rectified flow while using half the number of sampling steps (Section 4.2, Table 1). The paper includes derivations of the forward conditional q(z_t|z0) in Eq(8) and Eq(9), a toy example, and unconditional image generation experiments on CIFAR-10, CelebA-HQ, and ImageNet.

Significance. If the central claim were established, the idea of replacing a single straight transport path by a sequence of momentum-perturbed sub-paths would be a useful contribution to the flow-matching literature, offering a concrete way to trade a small amount of sampling speed for more diverse trajectories while keeping each sub-path linear. The paper has some strengths: the forward conditional derivation in Appendix A.2 is worked out in detail, the use of beta = (sqrt(gamma)-1)/(sqrt(gamma)^T-1) to cancel the z0 coefficient in Eq(9) is algebraically clever, and the experiments span several datasets with a promised code release. However, the theoretical core is not yet established. The reverse training objective in Eq(13) omits the segment index, Algorithm 2 is not an executable generative procedure, and the paper does not show that per-segment flow matching on the intermediate marginals composes into a correct reverse trajectory. The reported FID/recall numbers therefore cannot be traced back to a well-defined algorithm.

major comments (5)
  1. [Section 3.3, Eq(13)] The per-segment objective trains a single network u_theta(z,m) on all T segments, but the input does not contain the segment index t. For T>1, the same pair (z,m) can arise in different segments with different target velocities z_t - z_{t-1}; the minimizer of Eq(13) is then an average over segments, not the segment-specific conditional expectation that the reverse ODE needs. The paper does not show that the marginals pi(z_t) have disjoint supports or that a single network without a segment embedding can represent the piecewise flow.
  2. [Algorithm 2] Algorithm 2 is not an executable generative procedure. Step 4 says to 'draw (z_{t-1}, z_t) from pi(z_{t-1}) x pi(z_t)' and simultaneously 'solve ODE ... with z0 ~ pi_0'; at inference time a sampler does not have access to paired z_{t-1} from the forward process, and the initial condition is written as z0 instead of z_T. Because the reported FID and recall values are produced by this underspecified reverse procedure, the central empirical claim is not reproducible from the manuscript as written.
  3. [Section 3.2, Algorithm 1] The forward chain is internally inconsistent: z_T is declared as an input epsilon0 ~ pi_1 but is then overwritten by the loop z_t = z_{t-1} + v_{t-1}, and v_T is defined in Eq(3) but never used to update any z_t. More importantly, no argument is given that the learned per-segment vector field transports the marginal pi(z_t) to pi(z_{t-1}), so the composition of segments is assumed rather than demonstrated.
  4. [Eq(9)] The claimed endpoint is not exactly pi_1 = N(0,I). With beta = (sqrt(gamma)-1)/(sqrt(gamma)^T-1), the variance of q(z_T|z0) equals 1 + beta^2 [T - (gamma^T - 1)/(gamma - 1)], which exceeds 1 for gamma < 1. If the implementation samples from N(0,I) rather than from the derived endpoint, the forward starting distribution and the reverse initial condition are mismatched; this should be corrected or explicitly justified.
  5. [Section 4.2, Table 1, Figure 5] The empirical claim that momentum flow 'consistently matches or even outperforms' rectified flow is not supported by the full table. The N=5 rows in Table 1 are substantially worse on both datasets at NFE=10 (CIFAR-10 FID 41.28 vs 36.73; CelebA-HQ FID 110.72 vs 98.98), and on CIFAR-10 even the N=2 row at NFE=10 is worse (40.38 vs 36.73). In addition, gamma and N appear to be selected on the test datasets (e.g., gamma=0.99 for N=2 and gamma=0.999 for N=5 in Table 2), which inflates the reported numbers. The claim should be restricted to the configurations actually supported, and hyperparameters should be selected on a validation split.
minor comments (5)
  1. [Notation in Eq(5) and Eq(13)] The training objective in Eq(5) writes t ~ U[0,1], while Eq(13) samples m ~ U[0,1]; the relationship between the global time t and the local segment time m is never defined, even though the network input is m in one equation and t in the other.
  2. [Algorithm 2, line 4] The ODE line 'dz_t/dt = u_theta(z^m_t, m), with z0 ~ pi_0' uses z0 for both the data distribution and the initial condition; rename the initial condition, e.g., z_T^{(0)}, and specify how the segment index t enters the network input.
  3. [Eq(10)] The posterior p_theta(v_{t-1}|v_t) is presented without derivation and without explaining how it follows from the DDPM-style forward process; since the experiments use the second method, either derive the formula carefully or remove it.
  4. [Figure 5] The axes of Figure 5 are not labeled clearly and the legend is hard to read; the text assertion that gamma=0.99 gives 'significantly lower' FID and higher recall would be easier to verify with clearly labeled curves for each N and step configuration.
  5. [Section 1] The bullet claiming that the momentum field is 'the optimal approximation of multi-scale noise between a straight line and a fluctuating line' is stated as fact without a definition of optimality; it should be presented as an empirical motivation or proved in the appendix.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the momentum-flow matching objective is a standard regression on generated pairs, and the reported gains are external benchmark results.

full rationale

The claimed derivation chain is self-contained rather than circular. The forward momentum field is defined by the recurrence v_t = sqrt(gamma) v_{t-1} + sqrt(1-gamma) beta epsilon_t (Eq. 3), and the normalization beta = (sqrt(gamma)-1)/(sqrt(gamma^T)-1) is an algebraic choice that makes q(z_T|z0) zero-mean (Eq. 9); it is not fitted to the target prediction. The training objective (Eq. 13) is a standard conditional flow-matching regression of u_theta(m z_t + (1-m) z_{t-1}, m) onto the known pair difference z_t - z_{t-1}, so the learned velocity is not defined in terms of the generated output. Evaluation is against external benchmarks (CIFAR-10, CelebA-HQ, ImageNet) using FID and recall, so the central claim that momentum flow matches or outperforms rectified flow at half the steps does not reduce to any fitted value; gamma and N are hyperparameters, not predictions. The paper's self-citations in Related Work are contextual and not load-bearing. A reviewer concern about Algorithm 2's 'Draw (z_{t-1}, z_t) from pi(z_{t-1}) x pi(z_t)' is a procedural/identifiability defect in the pseudo-code, not a circularity: the empirical results rest on the Eq. 13-trained ODE, and the text does not define the predicted samples as equal to their training inputs by construction.

Assumptions & free parameters 3 free parameters · 3 assumptions · 1 invented entities

The central method rests on a small number of design choices: the momentum noise schedule (gamma, beta, N) and the assumption that per-segment flow matching composes into a valid transport. These are not independently derived, and the claims of optimality and diversity are asserted rather than proven.

free parameters (3)
  • Momentum decay coefficient gamma = 0.98, 0.99, 0.999, 1.0, chosen per N and dataset
    Gamma controls how quickly the initial velocity v0 decays and how much random velocity noise is added. It is grid-searched on the evaluation datasets using FID and recall curves.
  • Number of discretized anchor points N = 2 and 5, with N=1 as the rectified flow baseline
    N determines the number of sub-paths in the piecewise flow. It is selected per dataset based on the same FID and recall ablations.
  • Velocity normalization coefficient beta = beta = (sqrt(gamma) - 1) / (sqrt(gamma^T) - 1)
    Beta is derived from gamma and T to cancel the x0 coefficient at t=T, but it remains a design choice of the noise schedule, not a quantity fixed by any external principle.
assumptions (3)
  • domain assumption The endpoint z_T of the momentum forward process is a valid sample from the noise distribution pi_1.
    Section 3.2 uses q(z_T|z0) = N(0, ...) as the prior for reverse generation, but the derived variance is not exactly 1 and Algorithm 1 labels z_T as eps0; the paper does not prove or verify that the empirical endpoint distribution matches N(0,I).
  • domain assumption For each discretized pair (pi(z_t), pi(z_{t-1})), flow matching on the linear subpath transports pi(z_t) to pi(z_{t-1}).
    The reverse Algorithm 2 relies on per-segment rectified flow. The paper does not verify that the marginal distributions of the forward chain satisfy the coupling needed for the segment flows to compose correctly.
  • ad hoc to paper The momentum field provides an optimal approximation of multi-scale noise between a straight line and a fluctuating line.
    This is asserted in Section 3 and in the contributions without a formal definition of optimality or a proof; no diversity objective is minimized or even stated in the derivation.
invented entities (1)
  • Momentum field (acceleration field) v_t
    purpose: Introduces random velocity perturbations along flow trajectories to expand sampling diversity while preserving piecewise linear efficiency.
    The construct is defined by Eq(3) and Eq(4) and evaluated only through the paper's own FID and recall experiments; no external benchmark or falsifiable prediction outside the paper supports the claim that there is such an optimal field.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Flow Diverse and Efficient: Learning Momentum Flow Matching via Stochastic Velocity Field Sampling." pith.science (2026). https://pith.science/paper/MDUGY7NX

@misc{pith2026250608796,
  author       = {Pith},
  title        = {Pith review of: Flow Diverse and Efficient: Learning Momentum Flow Matching via Stochastic Velocity Field Sampling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MDUGY7NX}},
  note         = {Machine review of arXiv:2506.08796}
}
abstract

Recently, the rectified flow (RF) has emerged as the new state-of-the-art among flow-based diffusion models due to its high efficiency advantage in straight path sampling, especially with the amazing images generated by a series of RF models such as Flux 1.0 and SD 3.0. Although a straight-line connection between the noisy and natural data distributions is intuitive, fast, and easy to optimize, it still inevitably leads to: 1) Diversity concerns, which arise since straight-line paths only cover a fairly restricted sampling space. 2) Multi-scale noise modeling concerns, since the straight line flow only needs to optimize the constant velocity field $\bm v$ between the two distributions $\bm\pi_0$ and $\bm\pi_1$. In this work, we present Discretized-RF, a new family of rectified flow (also called momentum flow models since they refer to the previous velocity component and the random velocity component in each diffusion step), which discretizes the straight path into a series of variable velocity field sub-paths (namely ``momentum fields'') to expand the search space, especially when close to the distribution $p_\text{noise}$. Different from the previous case where noise is directly superimposed on $\bm x$, we introduce noise on the velocity $\bm v$ of the sub-path to change its direction in order to improve the diversity and multi-scale noise modeling abilities. Experimental results on several representative datasets demonstrate that learning momentum flow matching by sampling random velocity fields will produce trajectories that are both diverse and efficient, and can consistently generate high-quality and diverse results. Code is available at https://github.com/liuruixun/momentum-fm.

Figures

Figures reproduced from arXiv: 2506.08796 by the authors.

Figure 1
Figure 1. Graphical momentum flow trajectories. Momentum Flow (blue) vs. Rectified Flow (green). straight flow only needs to directly optimize the constant velocity field vθ → (x1 −x0) between the two distributions π0 and π1, instead of con￾sidering multi-scale progressive denoising. At the other extreme, the diffusion probability mod￾els (e.g., DDPM) based on fluctuation trajecto￾ries have extremely strong diversity and mult… view at source ↗
Figure 2
Figure 2. Overview of Momentum Flow. Compared with Rectified Flow [ [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Forward and reverse trajectories of momentum flow with different numbers [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: (a) shows the relationship between model performance (FID, Recall) and the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 4
Figure 4. Figure 4: Face generation by Rectified Flow and Momentum Flow under different sampling steps. The Acceleration Process of Momentum Flow. We empirically evaluate the efficiency of mo￾mentum flow in image generation. Although ad￾ditional noise is injected into the velocity field, …
Figure 6
Figure 6. Figure 6: Reverse trajectories of momentum flow and rectified flow at different denoising steps, [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: The impact of adjusting N and γ on image details in Momentum Flow (Step = 50). ’-’ means γ do not influence the trajectories while N = 1. 𝜸 =0.95 𝜸 =0.98 𝜸 =0.99 𝜸 =0.999 𝜸 = 𝟏. 𝟎 [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Generated images under different values of gamma, where both excessively large and [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Deep Neural Networks Inspired by Differential Equations

    cs.LG 2025-10 unverdicted

    A review of differential-equation-inspired neural networks that compiles known results into a taxonomy, with no new experiments or theory.

Reference graph

Works this paper leans on

55 extracted references · 33 canonical work pages · cited by 1 Pith paper

  1. [1]

    Neural flow diffusion models: Learnable forward process for improved diffusion modelling.Advances in Neural Information Processing Systems, 37:73952–73985, 2024

    Grigory Bartosh, Dmitry P Vetrov, and Christian Andersson Naesseth. Neural flow diffusion models: Learnable forward process for improved diffusion modelling.Advances in Neural Information Processing Systems, 37:73952–73985, 2024

    work page 2024

  2. [2]

    Fluxspace: Disentangled semantic editing in rectified flow transformers, 2024

    Yusuf Dalva, Kavana Venkatesh, and Pinar Yanardag. Fluxspace: Disentangled semantic editing in rectified flow transformers, 2024

    work page 2024

  3. [3]

    Flow matching in latent space, 2023

    Quan Dao, Hao Phung, Binh Nguyen, and Anh Tran. Flow matching in latent space, 2023

    work page 2023

  4. [4]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. InForty-first International Conference on Machine Learning, 2024

    work page 2024

  5. [5]

    Scaling rectified flow transformers for high-resolution image synthesis, 2024

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024

    2024

  6. [6]

    Itai Gat, Tal Remez, Neta Shaul, Felix Kreuk, Ricky T. Q. Chen, Gabriel Synnaeve, Yossi Adi, and Yaron Lipman. Discrete flow matching. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors,Advances in Neural Information Processing Systems, volume 37, pages 133345–133385. Curran Associates, Inc., 2024

    work page 2024

  7. [7]

    Seeds: Exponential sde solvers for fast high-quality sampling from diffusion models

    Martin Gonzalez, Nelson Fernandez Pinto, Thuy Tran, elies Gherbi, Hatem Hajri, and Nader Masmoudi. Seeds: Exponential sde solvers for fast high-quality sampling from diffusion models. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors,Advances in Neural Information Processing Systems, volume 36, pages 68061–68120. Curran Assoc...

    work page 2023

  8. [8]

    Pengsheng Guo and Alexander G. Schwing. Variational rectified flow matching, 2025

    work page 2025

Show all 55 references
  1. [9]

    Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  2. [10]

    Progressive growing of gans for improved quality, stability, and variation

    Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. InProceedings of the International Conference on Learning Representations (ICLR), 2018

    2018

  3. [11]

    Denoising diffusion restoration models.arXiv preprint arXiv:2201.11793, 2022

    Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models.arXiv preprint arXiv:2201.11793, 2022

    2022 arXiv

  4. [12]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical Report TR-2009, University of Toronto, 2009. 10

    2009

  5. [13]

    Flux.https://github.com/black-forest-labs/flux, 2024

    Black Forest Labs. Flux.https://github.com/black-forest-labs/flux, 2024

    2024

  6. [14]

    Improving the training of rectified flows.Advances in Neural Information Processing Systems, 37:63082–63109, 2024

    Sangyun Lee, Zinan Lin, and Giulia Fanti. Improving the training of rectified flows.Advances in Neural Information Processing Systems, 37:63082–63109, 2024

    2024

  7. [15]

    Distrifusion: Distributed parallel inference for high-resolution diffusion models

    Muyang Li, Tianle Cai, Jiaxin Cao, Qinsheng Zhang, Han Cai, Junjie Bai, Yangqing Jia, Kai Li, and Song Han. Distrifusion: Distributed parallel inference for high-resolution diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p...

    2024

  8. [16]

    Faster diffusion: Rethinking the role of the encoder for diffusion model inference

    Senmao Li, Taihang Hu, Joost van de Weijer, Fahad Shahbaz Khan, Tao Liu, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster diffusion: Rethinking the role of the encoder for diffusion model inference. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U....

    2024

  9. [17]

    Omniflow: Any-to-any generation with multi-modal rectified flows

    Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Zichun Liao, Yusuke Kato, Kazuki Kozuka, and Aditya Grover. Omniflow: Any-to-any generation with multi-modal rectified flows. arXiv preprint arXiv:2412.01169, 2024

    2024 arXiv

  10. [18]

    Snapfusion: Text-to-image diffusion model on mobile devices within two seconds.Advances in Neural Information Processing Systems, 36:20662–20678, 2023

    Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds.Advances in Neural Information Processing Systems, 36:20662–20678, 2023

    2023

  11. [19]

    Flow matching for generative modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. InThe Eleventh International Conference on Learning Representations

  12. [20]

    Pseudo numerical methods for diffusion models on manifolds, 2022

    Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds, 2022

    2022

  13. [21]

    Rfwave: Multi-band rectified flow for audio waveform reconstruction, 2024

    Peng Liu, Dongyang Dai, and Zhiyong Wu. Rfwave: Multi-band rectified flow for audio waveform reconstruction, 2024

    2024

  14. [22]

    Rectified flow: A marginal preserving approach to optimal transport, 2022

    Qiang Liu. Rectified flow: A marginal preserving approach to optimal transport, 2022

    2022

  15. [23]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. InThe Eleventh International Conference on Learning Representations (ICLR), 2023

    2023

  16. [24]

    Instaflow: One step is enough for high-quality diffusion-based text-to-image generation

    Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. InThe Twelfth International Conference on Learning Representations, 2023

    2023

  17. [25]

    Fixing weight decay regularization in adam.arXiv preprint arXiv:1711.05101, 5:5, 2017

    Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam.arXiv preprint arXiv:1711.05101, 5:5, 2017

    2017 arXiv

  18. [26]

    Flowdiffuser: Advancing optical flow estimation with diffusion models

    Ao Luo, Xin Li, Fan Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Flowdiffuser: Advancing optical flow estimation with diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19167–19176, 2024

    2024

  19. [27]

    Deepcache: Accelerating diffusion models for free

    Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15762–15772, June 2024

    2024

  20. [28]

    Safe-sd: Safe and traceable stable diffusion with text prompt trigger for invisible generative watermarking

    Zhiyuan Ma, Guoli Jia, Biqing Qi, and Bowen Zhou. Safe-sd: Safe and traceable stable diffusion with text prompt trigger for invisible generative watermarking. InACM Multimedia 2024

    2024

  21. [29]

    Adapedit: Spatio-temporal guided adaptive edit- ing algorithm for text-based continuity-sensitive image editing

    Zhiyuan Ma, Guoli Jia, and Bowen Zhou. Adapedit: Spatio-temporal guided adaptive edit- ing algorithm for text-based continuity-sensitive image editing. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 4154–4161, 2024

    2024

  22. [30]

    Lmd: faster image reconstruction with latent masking diffusion

    Zhiyuan Ma, Zhihuan Yu, Jianjun Li, and Bowen Zhou. Lmd: faster image reconstruction with latent masking diffusion. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 4145–4153, 2024

    2024

  23. [31]

    Efficient diffusion models: A comprehensive survey from principles to practices.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    Zhiyuan Ma, Yuzhu Zhang, Guoli Jia, Liangliang Zhao, Yichao Ma, Mingjie Ma, Gaofeng Liu, Kaiyan Zhang, Ning Ding, Jianjun Li, et al. Efficient diffusion models: A comprehensive survey from principles to practices.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 11

    2025

  24. [32]

    Neural residual diffusion models for deep scalable vision generation

    Zhiyuan Ma, Liangliang Zhao, Biqing Qi, and Bowen Zhou. Neural residual diffusion models for deep scalable vision generation. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  25. [33]

    On distillation of guided diffusion models

    Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14297–14306, June 2023

    2023

  26. [34]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. InInternational conference on machine learning, pages 8162–8171. PMLR, 2021

    2021

  27. [35]

    Posterior-mean rectified flow: Towards minimum mse photo-realistic image restoration, 2025

    Guy Ohayon, Tomer Michaeli, and Michael Elad. Posterior-mean rectified flow: Towards minimum mse photo-realistic image restoration, 2025

    2025

  28. [36]

    Mehdi S. M. Sajjadi, Olivier Bachem, Mario Lucic, Olivier Bousquet, and Sylvain Gelly. Assessing generative models via precision and recall. InAdvances in Neural Information Processing Systems (NeurIPS), 2018

    2018

  29. [37]

    Progressive distillation for fast sampling of diffusion models, 2022

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models, 2022

    2022

  30. [38]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. InEuropean Conference on Computer Vision, pages 87–103. Springer, 2024

    2024

  31. [39]

    pytorch-fid: FID Score for PyTorch

    Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/ pytorch-fid, August 2020. Version 0.3.0

    2020

  32. [40]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  33. [41]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. 2023

    2023

  34. [42]

    Generative modeling by estimating gradients of the data distribution.Advances in neural information processing systems, 32, 2019

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution.Advances in neural information processing systems, 32, 2019

    2019

  35. [43]

    Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    2011 arXiv

  36. [44]

    Rectified diffusion: Straightness is not your need in rectified flow, 2024

    Fu-Yun Wang, Ling Yang, Zhaoyang Huang, Mengdi Wang, and Hongsheng Li. Rectified diffusion: Straightness is not your need in rectified flow, 2024

    2024

  37. [45]

    Taming rectified flow for inversion and editing, 2024

    Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Taming rectified flow for inversion and editing, 2024

    2024

  38. [46]

    Frieren: Efficient video-to-audio generation with rectified flow matching

    Yongqi Wang, Wenxiang Guo, Rongjie Huang, Jiawei Huang, Zehan Wang, Fuming You, Ruiqi Li, and Zhou Zhao. Frieren: Efficient video-to-audio generation with rectified flow matching. arXiv e-prints, pages arXiv–2406, 2024

    2024

  39. [47]

    Ufogen: You forward once large scale text-to-image generation via diffusion gans

    Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8196–8206, 2024

    2024

  40. [48]

    Perflow: Piecewise rectified flow as universal plug-and-play accelerator.arXiv preprint arXiv:2405.07510, 2024

    Hanshu Yan, Xingchao Liu, Jiachun Pan, Jun Hao Liew, Qiang Liu, and Jiashi Feng. Perflow: Piecewise rectified flow as universal plug-and-play accelerator.arXiv preprint arXiv:2405.07510, 2024

    2024 arXiv

  41. [49]

    Improving training efficiency of diffusion models via multi-stage framework and tailored multi-decoder architecture

    Huijie Zhang, Yifu Lu, Ismail Alkhouri, Saiprasad Ravishankar, Dogyoon Song, and Qing Qu. Improving training efficiency of diffusion models via multi-stage framework and tailored multi-decoder architecture. InProceedings of the IEEE/CVF Conference on Computer Vision and Patter...

    2024

  42. [50]

    Mobilediffusion: Instant text-to-image generation on mobile devices

    Yang Zhao, Yanwu Xu, Zhisheng Xiao, Haolin Jia, and Tingbo Hou. Mobilediffusion: Instant text-to-image generation on mobile devices. InEuropean Conference on Computer Vision, pages 225–242. Springer, 2024

    2024

  43. [51]

    Flowie: Efficient image enhancement via rectified flow

    Yixuan Zhu, Wenliang Zhao, Ao Li, Yansong Tang, Jie Zhou, and Jiwen Lu. Flowie: Efficient image enhancement via rectified flow. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13–22, June 2024. 12 A Appendix A.1 Proof of the On...

    2024

  44. [52]

    CIFAR-10: Images with a resolution of32×32from the CIFAR-10 training set

  45. [53]

    CelebA-HQ: Images from the ‘img_align_celeba_png.7z‘ version of the CelebA-HQ dataset, resized to256×256

  46. [54]

    During training, images are normalized to have zero mean and unit variance

  47. [55]

    During training, images are normalized to have zero mean and unit variance

    ImageNet: Images from ImageNet resized to32×32and64×64. During training, images are normalized to have zero mean and unit variance. Training Details:The model is trained using the Adam optimizer with a learning rate of 3×10 −4. For ImageNet, we use a batch size of 64, whereas ...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.