Pith. sign in

REVIEW 5 major objections 5 minor 5 cited by

SplitMeanFlow: Interval Splitting Consistency in Few-Step Generative Modeling

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read SplitMeanFlow replaces MeanFlow's derivative identity with an algebraic interval-splitting identity, giving a JVP-free training objective for few-step generation.

desk verdict Clean algebraic identity and a correct limit proof, but the training objective rests on an unproven bootstrap and the experiments need stronger evidence. read the letter →

arxiv 2507.16884 v1 pith:F3KR56FV submitted 2025-07-22 cs.LG cs.AI

Yi Guo , Wei Wang , Zhihang Yuan , Rong Cao , Kuan Chen , Zhengyang Chen , Yuanyuan Huo , Yang Zhang
show 3 more authors
Yuping Wang Shouda Liu Yuxuan Wang
This is my paper · ORCID
classification cs.LGcs.AI
keywords few-stepgenerativemodelingaveragevelocityfieldIntervalSplittingConsistencyMeanFlowmatchingone-stepgenerationJVP-freetrainingspeechsynthesis
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 the right learning target for few-step generative models is the average velocity field, and that this field is governed by an exact algebraic identity it calls Interval Splitting Consistency. The identity states that a displacement over $[r,t]$ equals the sum of displacements over $[r,s]$ and $[s,t]$ in terms of average velocities, a direct consequence of additivity of definite integrals. The paper proves that MeanFlow's differential identity $u = v - (t-r)\frac{d}{dt}u$ is the infinitesimal limit of this algebraic relation, making MeanFlow a special case. Training SplitMeanFlow by enforcing the identity requires only standard forward and backward passes, no Jacobian-vector products, and the reported one- and two-step speech synthesis results match a ten-step flow matching baseline. The paper's claim is that this algebraic consistency is the more fundamental principle behind average-velocity learning.

What carries the argument

The load-bearing object is the Interval Splitting Consistency identity, Eq. (10): $$(t-r)u(z_t,r,t) = (s-r)u(z_s,r,s) + (t-s)u(z_t,s,t).$$ It is derived purely from the first-principles definition of average velocity as an integral and the additivity of definite integrals, without differential operators. The identity provides a self-referential training signal: the average velocity over the full interval is expressed as a weighted sum of average velocities over two sub-intervals, so the network supervises itself at self-generated intermediate points. The proof that MeanFlow is a special case works by writing $g(t) = (t-r)u(z_t,r,t)$, recognizing the left side of the rearranged identity as $(g(t)-g(s))/(t-s)$, and identifying its $s \to t$ limit as $g'(t) = v(z_t,t)$. The boundary condition $u(z_t,t,t) = v(z_t,t)$, anchored by the teacher network for a fraction of each batch, is what keeps this self-supervision from collapsing to a degenerate solution.

What would settle it

Train SplitMeanFlow from scratch with no teacher boundary condition and check whether the consistency loss can reach zero while the network fails to generate data-like samples; if a degenerate field such as a constant velocity satisfies the identity but produces poor samples, the boundary anchor rather than the algebraic identity is carrying the method. Alternatively, on a fixed pretrained flow model, compute the true average velocity by numerical integration, train SplitMeanFlow on the consistency loss, and compare the learned field to the true average velocity on held-out $(z_t, r, s, t)$ tuples; agreement would confirm the identity, while a small consistency loss with large deviation from the true integral would show the objective is self-consistent but wrong.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the average velocity $u(z_t,r,t)$, defined as the integral of instantaneous velocity over $[r,t]$ divided by $t-r$, satisfies $$(t-r)u(z_t,r,t) = (s-r)u(z_s,r,s) + (t-s)u(z_t,s,t)$$ for any $r \le s \le t$, because the displacement integral splits additively. SplitMeanFlow turns this identity into a training objective: for a sampled split with $\lambda = (t-s)/(t-r)$, it predicts $u_2 = u_\theta(z_t,s,t)$, constructs the intermediate flow point $z_s = z_t - (t-s)u_2$, predicts $u_1 = u_\theta(z_s,r,s)$, and regresses $u_\theta(z_t,r,t)$ toward $(1-\lambda)u_1 + \lambda u_2$ with stop-gradient, while a fraction of each batch anchors the boundary condition $u(z_t,t,t) = v(z_t,t)$ using a pretrained flow-matching teacher. The paper further shows that taking $s \to t$ in the identity recovers the MeanFlow identity $u = v - (t-r)\frac{d}{dt}u$, so the differential formulation is a limiting case rather than an independent law.

Load-bearing premise

Training constructs the intermediate point $z_s = z_t - (t-s)u_2$ using the model's own predicted average velocity, so the identity is enforced at a point that is only as faithful to the true flow as that prediction; the method relies on this bootstrap point and the teacher-anchored boundary condition being accurate enough to pull the network toward the true average velocity instead of a degenerate self-consistent solution.

Editorial extensions

If this is right

  • If the identity is correct and exact, any average-velocity model can be trained with a small number of standard forward passes and one backward pass per sample, removing JVP as a requirement for this class of few-step methods.
  • MeanFlow's differential identity becomes a corollary, so theoretical guarantees or intuitions developed for MeanFlow transfer to SplitMeanFlow, while SplitMeanFlow's validity does not depend on differentiability of the learned field in the same way.
  • One- and two-step sampling can match the quality of a ten-step flow matching baseline in the reported speech synthesis tasks, with classifier-free guidance turned off; that is a directly deployable speedup for latency-sensitive generation.
  • Because the identity holds for every intermediate split point $s$, the training procedure can sample $r$, $s$, $t$ continuously, generalizing midpoint-only shortcut-model formulations.
  • The two-stage recipe of flow matching pretraining followed by consistency distillation with a teacher-anchored boundary condition becomes a stable practical recipe for large-scale few-step deployment.

Reading between the lines

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

  • An implicit consequence: the accuracy of the self-supervised target is only as good as the self-generated intermediate point $z_s = z_t - (t-s)u_2$; a natural test is to replace $u_2$ with the teacher's true average velocity during training and measure whether one-step quality improves, which would locate the residual error in the bootstrap rather than in the identity.
  • The same integral-additivity argument applies to any quantity defined as a path integral, such as the score function or the drift of a probability-flow ODE; an analogous splitting consistency could yield JVP-free objectives in those settings.
  • The paper does not quantify how the flow ratio $p$ or the choice of split-point distribution affects convergence; an extension would sweep $p$ and the distribution of $s$ to find where the boundary anchor stops mattering.
  • If the identity is as fundamental as claimed, then the distance between $u_\theta(z_t,r,t)$ and the true average velocity should be controlled by the consistency residual plus the boundary error; measuring both separately would give a sharper training objective.
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 introduces SplitMeanFlow, a training framework for few-step generative modeling based on an algebraic identity called Interval Splitting Consistency. For the average velocity field u(z_t,r,t), the identity (t-r)u(z_t,r,t) = (s-r)u(z_s,r,s) + (t-s)u(z_t,s,t) is derived directly from additivity of definite integrals. The authors prove that the MeanFlow differential identity u = v - (t-r) d/dt u is recovered in the limit s→t, and propose a JVP-free training objective in which a network regresses the full-interval average velocity against a stop-gradient convex combination of its own predictions on two sub-intervals, anchored by a teacher velocity boundary condition at r=t. Experiments on Seed-TTS speech synthesis report two-step and one-step performance comparable to a 10-step Flow Matching baseline, with the method deployed in a large-scale product.

Significance. The algebraic identity in Eq. (10) is a clean and correct observation, and the limit argument in Section 4.3 is a genuine deduction rather than a fitted consistency relation. The JVP-free objective is practically attractive, and the claimed large-scale deployment indicates real-world viability. However, the central claim that this training procedure learns the true average velocity field relies on a bootstrap assumption that is not analyzed, and the experimental evidence is too thin to establish the method's advantages over the closest prior baseline. The paper would be strengthened substantially by an identifiability or convergence analysis for the self-generated intermediate points and by a matched comparison against MeanFlow.

major comments (5)
  1. [Algorithm 1; Section 5.1] The central training procedure is not fully specified. Algorithm 1 only shows the consistency branch and explicitly says the boundary conditions are omitted, while Section 5.1 gives a flow ratio p but does not state the exact loss composition, how the r=t samples are constructed, or how the boundary and consistency terms are weighted or alternated. Because the boundary anchor is the only mechanism cited to prevent drift of the self-supervised target, a complete and reproducible specification of this branch is required.
  2. [Algorithm 1, lines 4-6; Eq. (10)] Eq. (10) is an exact identity for points on the true flow trajectory, but Algorithm 1 constructs the intermediate point as z_s = z_t - (t-s)u_2 using the model's own prediction, with stop-gradient on u_1, u_2, and z_s. The consistency loss is therefore evaluated at self-generated, potentially off-trajectory points, and there is no direct mechanism pulling z_s onto the true trajectory. The paper provides no fixed-point or identifiability argument showing that the bootstrap objective plus the boundary condition has the true average velocity field as its unique or attracting solution. A concrete test would be to run the training on a synthetic ODE with known instantaneous velocity and compare the learned u against the true average velocity, or to ablate self-generated z_s against teacher-generated or ground-truth z_s.
  3. [Tables 1 and 2] The experimental results are reported without error bars, confidence intervals, significance tests, or sample sizes. The claim that the one-step model is 'statistically on par' with the 10-step baseline is unsupported, and CMOS values of 0 are interpreted as perceptual equivalence without any measure of uncertainty. The experiments should be rerun over multiple seeds and evaluator sets, with confidence intervals reported for SIM, WER, and CMOS.
  4. [Section 5.2; Tables 1 and 2] MeanFlow, the method that this paper claims to generalize and improve upon, is not included among the baselines. Since the paper's theoretical contribution is that SplitMeanFlow is a more general and more practical formulation than MeanFlow, the tables cannot support that comparative claim without a matched evaluation of MeanFlow under the same training budget, including quality metrics, training stability, and wall-clock time.
  5. [Section 5.1] The method depends on two free parameters, the boundary-condition mixing ratio p and the teacher CFG scale, but no sensitivity analysis or schedule is reported. The paper states only that p ≥ 0.5 is required for stability; without ablations, it is unclear how robust the objective is to these choices or how they should be set for other domains.
minor comments (5)
  1. [Eq. (11)] The definition of λ as (t-s)/(t-r) is correct, but the text says 'it gives s = (1-λ)t + λr' without showing the algebra; a short derivation would avoid confusion with the weighting coefficients in the convex combination.
  2. [Algorithm 1; Section 5.1] The network is written as u_θ(z, r, t), but the architecture and the way the interval endpoints r and t are encoded are not described; this is needed for reproducibility.
  3. [Section 4.3.2] The sentence beginning 'Second, the backpropagation step demands high numerical precision...' is incomplete and does not clearly connect to the JVP computation described earlier; please rewrite for clarity.
  4. [Tables 1 and 2] The tables do not report real-time factor or wall-clock inference time, so the claimed 20x speedup is not directly measured in the experiments.
  5. [Abstract and Section 1] The product name is spelled 'Doubao' in the abstract and 'DouBao' in the contributions list; please use a consistent spelling.

Circularity Check

0 steps flagged · score 1.0 of 10

Derivation is self-contained; no circular reduction found. Eq. 10 follows from the definition of average velocity and integral additivity, and the MeanFlow-limit proof is a genuine deduction; the self-referential training target is a stability concern, not a circularity.

full rationale

The central identity (Eq. 10) is derived directly from the definition u(zt,r,t) = (1/(t-r))∫_r^t v dτ (Eq. 7) and the additivity of definite integrals (Eq. 9). No parameter is fitted to make the identity true; it is exact for the true field by construction. The proof that MeanFlow's differential identity is recovered in the limit s→t (Eqs. 12-18) is a self-contained mathematical argument using the fundamental theorem of calculus and the product rule; it does not import the MeanFlow identity as an assumption. There is no load-bearing self-citation chain: MeanFlow [9] is external prior work, and the Seed-TTS [2] and speaker-verification [4] citations are tools or baselines, not premises of the derivation. The paper explicitly concedes that the Shortcut Model [7] already achieves partial equivalence for s=(r+t)/2, so the algebraic consistency idea is not presented as entirely unprecedented; this is a prior-art limitation rather than a circular step. The main residual concern is algorithmic, not circular: in Algorithm 1, zs is constructed from the model's own prediction u2, so the consistency loss is enforced at a self-generated point rather than at a verifiably exact trajectory point. The paper acknowledges the need for a boundary condition (Section 5.1: 'The loss requires a boundary condition to avoid collapsing to a trivial solution') and a teacher-anchored flow ratio p≥0.5 to prevent drift, but it does not prove that the self-bootstrap fixed point is unique or equals the true average velocity. This is an identifiability or stability gap in the training objective, not a case where a predicted quantity reduces by construction to its input. Because the theoretical derivation is honest and the empirical claims are benchmarked externally (SIM/WER/CMOS against Flow Matching and DMD), the paper does not exhibit self-definitional, fitted-input-as-prediction, or self-citation circularity. Score 1 reflects the minor self-referential training design, which is not load-bearing on the paper's mathematical claims.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central identity rests only on standard integral calculus and the flow matching path convention. The empirical method adds two hand-chosen training hyperparameters, the boundary mixing ratio and the teacher CFG scale, and assumes that self-generated sub-interval targets, anchored by a teacher boundary condition, converge to the true average velocity. No new entities are introduced.

free parameters (2)
  • Boundary condition mixing ratio p = p >= 0.5, exact value not stated
    Section 5.1 states 'a stable and effective training regime requires p >= 0.5'; this hand-chosen ratio controls the fraction of the batch anchored to the teacher's instantaneous velocity, and it is tuned for training stability.
  • Teacher CFG scale = Not stated
    Section 5.1 says the teacher generates instantaneous velocity with a fixed CFG scale, but the value is not reported; this scale affects the boundary targets and therefore the learned velocity field.
assumptions (5)
  • standard math Additivity of definite integrals over partitioned intervals
    Eq. 9: the integral over [r, t] equals the sum over [r, s] and [s, t]; this is the foundation of the Interval Splitting Consistency identity.
  • standard math Definition of average velocity as the integral of instantaneous velocity divided by interval length
    Eq. 7, used to translate integral additivity into Eq. 10.
  • domain assumption The average velocity field is smooth enough for the limit s -> t and the total derivative to be taken
    Section 4.3.1 uses (g(t) - g(s))/(t - s) -> g'(t) and the chain rule; this requires differentiability of the field and the trajectory.
  • domain assumption The linear interpolated path z_t = (1 - t)x + t epsilon represents the probability flow
    Eq. 2, standard flow matching setup; the conditional velocity v_t = epsilon - x is used to build the teacher.
  • domain assumption Self-generated sub-interval targets with a teacher boundary condition converge to the true average velocity
    Algorithm 1 builds the target from the student's own predictions u_1 and u_2; the paper offers empirical evidence but no convergence proof.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SplitMeanFlow: Interval Splitting Consistency in Few-Step Generative Modeling." pith.science (2026). https://pith.science/paper/F3KR56FV

@misc{pith2026250716884,
  author       = {Pith},
  title        = {Pith review of: SplitMeanFlow: Interval Splitting Consistency in Few-Step Generative Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F3KR56FV}},
  note         = {Machine review of arXiv:2507.16884}
}
read the original abstract

Generative models like Flow Matching have achieved state-of-the-art performance but are often hindered by a computationally expensive iterative sampling process. To address this, recent work has focused on few-step or one-step generation by learning the average velocity field, which directly maps noise to data. MeanFlow, a leading method in this area, learns this field by enforcing a differential identity that connects the average and instantaneous velocities. In this work, we argue that this differential formulation is a limiting special case of a more fundamental principle. We return to the first principles of average velocity and leverage the additivity property of definite integrals. This leads us to derive a novel, purely algebraic identity we term Interval Splitting Consistency. This identity establishes a self-referential relationship for the average velocity field across different time intervals without resorting to any differential operators. Based on this principle, we introduce SplitMeanFlow, a new training framework that enforces this algebraic consistency directly as a learning objective. We formally prove that the differential identity at the core of MeanFlow is recovered by taking the limit of our algebraic consistency as the interval split becomes infinitesimal. This establishes SplitMeanFlow as a direct and more general foundation for learning average velocity fields. From a practical standpoint, our algebraic approach is significantly more efficient, as it eliminates the need for JVP computations, resulting in simpler implementation, more stable training, and broader hardware compatibility. One-step and two-step SplitMeanFlow models have been successfully deployed in large-scale speech synthesis products (such as Doubao), achieving speedups of 20x.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 5 Pith papers

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

  1. Genuine pair density wave order on the kagome lattice

    cond-mat.supr-con 2026-04 unverdicted novelty 7.0 of 10

    A genuine primary pair-density-wave phase emerges as a competing ground state in a two-orbital kagome Hubbard model over a wide parameter range, driven by sublattice- and orbital-polarized Fermi pockets.

  2. Expanding Flow Maps

    cs.LG 2026-07 conditional novelty 6.0 of 10

    Expanding Flow Maps make a single flow map grow its state dimensionality during inference, enabling few-step variable-size generation over continuous and discrete data.

  3. High-Fidelity One-Step Generative Visuomotor Policy via Recursive Correction, Frequency Consistency, and Contrastive Flow Matching

    cs.RO 2026-07 conditional novelty 6.0 of 10

    One-step flow-matching visuomotor policy with recursive correction, dual-timestep spectral consistency, and contrastive mode separation matches or exceeds 10-step baselines at 1 NFE.

  4. CaloTrilogy: Toward a Breakthrough in One-Step, End-to-End, Physics-Guided Shower Generation for Modern Calorimeters

    hep-ex 2026-06 unverdicted novelty 6.0 of 10

    Presents CaloTrilogy, a unified one-step generative model for high-granularity calorimeter showers that combines velocity field integration, learned priors, and physics losses to match SOTA quality.

  5. Dual-End Consistency Model

    cs.CV 2026-02 unverdicted novelty 6.0 of 10

    DE-CM trains a flow-map consistency model on three sub-trajectories (coupling, instantaneous, noise-to-noisy) and reports 1.70 FID one-step on ImageNet 256.

Reference graph

Works this paper leans on

34 extracted references · 8 canonical work pages · cited by 5 Pith papers

  1. [1]

    Stochastic interpolants: A unifying framework for flows and diffusions.arXiv preprint arXiv:2303.08797, 2023

    Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions.arXiv preprint arXiv:2303.08797, 2023

    arXiv 2023

  2. [2]

    Seed-tts: A family of high-quality versatile speech generation models.arXiv preprint arXiv:2406.02430, 2024

    Philip Anastassiou, Jiawei Chen, Jitong Chen, Yuanzhe Chen, Zhuo Chen, Ziyi Chen, Jian Cong, Lelai Deng, Chuang Ding, Lu Gao, et al. Seed-tts: A family of high-quality versatile speech generation models.arXiv preprint arXiv:2406.02430, 2024

    arXiv 2024

  3. [3]

    Lumiere: A space-time diffusion model for video generation.arXiv preprint arXiv:2401.12945, 2024

    Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, et al. Lumiere: A space-time diffusion model for video generation.arXiv preprint arXiv:2401.12945, 2024

    arXiv 2024

  4. [4]

    Large-scale self-supervised speech representation learning for automatic speaker verification

    Zhengyang Chen, Sanyuan Chen, Yu Wu, Yao Qian, Chengyi Wang, Shujie Liu, Yanmin Qian, and Michael Zeng. Large-scale self-supervised speech representation learning for automatic speaker verification. InICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6147–6151. IEEE, 2022

    work page 2022

  5. [5]

    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, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-firstinternational conference on machine learning, 2024

    2024

  6. [6]

    Music consistency models.arXiv preprint arXiv:2404.13358, 2024

    Zhengcong Fei, Mingyuan Fan, and Junshi Huang. Music consistency models.arXiv preprint arXiv:2404.13358, 2024

    arXiv 2024

  7. [7]

    One step diffusion via shortcut models

    Kevin Frans, Danijar Hafner, Sergey Levine, and Pieter Abbeel. One step diffusion via shortcut models. In International Conference on Learning Representations (ICLR), 2025

    2025

  8. [8]

    Funasr: A fundamental end-to-end speech recognition toolkit.arXiv preprint arXiv:2305.11013, 2023

    Zhifu Gao, Zerui Li, Jiaming Wang, Haoneng Luo, Xian Shi, Mengzhe Chen, Yabin Li, Lingyun Zuo, Zhi- hao Du, Zhangyu Xiao, et al. Funasr: A fundamental end-to-end speech recognition toolkit.arXiv preprint arXiv:2305.11013, 2023

    arXiv 2023

Show all 34 references
  1. [9]

    Mean flows for one-step generative modeling

    Zhengyang Geng, Mingyang Deng, Xingjian Bai, J Zico Kolter, and Kaiming He. Mean flows for one-step generative modeling. arXiv preprint arXiv:2505.13447, 2025

    2025 arXiv

  2. [10]

    Lafma: A latent flow matching model for text-to-audio generation.arXiv preprint arXiv:2406.08203, 2024

    Wenhao Guan, Kaidi Wang, Wangjin Zhou, Yang Wang, Feng Deng, Hui Wang, Lin Li, Qingyang Hong, and Yong Qin. Lafma: A latent flow matching model for text-to-audio generation.arXiv preprint arXiv:2406.08203, 2024

    2024 arXiv

  3. [11]

    Multistep consistency models

    Jonathan Heek, Emiel Hoogeboom, and Tim Salimans. Multistep consistency models. arXiv preprint arXiv:2403.06807, 2024

    2024 arXiv

  4. [12]

    Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    2022 arXiv

  5. [13]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Neural Information Processing Systems (NeurIPS), 2020

    2020

  6. [14]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. InNeural Information Processing Systems (NeurIPS), 2022

    2022

  7. [15]

    Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    2024 arXiv

  8. [16]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. InInternational Conference on Learning Representations (ICLR), 2023

    2023

  9. [17]

    Audiolcm: Text-to-audio generation with latent consistency models.arXiv preprint arXiv:2406.00356, 2024

    Huadai Liu, Rongjie Huang, Yang Liu, Hengyuan Cao, Jialei Wang, Xize Cheng, Siqi Zheng, and Zhou Zhao. Audiolcm: Text-to-audio generation with latent consistency models.arXiv preprint arXiv:2406.00356, 2024

    2024 arXiv

  10. [18]

    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. InInternational Conference on Learning Representations (ICLR), 2023

    2023

  11. [19]

    Simplifying, stabilizing and scaling continuous-time consistency models.arXiv preprint arXiv:2410.11081, 2024

    Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models.arXiv preprint arXiv:2410.11081, 2024

    2024 arXiv

  12. [20]

    Latent consistency models: Synthesizing high-resolution images with few-step inference, 2023

    Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference, 2023. 12

    2023

  13. [21]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning (ICML). PMLR, 2021

    2021

  14. [22]

    Rl for consistency models: Faster reward guided text-to-image generation.arXiv preprint arXiv:2404.03673, 2024

    Owen Oertell, Jonathan D Chang, Yiyi Zhang, Kianté Brantley, and Wen Sun. Rl for consistency models: Faster reward guided text-to-image generation.arXiv preprint arXiv:2404.03673, 2024

    2024 arXiv

  15. [23]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  16. [24]

    Flow-anchored consistency models

    Yansong Peng, Kai Zhu, Yu Liu, Pingyu Wu, Hebei Li, Xiaoyan Sun, and Feng Wu. Flow-anchored consistency models. arXiv preprint arXiv:2507.03738, 2025

    2025

  17. [25]

    Robust speech recognition via large-scale weak supervision

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. InInternational conference on machine learning, pages 28492–28518. PMLR, 2023

    2023

  18. [26]

    High-resolutionimagesyn- thesis with latent diffusion models

    RobinRombach, AndreasBlattmann, DominikLorenz, PatrickEsser, andBjörnOmmer. High-resolutionimagesyn- thesis with latent diffusion models. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021

    2021

  19. [27]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  20. [28]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. InInternational Conference on Machine Learning (ICML), 2015

    2015

  21. [29]

    Generative modeling by estimating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Neural Information Processing Systems (NeurIPS), 2019

    2019

  22. [30]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021

    2021

  23. [31]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In International Conference on Machine Learning (ICML), 2023

    2023

  24. [32]

    Audiox: Diffusion transformer for anything-to-audio generation.arXiv preprint arXiv:2503.10522, 2025

    Zeyue Tian, Yizhu Jin, Zhaoyang Liu, Ruibin Yuan, Xu Tan, Qifeng Chen, Wei Xue, and Yike Guo. Audiox: Diffusion transformer for anything-to-audio generation.arXiv preprint arXiv:2503.10522, 2025

    2025 arXiv

  25. [33]

    Diffusion models: A comprehensive survey of methods and applications.ACM computing surveys, 56(4):1–39, 2023

    Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications.ACM computing surveys, 56(4):1–39, 2023

    2023

  26. [34]

    One-step diffusion with distribution matching distillation

    Tianwei Yin, Michaël Gharbi, Richard Zhang, Eli Shechtman, Frédo Durand, William T Free- man, and Taesung Park. One-step diffusion with distribution matching distillation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 13

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.