Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

Advancing Wasserstein Convergence Analysis of Score-Based Models: Insights from Discretization and Second-Order Acceleration

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Using Hessian information through a local-linearization update, this paper claims a score-based diffusion sampler that reaches Wasserstein-2 accuracy ε in O~(1/ε) iterations instead of O~(1/ε²).

desk verdict Real but conditional advance: the O~(1/ε) Wasserstein rate for a Hessian-based sampler is new, but the proofs as stated require m0 > 1/2 and one step in Proposition 17 is under-justified. read the letter →

arxiv 2502.04849 v1 pith:ZRFBHAFF submitted 2025-02-07 stat.ML cs.LGmath.PR

Yifeng Yu , Lu Yu This is my paper · ORCID
classification stat.MLcs.LGmath.PR MSC 60H1065C3060J60
keywords score-baseddiffusionmodelsWasserstein-2convergencediscretizationanalysisexponentialintegratorrandomizedmidpointmethodlocallinearizationHessianaccelerationstronglog-concavity
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 analyzes score-based diffusion models in Wasserstein-2 distance when the data distribution is smooth and strongly log-concave. It proves that Euler, exponential-integrator, randomized-midpoint, and hybrid exponential-midpoint discretizations all need $\widetilde{\mathcal{O}}(1/\varepsilon^2)$ iterations to reach accuracy $\varepsilon$, differing only in constants. The main claim is a new accelerated sampler that uses Hessian information via the local linearization method: with accurate estimates of the score and of spatial and temporal derivative terms, it achieves $W_2(\mathcal{L}(\vartheta_N^{\mathrm{SO}}), p_0) \lesssim e^{-m_{\min}T}\|X_0\|_{L_2} + O(h) + O(\varepsilon_{\mathrm{sc}} + h^{1/2}\varepsilon_{\mathrm{sc}}^{(L)} + h\,\varepsilon_{\mathrm{sc}}^{(M)})$, so $N = \widetilde{\mathcal{O}}(1/\varepsilon)$ steps with step size $h = O(\varepsilon)$ suffice. A sympathetic reader should care because, if correct, it is the first Wasserstein-2 analysis of an accelerated SDE-based diffusion sampler and gives a quantitative reason to invest in second-order score information.

What carries the argument

Key machinery is the local linearization method [Sho98] applied to the reverse SDE. On each interval $[nh,(n+1)h]$, the drift is approximated as $\gamma(T-t,x_t)\approx \gamma(T-nh,\vartheta_n^{\mathrm{SO}}) + L_n(x_t-\vartheta_n^{\mathrm{SO}}) + M_n(t-nh)$, with $L_n=\tfrac12 I_d+\nabla^2\log p_{T-nh}(\vartheta_n^{\mathrm{SO}})$ and $M_n$ built from the spatial second derivatives and the temporal derivative of the score. The resulting linear SDE is integrated analytically, producing an update with matrix exponentials $e^{s_n h}$ and a Gaussian stochastic integral. The Hessian in $L_n$ is what turns the local drift into a contracting linear system, making the one-step error $O(h^2)$ so that after $N=T/h$ steps the accumulated discretization error is $O(h)$ instead of $O(\sqrt{dh})$.

What would settle it

For an isotropic Gaussian target $N(0,2I_d)$ (so $m_0=1/2$) and for $N(0,I_d)$ ($m_0=1$), implement the SO update using exact scores, Hessians, and derivative terms, and compare the empirical $W_2$ after $N=O((1/\varepsilon)\log(1/\varepsilon))$ steps with the bound in Theorem 6; the Gaussian transition is closed-form, so a mismatch would reveal missing terms, and the $m_0=1/2$ case directly probes the hidden $m_{\min}>1/2$ condition.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the one-step discretization error, rather than the score-matching or initialization error, is what separates first-order from accelerated diffusion sampling in $W_2$. For four standard discretizations the accumulated discretization error is $\widetilde{O}(\sqrt{dh})$, forcing $h=O(\varepsilon^2)$ and hence $N=\widetilde{O}(1/\varepsilon^2)$. The proposed second-order scheme instead linearizes the drift $\gamma(T-t,x)=\tfrac12 x+\nabla\log p_{T-t}(x)$ with an Itô expansion around each state, keeps the linear part including the Hessian of the log-density, solves the resulting linear SDE exactly on each step, and thereby cuts the accumulated discretization error to $O(h)$, allowing $h=O(\varepsilon)$ and $N=\widetilde{O}(1/\varepsilon)$. The paper states this rate as Theorem 6 and Corollary 7, with total error decomposed into the same initialization term $e^{-m_{\min}T}\|X_0\|_{L_2}$ plus $O(h)$ and score-estimation terms $\varepsilon_{\mathrm{sc}}, \varepsilon_{\mathrm{sc}}^{(L)}, \varepsilon_{\mathrm{sc}}^{(M)}$.

Load-bearing premise

The proof's load-bearing premise is that the data distribution is strongly log-concave with parameter $m_0>1/2$ (so $m_{\min}-1/2>0$); every constant divides by this difference, the stepwise contraction factor $e^{-h(m_{\min}-1/2)}$ decays only then, and the estimated Hessian matrices $s^{(L)}_*$ must stay invertible.

Editorial extensions

If this is right

  • For a fixed accuracy $\varepsilon$, the Hessian-based sampler reaches $W_2<\varepsilon$ in $N=O((1/\varepsilon)\log(1/\varepsilon))$ iterations with step size $h=O(\varepsilon)$, whereas the four first-order schemes need $N=O((1/\varepsilon^2)\log(1/\varepsilon))$ and $h=O(\varepsilon^2)$.
  • The Wasserstein error decomposes additively: initialization error $e^{-m_{\min}T}\|X_0\|_{L_2}$, discretization error $O(h)$ for the second-order scheme versus $O(\sqrt{h d})$ for the first-order schemes, and score-estimation errors; in both cases $T=O(\log(1/\varepsilon))$ suffices.
  • Among Euler, exponential integrator, randomized midpoint, and exponential-midpoint, the paper's bounds give the same $W_2$ rate; randomized midpoint does not improve the order here, but the paper notes it remains attractive because the midpoint evaluation is parallelizable.
  • The acceleration requires estimating not only the score but also the Hessian and a temporal derivative term, with errors $\varepsilon_{\mathrm{sc}}^{(L)}$ and $\varepsilon_{\mathrm{sc}}^{(M)}$ entering at orders $h^{1/2}$ and $h$; this quantifies the price of second-order information.

Reading between the lines

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

  • Editorial extension: the proof's denominator $m_{\min}-1/2$ suggests a sharp threshold: as $m_0\downarrow 1/2$ the constants blow up, so the practical gain of the second-order scheme may shrink for weakly log-concave targets; a Gaussian with variance $2$ isolates this edge.
  • Editorial extension: the local-linearization update is a template for higher-order drift approximation; a quadratic expansion would plausibly give $O(h^2)$ discretization error and $N=\widetilde{O}(1/\varepsilon^{1/2})$ at the cost of estimating third-order score derivatives, which the paper does not analyze.
  • Editorial extension: in high dimension the $d$-dependent constants (such as $3dL_F/2$) may dominate, so the proven gain may only appear when Hessian-vector products or low-rank structure make the matrix exponentials tractable; the paper's numerical test is only $d=2$.
  • Editorial extension: the randomized-midpoint schemes' parallelizability could be combined with the second-order update, since the midpoint evaluation is independent of the step's Brownian increment, yielding a parallel $O(1/\varepsilon)$ sampler; this is not claimed in the paper.
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

3 major / 5 minor

Summary. The paper studies Wasserstein-2 convergence of score-based diffusion models in the strongly log-concave setting. For the Ornstein-Uhlenbeck forward process it analyzes four discretizations of the reverse SDE—Euler-Maruyama, exponential integrator, randomized midpoint, and randomized midpoint combined with an exponential integrator—and derives W2 bounds of the form e^{-mmin T}∥X0∥ + C1√d h + C2 εsc for each. It then proposes a Hessian-based accelerated sampler using the local linearization method and claims, in Theorem 6 and Corollary 7, a W2 error bound that yields iteration complexity N = O((1/ε) log(1/ε)) with step size h = O(ε), improving on the O(1/ε²) complexity of the first-order schemes. The proofs are stochastic calculus arguments built on contraction estimates for the reverse SDE together with one-step discretization error recursions.

Significance. If the claimed results can be made fully rigorous, the paper would be a worthwhile contribution: it gives the first Wasserstein convergence analysis of a second-order accelerated score-based sampler, it provides a systematic comparison of several discretization schemes in W2, and its derivation is self-contained, with the O~(1/ε) rate derived rather than imposed. The main conceptual novelty—using Hessian information via local linearization to improve the step-size scaling from h = O(ε²) to h = O(ε)—is interesting and potentially useful for practitioners. However, the stated theorems are not valid under the assumptions as written, because every bound divides by mmin - 1/2 without assuming mmin > 1/2, and the second-order proof contains an unjustified estimate. These issues are load-bearing but appear fixable, so the paper merits major revision rather than rejection.

major comments (3)
  1. [Assumption 1; Theorems 1, 3, 4, 5, 6; Lemma 10; Eq. (A.10)] Every stated Wasserstein bound contains a denominator mmin - 1/2, and the contraction factor used throughout is e^{-h(mmin - 1/2)}, but Assumption 1 only assumes m0-strong log-concavity and defines mmin = min(1, m0). If m0 ≤ 1/2, then mmin - 1/2 ≤ 0, the denominators in Theorems 1, 3, 4, 5, and 6 are negative or undefined, and the geometric sum in Eq. (A.10), whose ratio is e^{-h(mmin - 1/2)}, is not bounded. In that regime the induction argument in Appendix A collapses. Concretely, Lemma 10 and inequality (A.2) give contraction only when the exponent ∫(m(T-s) - 1/2)ds is positive; when m(t) ≤ 1/2 the 'contraction' factor is at least 1, so the bound 1/(1 - e^{-h(mmin - 1/2)}) ≲ 1/(mmin - 1/2) is invalid. The statement of the theorems is therefore not established under the stated Assumption 1. The fix is to strengthen Assumption 1 to m0 > 1/2, or to handle the case m0 ≤ 1/2 separately with different arguments, and to state this condition explicitly before Theorems 1, 3, 4, 5, and 6.
  2. [Section 4, Assumption 6, and the definition of ϑ^SO_{n+1}] The second-order update in Section 4 requires the estimated matrix s^(L)_*(T - nh, ϑ^SO_n) to be invertible. The true matrix L_n = (1/2)I + ∇² log p_{T-nh}(ϑ^SO_n) is negative definite only when m(T-nh) > 1/2, and under Assumption 1 this is not guaranteed. Even when m0 > 1/2, Assumption 6 only asserts a bound ε^(L)_sc on the estimation error and does not require ε^(L)_sc to be smaller than the spectral gap (mmin - 1/2); without such a smallness condition, a perturbation of a negative definite matrix need not remain invertible. Thus the sampler and Theorem 6 are not well-defined under the assumptions as stated. The authors should add an explicit invertibility or small-ε^(L)_sc condition to Assumption 6 or to Theorem 6.
  3. [Proposition 17 and Appendix B (proof of Theorem 6)] In the proof of Proposition 17, after invoking Assumptions 7 and 8, the displayed estimate ∥∇² log p_{T-nh-u}(tilde Y_u) - ∇² log p_{T-nh}(tilde Y_0)∥_{L2} ≲ L_F√u drops the M2 h(1 + ∥tilde Y_u∥_{L2}) term from Assumption 8 and the L_F∥tilde Y_u - tilde Y_0∥_{L2} term from Assumption 7. These are not negligible by inspection: the subsequent Gronwall argument for term II uses exactly these quantities, and the final constant C^SO_1 in Theorem 6 depends on their contribution being of order h² in r^SO_n and order h in the summed bound. As written, the '≲ L_F√u' step is unjustified, and without a careful account of the dropped terms the h coefficient in Theorem 6 is not fully established. Please either provide the missing estimates or revise the proof to show that these terms contribute only at higher order.
minor comments (5)
  1. [Section 5] The text 'we compare the performance of the SGMs uner EM, EI, REM, REIdiscretization schemes' contains a typo: 'uner' should be 'under' and 'REI' should be followed by a space.
  2. [Proposition 16 and Section 3.4] Proposition 16 states that it assumes Assumptions 1, 2, and 4, but the exponential-integrator randomized-midpoint scheme is governed by Assumption 5, not Assumption 4. The theorem statement should be corrected to reference Assumption 5.
  3. [Sections 2 and 3.3] The randomized midpoint update is defined with inconsistent noise scales: Section 2 writes √(hU_n)ξ'_n in the intermediate step, while Section 3.3 writes √(2hU_n)ξ'_n. The proof appears to use √(hU_n). Please make the definitions consistent.
  4. [Introduction] The reference list contains a duplicated entry '[GNZ23a, GNZ23a]' in the introductory overview; one occurrence should be removed or relabeled.
  5. [Section 5] The numerical experiments only report the Wasserstein distance along the first coordinate for d = 2 and do not provide empirical iteration-count comparisons. The sentence 'These numerical results support our theoretical findings' is therefore stronger than what the figures demonstrate; a more cautious wording would be appropriate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the Wasserstein bounds are derived from stated assumptions in a self-contained stochastic-calculus argument, with no fitted input called a prediction and no load-bearing self-citation.

full rationale

The paper's central results (Theorems 1, 3, 4, 5, and 6 with Corollaries 2 and 7) are obtained by direct contraction and one-step-discretization arguments. The Wasserstein bounds depend on m0, L0, d, h, and score/Hessian estimation errors as explicit assumptions, and no parameter is fitted to a subset of the data and then renamed as a prediction. The cited lemmas from [GNZ23b] (Lemmas 8 and 9) are external results about strong log-concavity propagation for the Ornstein-Uhlenbeck forward process, not results from this paper's authors, and they are used as ingredients rather than as a substitute for the main proof. The local-linearization sampler in Section 4 is analyzed by comparing the discretized process ϑ^SO_n to the continuous-time local-linearization process and then to the true reverse process; the final bound contains the estimation errors ε_sc, ε_sc^(L), and ε_sc^(M) as assumed inputs, which is the standard structure of an error decomposition rather than a circular use of the conclusion. The cited consistency of the O~(1/ε) rate with [LJLS24] and [HHL24] is presented as a comparison, not as the derivation of the rate. The only notable concern is that several constants contain denominators mmin − 1/2, while Assumption 1 only assumes m0-strong log-concavity; for m0 ≤ 1/2 the contraction factor e^{-h(mmin-1/2)} is not contracting and the stated constants are not finite. However, this is a gap between the stated assumptions and the proof's implicit requirement (m0 > 1/2), not a circularity: the result is not being assumed as an input, and it could be repaired by strengthening the assumption. Accordingly, no step in the derivation reduces to its own inputs by construction, and the derivation is self-contained against external benchmarks.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

The central claims rest on strong log-concavity, oracle score/Hessian estimates, and regularity conditions. No parameters are fitted to data in the theoretical results. The main gap is the unstated m0 > 1/2 requirement, which the paper does not list among its assumptions.

assumptions (6)
  • domain assumption p0 is m0-strongly log-concave and ∇log p0 is L0-Lipschitz (Assumption 1).
    This gives m(t) and L(t) bounds used throughout all proofs.
  • domain assumption The score time regularity bound in Assumption 2 holds with constant M1.
    Used to bound differences of the score across time in one-step discretization errors.
  • domain assumption Score estimator is accurate at discrete points with error εsc (Assumptions 3, 4, 5).
    Oracle assumption; the bounds scale linearly in εsc.
  • domain assumption Hessian and time-derivative estimates satisfy Assumption 6 with errors ε_sc^(L) and ε_sc^(M).
    Required for the second-order sampler; no construction or sample complexity is given.
  • domain assumption The Hessian of the log-density is LF-Lipschitz (Assumption 7) and satisfies time regularity (Assumption 8).
    Used to control the local linearization error in Appendix B.
  • ad hoc to paper The strong log-concavity constant satisfies m0 > 1/2, equivalently mmin > 1/2.
    Not stated in Assumption 1 but required for positive denominators mmin - 1/2 and decaying contraction factors in Theorems 1, 3, 4, 5, and 6; also guarantees negative-definiteness and invertibility of L_n = 1/2 I + Hessian.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Advancing Wasserstein Convergence Analysis of Score-Based Models: Insights from Discretization and Second-Order Acceleration." pith.science (2026). https://pith.science/paper/ZRFBHAFF

@misc{pith2026250204849,
  author       = {Pith},
  title        = {Pith review of: Advancing Wasserstein Convergence Analysis of Score-Based Models: Insights from Discretization and Second-Order Acceleration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZRFBHAFF}},
  note         = {Machine review of arXiv:2502.04849}
}
abstract

Score-based diffusion models have emerged as powerful tools in generative modeling, yet their theoretical foundations remain underexplored. In this work, we focus on the Wasserstein convergence analysis of score-based diffusion models. Specifically, we investigate the impact of various discretization schemes, including Euler discretization, exponential integrators, and midpoint randomization methods. Our analysis provides a quantitative comparison of these discrete approximations, emphasizing their influence on convergence behavior. Furthermore, we explore scenarios where Hessian information is available and propose an accelerated sampler based on the local linearization method. We demonstrate that this Hessian-based approach achieves faster convergence rates of order $\widetilde{\mathcal{O}}\left(\frac{1}{\varepsilon}\right)$ significantly improving upon the standard rate $\widetilde{\mathcal{O}}\left(\frac{1}{\varepsilon^2}\right)$ of vanilla diffusion models, where $\varepsilon$ denotes the target accuracy.

Figures

Figures reproduced from arXiv: 2502.04849 by the authors.

Figure 1
Figure 1. Error of various discretization schemes and second-order sampler with different choice of [PITH_FULL_IMAGE:figures/full_fig_p014_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Assessing the Quality of Denoising Diffusion Models in Wasserstein Distance: Noisy Score and Optimal Bounds

    stat.ML 2025-06 conditional novelty 7.0 of 10

    Denoising diffusion models achieve Wasserstein-2 sampling error of order √D/K up to logarithmic factors for a broad class of distributions, matching the Gaussian lower bound, and score-evaluation noise vanishes as the...

  2. Fast Convergence for High-Order ODE Solvers in Diffusion Probabilistic Models

    cs.LG 2025-06 conditional novelty 6.0 of 10

    A TV convergence bound O(d^{7/4} ε^{1/2} + d(dH)^p) is proved for p-th order (exponential) Runge-Kutta samplers of probability-flow ODEs under C² smoothness of the learned score.

Reference graph

Works this paper leans on

47 extracted references · 14 canonical work pages · cited by 2 Pith papers

  1. [1]

    Namrata Anand and Tudor Achim, Protein structure and sequence generation with equivariant denoising diffusion probabilistic models, arXiv preprint arXiv:2205.15019 (2022)

    arXiv 2022

  2. [2]

    3, 313--326

    Brian DO Anderson, Reverse-time diffusion equation models, Stochastic Processes and their Applications 12 (1982), no. 3, 313--326

    1982

  3. [3]

    Stefano Bruno, Ying Zhang, Dong-Young Lim, \"O mer Deniz Akyildiz, and Sotirios Sabanis, On diffusion-based generative models and their error bounds: The log-concave case with full convergence estimates, arXiv preprint arXiv:2311.13584 (2023)

    arXiv 2023

  4. [4]

    300--323

    Xiang Cheng, Niladri S Chatterji, Peter L Bartlett, and Michael I Jordan, Underdamped langevin mcmc: A non-asymptotic analysis, Conference on learning theory, PMLR, 2018, pp. 300--323

    work page 2018

  5. [5]

    59, Institut Henri Poincar \'e , 2023, pp

    Patrick Cattiaux, Giovanni Conforti, Ivan Gentil, and Christian L \'e onard, Time reversal of diffusion processes under a finite entropy condition, Annales de l'Institut Henri Poincar \'e (B) Probabilit \'e s et Statistiques, vol. 59, Institut Henri Poincar \'e , 2023, pp. 1844--1881

    2023

  6. [6]

    Sitan Chen, Sinho Chewi, Jerry Li, Yuanzhi Li, Adil Salim, and Anru R Zhang, Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions, arXiv preprint arXiv:2209.11215 (2022)

    arXiv 2022

  7. [7]

    Sitan Chen, Sinho Chewi, Holden Lee, Yuanzhi Li, Jianfeng Lu, and Adil Salim, The probability flow ode is provably fast, Advances in Neural Information Processing Systems 36 (2024)

    work page 2024

  8. [8]

    4735--4763

    Hongrui Chen, Holden Lee, and Jianfeng Lu, Improved analysis of score-based generative modeling: User-friendly bounds under minimal smoothness assumptions, International Conference on Machine Learning, PMLR, 2023, pp. 4735--4763

    2023

Show all 47 references
  1. [9]

    Minshuo Chen, Song Mei, Jianqing Fan, and Mengdi Wang, An overview of diffusion models: Applications, guided generation, statistical rates and optimization, arXiv preprint arXiv:2404.07771 (2024)

    2024 arXiv

  2. [10]

    Valentin De Bortoli, Convergence of denoising diffusion models under the manifold hypothesis, arXiv preprint arXiv:2208.05314 (2022)

    2022 arXiv

  3. [11]

    Prafulla Dhariwal and Alexander Nichol, Diffusion models beat gans on image synthesis, Advances in neural information processing systems 34 (2021), 8780--8794

    2021

  4. [12]

    Shivam Gupta, Linda Cai, and Sitan Chen, Faster diffusion-based sampling with randomized midpoints: Sequential and parallel, arXiv preprint arXiv:2406.00924 (2024)

    2024 arXiv

  5. [14]

    , Wasserstein convergence guarantees for a general class of score-based generative models, arXiv preprint arXiv:2311.11003 (2023)

    2023 arXiv

  6. [15]

    Marta Gentiloni-Silveri and Antonio Ocello, Beyond log-concavity and score regularity: Improved convergence bounds for score-based generative models in w2-distance, arXiv preprint arXiv:2501.02298 (2025)

    2025

  7. [16]

    Xuefeng Gao and Lingjiong Zhu, Convergence analysis for general probability flow odes of diffusion models in wasserstein distances, arXiv preprint arXiv:2401.17958 (2024)

    2024 arXiv

  8. [17]

    Ye He, Krishnakumar Balasubramanian, and Murat A Erdogdu, On the ergodicity, bias and asymptotic normality of randomized midpoint sampling method, Advances in Neural Information Processing Systems 33 (2020), 7366--7376

    2020

  9. [18]

    Aapo Hyv \"a rinen and Peter Dayan, Estimation of non-normalized statistical models by score matching., Journal of Machine Learning Research 6 (2005), no. 4

    2005

  10. [19]

    Daniel Zhengyu Huang, Jiaoyang Huang, and Zhengjiang Lin, Convergence analysis of probability flow ode for score-based generative models, arXiv preprint arXiv:2404.09730 (2024)

    2024 arXiv

  11. [20]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel, Denoising diffusion probabilistic models, Advances in neural information processing systems 33 (2020), 6840--6851

    2020

  12. [21]

    Marlis Hochbruck and Alexander Ostermann, Exponential integrators, Acta Numerica 19 (2010), 209--286

    2010

  13. [22]

    Saravanan Kandasamy and Dheeraj Nagaraj, The poisson midpoint method for langevin dynamics: Provably efficient discretization for diffusion models, arXiv preprint arXiv:2405.17068 (2024)

    2024

  14. [23]

    Gen Li and Changxiao Cai, Provable acceleration for diffusion models under minimal assumptions, arXiv preprint arXiv:2410.23285 (2024)

    2024 arXiv

  15. [24]

    Jiadong Liang, Zhihan Huang, and Yuxin Chen, Low-dimensional adaptation of diffusion models: Convergence in total variation, arXiv preprint arXiv:2501.12982 (2025)

    2025 arXiv

  16. [25]

    Gen Li, Yu Huang, Timofey Efimov, Yuting Wei, Yuejie Chi, and Yuxin Chen, Accelerating convergence of score-based diffusion models, provably, arXiv preprint arXiv:2403.03852 (2024)

    2024 arXiv

  17. [26]

    Gen Li and Yuchen Jiao, Improved convergence rate for diffusion probabilistic models, arXiv preprint arXiv:2410.13738 (2024)

    2024 arXiv

  18. [27]

    Y Liang, P Ju, Y Liang, and N Shroff, Broadening target distributions for accelerated diffusion models via a novel analysis approach, arXiv preprint arXiv:2402.13901 (2024)

    2024 arXiv

  19. [28]

    Gen Li, Yuting Wei, Yuejie Chi, and Yuxin Chen, A sharp convergence theory for the probability flow odes of diffusion models, arXiv preprint arXiv:2408.02320 (2024)

    2024 arXiv

  20. [29]

    Chenlin Meng, Yang Song, Wenzhe Li, and Stefano Ermon, Estimating high order gradients of the data distribution by denoising, Advances in Neural Information Processing Systems 34 (2021), 25359--25369

    2021

  21. [30]

    Nikiforos Mimikos-Stamatopoulos, Benjamin J Zhang, and Markos A Katsoulakis, Score-based generative models are provably robust: an uncertainty quantification perspective, arXiv preprint arXiv:2405.15754 (2024)

    2024 arXiv

  22. [31]

    8599--8608

    Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov, Grad-tts: A diffusion probabilistic model for text-to-speech, International Conference on Machine Learning, PMLR, 2021, pp. 8599--8608

    2021

  23. [32]

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen, Hierarchical text-conditional image generation with clip latents, arXiv preprint arXiv:2204.06125 1 (2022), no. 2, 3

    2022 arXiv

  24. [33]

    574--584

    Yang Song, Sahaj Garg, Jiaxin Shi, and Stefano Ermon, Sliced score matching: A scalable approach to density and score estimation, Uncertainty in Artificial Intelligence, PMLR, 2020, pp. 574--584

    2020

  25. [34]

    Isao Shoji, Approximation of continuous time stochastic processes by a local linearization method, Math. Comput. 67 (1998), 287--298

    1998

  26. [35]

    Ruoqi Shen and Yin Tat Lee, The randomized midpoint method for log-concave sampling, Advances in Neural Information Processing Systems 32 (2019)

    2019

  27. [36]

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

    2020 arXiv

  28. [37]

    Stanislas Strasman, Antonio Ocello, Claire Boyer, Sylvain Le Corff, and Vincent Lemaire, An analysis of the noise schedule for score-based generative models, arXiv preprint arXiv:2402.04650 (2024)

    2024 arXiv

  29. [38]

    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)

    2020 arXiv

  30. [39]

    Wenpin Tang and Hanyang Zhao, Contractive diffusion probabilistic models, arXiv preprint arXiv:2401.13115 (2024)

    2024 arXiv

  31. [40]

    , Score-based diffusion models via stochastic differential equations--a technical tutorial, arXiv preprint arXiv:2402.07487 (2024)

    2024 arXiv

  32. [41]

    7, 1661--1674

    Pascal Vincent, A connection between score matching and denoising autoencoders, Neural computation 23 (2011), no. 7, 1661--1674

    2011

  33. [42]

    Yuchen Wu, Yuxin Chen, and Yuting Wei, Stochastic runge-kutta methods: Provable acceleration of diffusion models, arXiv preprint arXiv:2410.04760 (2024)

    2024 arXiv

  34. [43]

    Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang, Geodiff: A geometric diffusion model for molecular conformation generation, arXiv preprint arXiv:2203.02923 (2022)

    2022 arXiv

  35. [44]

    Lu Yu and Arnak Dalalyan, Parallelized midpoint randomization for langevin monte carlo, arXiv preprint arXiv:2402.14434 (2024)

    2024 arXiv

  36. [45]

    Lu Yu, Avetik Karagulyan, and Arnak Dalalyan, Langevin monte carlo for strongly log-concave distributions: Randomized midpoint revisited, arXiv preprint arXiv:2306.08494 (2023)

    2023 arXiv

  37. [46]

    Lu Yu, Log-concave sampling on compact supports: A versatile proximal framework, arXiv preprint arXiv:2405.15379 (2024)

    2024 arXiv

  38. [47]

    4, 1--39

    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 (2023), no. 4, 1--39

    2023

  39. [48]

    Qinsheng Zhang and Yongxin Chen, Fast sampling of diffusion models with exponential integrator, arXiv preprint arXiv:2204.13902 (2022)

    2022 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.