Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

A Sharp KL-Convergence Analysis for Diffusion Models under Minimal Assumptions

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Under only L2 score-error and finite second moment, diffusion sampling reaches O~(ε²) KL in O~(d/ε) steps via an ODE-plus-noise scheme.

desk verdict A genuinely new ODE-plus-noise scheme and a clever derivative-identity analysis, but an algebra error in the key lemma leaves the stated linear-d KL bound unproven as written. read the letter →

arxiv 2508.16306 v1 pith:ZMBFAZFH submitted 2025-08-22 stat.ML cs.LGmath.APmath.STstat.TH

Nishant Jain , Tong Zhang This is my paper
classification stat.MLcs.LGmath.APmath.STstat.TH MSC 68T0760H1065C40
keywords diffusionmodelsKLdivergenceprobabilityflowODEminimalassumptionsscoreestimationerrordiscretizationanalysisdimensiondependencegenerativemodeling
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

The paper gives a refined convergence analysis for diffusion-model sampling under minimal assumptions: only an $L^2$ bound on the score-estimation error and a finite second moment for the data distribution. It claims that a sampler alternating one reverse probability-flow ODE step with one small forward noising step achieves $\varepsilon^2$-KL error in about $d \left(\log(1/\delta)\right)^{3/2}/\varepsilon$ steps, improving the previous best $d \log^2(1/\delta)/\varepsilon^2$ steps for the same metric. Because KL divergence dominates the square of TV distance, the result also strengthens the best known TV guarantee. The improvement comes from using the ODE step to control a Wasserstein-type error and then converting it into KL via the added noise, while new derivative identities keep the dimension dependence linear.

What carries the argument

The carrying mechanism is a two-step update per interval: an exponential-integrator reverse probability-flow ODE step (Algorithm 1, step 4) followed by a smaller forward noising step (step 6). The ODE step lets the per-interval error be controlled as a Wasserstein-type distance, which gives better step-size dependence; the subsequent noise step converts that Wasserstein error into KL via the closed-form Gaussian conditional KL formula of Lemma A.1. To make this work without smoothness assumptions, the proof introduces derivative identities for the rescaled score $s_r(t, z)$, where $z(t) = e^t x(t)$. Lemma A.11 relates $\frac{d}{dt} \mathbb{E}\|s_r\|^2$ to $-2e^{2t} \mathbb{E}\|\nabla s_r\|_F^2$; Lemma A.12 generalizes this to powers

What would settle it

For a known $d$-dimensional target (e.g., a Gaussian mixture), compute both sides of the identity $\frac{d}{dt} \mathbb{E}_{q_t}\|s_r\|^2 = -2e^{2t} \mathbb{E}_{q_t}\|\nabla s_r\|_F^2$ from Lemma A.11, and evaluate the Lemma A.15 bound on $\int \Delta q_t \|\nabla s_r\|_F^2 dz$. If the constant $C_d = \frac{(1+2\log(d)/d+6/d)^{\log d + 3}}{1+\log d}$ exceeds 12 for $d \ge 10$, or if either identity fails beyond numerical error, the linear-d discretization bound no longer follows.

Watch

Extended reading notes

Core claim

The central claim is Theorem 3.1 and Corollary 3.2: under Assumption 2.1 (time-averaged $L^2$ score-estimation error bounded by $\varepsilon_{\text{score}}^2$) and Assumption 2.2 (finite second moment $m_2$ of the data), Algorithm 1 achieves $KL(p_{t_1} \| \hat{p}_{t_1}) \lesssim (d + m_2)e^{-T} + d^2 c^3 K + T \varepsilon_{\text{score}}^2$. Choosing $T = \log\left((d+m_2)/\varepsilon_{\text{score}}\right)$, step-size multiplier $c = \Theta\left((\log(1/\delta)+T)/K\right)$, and iteration count $K = \Theta\left(d \log^{3/2}(1/\delta)/\varepsilon_{\text{score}}\right)$ yields KL error $\widetilde{O}(\varepsilon_{\text{score}}^2)$. This replaces the previous best iteration complexity $\widetilde{O}(d \log^2(1/\delta)/\varepsilon_{\text{score}}^2)$ with $\widetilde{O}(d \log^{3/2}(1/\delta)/\varepsilon_{\text{score}})$, and by Pinsker's inequality implies a TV bound that improves the prior $O(d/\varepsilon)$ TV result.

Load-bearing premise

The argument rests on the new score-derivative identities holding with constants that stay bounded as the dimension grows; if those constants grow with $d$, the claimed linear-d discretization error degrades to $d^{3/2}$ or worse.

Editorial extensions

If this is right

  • The same score-estimation budget now buys ε² KL error in about d(log(1/δ))^{3/2}/ε iterations, replacing the previous d log²(1/δ)/ε².
  • By Pinsker's inequality, the KL guarantee implies TV error O(ε), so the result dominates the previous TV-only complexity under the same minimal assumptions.
  • The guarantee holds for the early-stopped, δ-perturbed target distribution, with no smoothness or log-concavity assumptions on the data.
  • The ODE-plus-noising decomposition offers a new way to analyze diffusion samplers: as a Wasserstein-to-KL conversion per interval, rather than as a pure discretization of the reverse SDE.

Reading between the lines

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

  • The derivative identities in Lemmas A.11–A.17 may be reusable beyond this specific sampler: any OU-flow score-based scheme that alternates deterministic transport with small noise could convert its Wasserstein error into KL with the same improved step-size exponent.
  • The sharpness of the linear-d bound is tied to the dimension-uniform constant C_d in Lemma A.15; computing C_d numerically for d in the hundreds would show whether the d² term is truly irreducible or can be improved further.
  • The scheme resembles a one-step predictor-corrector, so a small forward noising step may also smooth the estimated score at early times; the paper does not explore this additional robustness, but it is a testable consequence of the construction.
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

4 major / 5 minor

Summary. The paper proposes Algorithm 1, a hybrid sampler that alternates a reverse Probability-Flow ODE step with a smaller forward noising step. Under Assumptions 2.1 and 2.2 (L2 score-estimation error and finite second moment), Theorem 3.1 claims KL(p_{t1} || \hat{p}_{t1}) ≲ (d+m_2)e^{-T} + d^2 c^3 K + T ε_score^2, and Corollary 3.2 derives an O~(d/ε) iteration complexity for ε^2 KL error, improving over the previous O~(d/ε^2). The proof introduces an auxiliary ideal process (exact PF ODE followed by Gaussian noising) whose marginals coincide with the forward OU process, then uses a chain-rule/DPI decomposition to compare the empirical sampler with this ideal process. The technical core is a set of derivative identities (Lemmas A.11–A.16) that convert the O(d²) Laplacian/Jacobian terms in the discretization error into O(d) score-moment derivatives, enabling the linear d-dependence.

Significance. If the main result is correct, it is a meaningful improvement: it is the first KL guarantee under minimal assumptions with O(1/ε) rather than O(1/ε²) dependence on the accuracy parameter, and the proof strategy of combining a reverse ODE step with a small forward noising step is elegant and potentially reusable. The paper is mostly self-contained, works under standard score-estimation assumptions, and contains no fitted parameters. The derivative-identity framework for the non-smooth ODE analysis is a genuine technical contribution. However, the central discretization bound currently contains a concrete coefficient error that invalidates the proof as written; the claim is plausible and likely repairable, but the manuscript in its present form does not rigorously establish the advertised complexity.

major comments (4)
  1. [A.4.3, Lemma A.17] The application of Lemma A.13 with m=4 is off by a factor of 2. Lemma A.13 gives X'_4 ≤ -(1/2)e^{-2t} d/dt E[||s||^4]. Multiplying by the 5/4 e^{4t} factor in the proof yields -(5/8)e^{2t} d/dt E[||s||^4], not -(5/16)e^{2t} d/dt E[||s||^4] as written. With the stated Lemma A.16 coefficient, the total coefficient on e^{2t} d/dt E[||s||^4] becomes -(125/104), not -(185/208). Since d/dt E[||s||^4] is negative, this term is positive and replacing it by -e^{2t} d/dt E[||s||^4] is not justified by '185/208 < 1'. The final bound of Lemma A.17, which is used to obtain the d^2 c^3 K term, is therefore not established as written. The authors should either correct the coefficient and verify that the subsequent aggregation in Lemma A.18 still yields the same d^2 c^3 K bound, or choose a different Young's inequality so that the total coefficient is indeed at most 1.
  2. [A.4.2, Lemma A.16 proof] In the derivation after defining a_i and b_i, the displayed inequality leads to +5∑ b_i^2, not +6∑ b_i^2. The step '+6b_i^2' is a valid loosening since 6T2 ≥ 5T2, so the resulting upper bound is still valid, but the proof should say so explicitly. More importantly, the subsequent use of Lemma A.13 for the 3/2 E||∇||s||^2||^2 term and the final coefficient -6/13 should be checked with this loosening; as written, the reader cannot tell whether the constants are deliberate or a typo. This is not independent of the previous comment, since the two errors interact in the final Lemma A.17 coefficient.
  3. [A.5, Corollary 3.2] The corollary states K = Θ(d log^{3/2}(1/δ)/ε), but the proof gives c ≲ (log(1/δ)+T)/K and hence d^2 c^3 K = d^2 (log(1/δ)+T)^3 / K^2. Since T = log((d+m2)/ε), the correct expression is K = Θ(d (log(1/δ)+T)^{3/2}/ε). This reduces to the displayed form only under an O-tilde convention that absorbs the T factor; the Θ statement as written is misleading. Please clarify whether the corollary is intended as an O-tilde statement or correct the exact dependence.
  4. [A.5, Theorem 3.1 proof, initialization error] The proof states KL(p_{t_{K+1}} || \hat{p}_{t_{K+1}}) ≤ (d+m_2)e^{-T} citing [4]. This is standard for the OU process, but the exact form (with m_2 and d) should be verified for the chosen T ≥ 1 and the specific discretization sequence; as written, the constant in '≲' is left implicit. This is a minor point, but since the final bound is stated with explicit constants, make the initialization estimate precise.
minor comments (5)
  1. [Abstract / Corollary 3.2] The abstract says O~(d log^{3/2}(1/δ)/ε), while Corollary 3.2 writes d(log(1/δ))^{3/2}/ε without the T term. Align the notation and state clearly whether the bound is up to logarithmic factors in d and ε.
  2. [Lemma A.1] The lemma defines p_{t_{k-1}|t_k} as the conditional of 'two steps of ODE and one step of noise similar to our algorithm', whereas Lemma A.2 uses p_{t_{k-1}|t_k} for the forward OU conditional. The proof of Theorem 3.1 depends on comparing the ideal PF-ODE+noise process to the empirical sampler and then using DPI; please spell out this two-step argument explicitly so the reader does not mistake Lemma A.1 for a statement about the forward OU transition.
  3. [A.4.3, Lemma A.17 proof] The inequality '2a·b ≤ ||a||^2/4 + 4||b||^2' is used with e^{4t} factors; the resulting 5/4 coefficient on the second term is correct, but the algebra is terse. Add a line showing ||∇s_r^T s_r||^2 = (1/4)||∇||s_r||^2||^2 to avoid confusion.
  4. [A.4.2, Lemma A.15] The bound C_d ≤ 12 for d ≥ 10 is stated without proof. This is easy to verify, but please include the short argument or a reference, since the uniform O(1) nature of C_d is used in the final d^2 c^3 K bound.
  5. [Throughout] There are several typos and notation inconsistencies, e.g., 'ptk−1,t' in the preliminaries, '˜O( d ε )' in the abstract region, and the use of both 'E' and 'E_q' for expectations. A careful proofreading pass is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the proof is a self-contained analysis under standard score-error and second-moment assumptions.

full rationale

The derivation chain is self-contained. Theorem 3.1 is proved from Assumptions 2.1 and 2.2 without fitting any parameter to the target quantity. The proof decomposes KL by the chain rule and data-processing inequality (Lemma A.2), converts each interval's Wasserstein-type error to KL via an exact Gaussian KL identity (Lemma A.1), separates score-estimation error from discretization error (Lemma A.3), and then bounds the discretization error through the rescaled process and the new derivative identities in Lemmas A.4-A.18. The final aggregation uses the explicit step-size choice h_k = c min{1,t_k} and the elementary monotonicity Lemma A.19. The only self-citation is [15], credited for the motivating idea of adding forward noise after an ODE step; the actual KL conversion is proven in Lemma A.1 rather than imported, so that citation is not load-bearing. The reader-flagged algebra issue in Lemma A.16 is a potential correctness/bug concern, not circularity: even if a constant is wrong, the argument does not assume its conclusion. Hence there is no circular step and the score is 0.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The argument relies only on the score-estimation assumption, finite second moment, and standard analytic inequalities. No fitted constants or new entities are introduced; the step-size parameter c is chosen from K, not fitted to data.

assumptions (3)
  • domain assumption L2 score estimation error (Assumption 2.1)
    The convergence guarantee is stated in terms of epsilon_score, the average L2 error between the estimated and true scores at the discretization points. This is a standard training-time assumption for score-based generative models.
  • domain assumption Finite second moment of data (Assumption 2.2)
    E||x0||^2 = m^2 < infinity is used to bound the initialization error (d+m^2)e^{-T}.
  • standard math Standard Gaussian moment bounds and Fokker-Planck identities
    Used throughout the appendix (Lemmas A.7, A.9, A.11-A.14) for bounding score moments and relating time and spatial derivatives.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Sharp KL-Convergence Analysis for Diffusion Models under Minimal Assumptions." pith.science (2026). https://pith.science/paper/ZMBFAZFH

@misc{pith2026250816306,
  author       = {Pith},
  title        = {Pith review of: A Sharp KL-Convergence Analysis for Diffusion Models under Minimal Assumptions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZMBFAZFH}},
  note         = {Machine review of arXiv:2508.16306}
}
abstract

Diffusion-based generative models have emerged as highly effective methods for synthesizing high-quality samples. Recent works have focused on analyzing the convergence of their generation process with minimal assumptions, either through reverse SDEs or Probability Flow ODEs. The best known guarantees, without any smoothness assumptions, for the KL divergence so far achieve a linear dependence on the data dimension $d$ and an inverse quadratic dependence on $\varepsilon$. In this work, we present a refined analysis that improves the dependence on $\varepsilon$. We model the generation process as a composition of two steps: a reverse ODE step, followed by a smaller noising step along the forward process. This design leverages the fact that the ODE step enables control in Wasserstein-type error, which can then be converted into a KL divergence bound via noise addition, leading to a better dependence on the discretization step size. We further provide a novel analysis to achieve the linear $d$-dependence for the error due to discretizing this Probability Flow ODE in absence of any smoothness assumptions. We show that $\tilde{O}\left(\tfrac{d\log^{3/2}(\frac{1}{\delta})}{\varepsilon}\right)$ steps suffice to approximate the target distribution corrupted with Gaussian noise of variance $\delta$ within $O(\varepsilon^2)$ in KL divergence, improving upon the previous best result, requiring $\tilde{O}\left(\tfrac{d\log^2(\frac{1}{\delta})}{\varepsilon^2}\right)$ steps.

Figures

Figures reproduced from arXiv: 2508.16306 by the authors.

Figure 1
Figure 1. Demonstrating the two updates: (a) along the generation process using [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗

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.

  1. The Value of Covariance Matching in Gaussian DDPMs and the Lanczos Sampler

    cs.LG 2026-05 unverdicted novelty 6.0 of 10

    Full covariance matching in Gaussian DDPMs yields O(1/T^2) path KL error and is enabled by the training-free Lanczos Gaussian sampler using Jacobian-vector products.

Reference graph

Works this paper leans on

32 extracted references · 17 canonical work pages · cited by 1 Pith paper

  1. [15]

    Multi-Step Consistency Models: Fast Generation with Theoretical Guarantees

    Nishant Jain, Xunpeng Huang, Yian Ma, and Tong Zhang. Multi-step consistency models: Fast generation with theoretical guarantees. arXiv preprint arXiv:2505.01049 , 2025

    work page Pith review arXiv 2025

  2. [4]

    Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions

    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. InInternational Conference on Learning Representations, 2023

    2023

  3. [1]

    Nearlyd-linear convergence bounds for diffusion models via stochastic localization.arXiv preprint arXiv:2308.03686 , 2023

    Joe Benton, Valentin De Bortoli, Arnaud Doucet, and George Deligiannidis. Nearlyd-linear convergence bounds for diffusion models via stochastic localization.arXiv preprint arXiv:2308.03686 , 2023

    arXiv 2023

  4. [2]

    Improved analysis of score-based generative modeling: User-friendly bounds under minimal smoothness assumptions

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

    work page 2023

  5. [3]

    The probability flow ode is provably fast

    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 , 36:68552–68575, 2023

    work page 2023

  6. [5]

    Control-a-video: Controllable text-to-video generation with diffusion models.arXiv e-prints, pages arXiv–2305, 2023

    Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models.arXiv e-prints, pages arXiv–2305, 2023

    2023

  7. [6]

    Improved analysis for a proximal algorithm for sampling

    Yongxin Chen, Sinho Chewi, Adil Salim, and Andre Wibisono. Improved analysis for a proximal algorithm for sampling. In Conference on Learning Theory, pages 2984–3014. PMLR, 2022

    work page 2022

  8. [7]

    Diffusion models in vision: A survey

    Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence , 45(9):10850–10869, 2023. 9

    work page 2023

Show all 32 references
  1. [8]

    Diffusion self-guidance for controllable image generation

    Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems , 36:16222–16239, 2023

    2023

  2. [9]

    Convergence analysis for general probability flow odes of diffusion models in wasserstein distances

    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

  3. [10]

    Protein design with guided discrete diffusion.Advances in neural information processing systems , 36:12489–12517, 2023

    Nate Gruver, Samuel Stanton, Nathan Frey, Tim GJ Rudner, Isidro Hotzel, Julien Lafrance-Vanasse, Arvind Rajpal, Kyunghyun Cho, and Andrew G Wilson. Protein design with guided discrete diffusion.Advances in neural information processing systems , 36:12489–12517, 2023

    2023

  4. [11]

    Diffusion models in bioinformatics and computational biology.Nature reviews bioengineering, 2(2):136–154, 2024

    Zhiye Guo, Jian Liu, Yanli Wang, Mengrui Chen, Duolin Wang, Dong Xu, and Jianlin Cheng. Diffusion models in bioinformatics and computational biology.Nature reviews bioengineering, 2(2):136–154, 2024

    2024

  5. [12]

    Denoising diffusion probabilistic models

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

    2020

  6. [13]

    Convergence rates for non-log-concave sampling and log-partition estimation

    David Holzmüller and Francis Bach. Convergence rates for non-log-concave sampling and log-partition estimation. arXiv preprint arXiv:2303.03237 , 2023

    2023 arXiv

  7. [14]

    Convergence analysis of probability flow ode for score-based generative models

    Daniel Zhengyu Huang, Jiaoyang Huang, and Zhengjiang Lin. Convergence analysis of probability flow ode for score-based generative models. IEEE Transactions on Information Theory , 2025

    2025

  8. [16]

    Convergence for score-based generative modeling with polynomial complexity

    Holden Lee, Jianfeng Lu, and Yixin Tan. Convergence for score-based generative modeling with polynomial complexity. arXiv preprint arXiv:2206.06227 , 2022

    2022 arXiv

  9. [17]

    Towards a mathematical theory for consistency training in diffusion models

    Gen Li, Zhihan Huang, and Yuting Wei. Towards a mathematical theory for consistency training in diffusion models. arXiv preprint arXiv:2402.07802 , 2024

    2024 arXiv

  10. [18]

    Towards non-asymptotic convergence for diffusion-based generative models

    Gen Li, Yuting Wei, Yuxin Chen, and Yuejie Chi. Towards non-asymptotic convergence for diffusion-based generative models. In The Twelfth International Conference on Learning Representations , 2023

    2023

  11. [19]

    A sharp convergence theory for the probability flow odes of diffusion models

    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

  12. [20]

    arXiv preprint arXiv:2409.18959 , 2024

    Gen Li and Yuling Yan.o(d/t) convergence theory for diffusion probabilistic models under minimal assumptions. arXiv preprint arXiv:2409.18959 , 2024

    2024 arXiv

  13. [21]

    Unified convergence analysis for score-based diffusion models with deterministic samplers

    Runjia Li, Qiwei Di, and Quanquan Gu. Unified convergence analysis for score-based diffusion models with deterministic samplers. arXiv preprint arXiv:2410.14237 , 2024

    2024 arXiv

  14. [22]

    Diffusion-lm improves controllable text generation

    Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. Advances in neural information processing systems , 35:4328–4343, 2022

    2022

  15. [23]

    Audioldm: Text-to-audio generation with latent diffusion models.arXiv preprint arXiv:2301.12503 , 2023

    Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. Audioldm: Text-to-audio generation with latent diffusion models.arXiv preprint arXiv:2301.12503 , 2023

    2023 arXiv

  16. [24]

    Repaint: Inpainting using denoising diffusion probabilistic models

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461–11471, 2022

    2022

  17. [25]

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 , 2021

    2021 arXiv

  18. [26]

    Denoising diffusion implicit models

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

    2010 arXiv

  19. [27]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2310.02279, 2023. 10

    2023 arXiv

  20. [28]

    Maximum likelihood training of score-based diffusion models

    Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. Advances in neural information processing systems , 34:1415–1428, 2021

    2021

  21. [29]

    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

  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. InInternational Conference on Learning Representations, 2020

    2020

  23. [31]

    Provably robust score-based diffusion posterior sampling for plug-and-play image reconstruction

    Xingyu Xu and Yuejie Chi. Provably robust score-based diffusion posterior sampling for plug-and-play image reconstruction. arXiv preprint arXiv:2403.17042 , 2024. 11 Contents 1 Introduction 1 1.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ...

    2024 arXiv

  24. [32]

    Z tk−2 tk dt Z t tk e2us′ r(u, z(u))du 2 2 # ≤ E

    Squaring both sides and using2ab ≤ a2 + b2, we will have: E[∥xk−0.5 − ˆxk−0.5∥2 2] ≤ 2E[∥xk−0.5 − ˜xk−0.5∥2 2] + 2E[∥˜xk−0.5 − ˆxk−0.5∥2 2] Bounding Test. Now, utilizing the Eq. 7, Eq. 9, we have: E[∥˜xk−0.5 − ˆxk−0.5∥2 2] = E∥(ehk+hk−1 − 1) (s(tk, xk) − ˆs(tk, xk)) ∥2 2 = (eh...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.