Pith. sign in

REVIEW 3 major objections 3 minor 2 cited by

The paper claims that infinitesimal learning errors in a wide class of generative models change the predicted density only on the data manifold, and that this robustness is caused by alignment of the fastest-growing perturbation directions

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Inexact generative models stay on the data manifold because infinitesimal learning errors perturb the density only along the manifold, when top Lyapunov vectors align with the support boundary.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Clever Lyapunov-alignment story for support robustness, but the exact-support claim likely overreaches a first-order analysis and the supplied text is unreadable—send to referees anyway. the 3 major comments →

arxiv 2508.07581 v1 pith:RBJVYHNP submitted 2025-08-11 cs.LG math.DSmath.PR

When and how can inexact generative models still sample from the data manifold?

Nisha Chandramoorthy , Adriaan de Clercq This is my paper
classification cs.LG math.DSmath.PR
keywords support robustnessgenerative modelsLyapunov vectorsprobability flowscore-based generative modelsflow matchingperturbation analysisdata manifold
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Generative models often produce samples that look plausible, even when the learned score or drift is imperfect. This paper tries to show that this is not luck: for a broad class of dynamical generative models, infinitesimal learning errors change the predicted density only along the data manifold, not perpendicular to it. The mechanism is dynamical alignment: the most sensitive perturbation directions of the generating flow line up with the tangent spaces of the manifold's boundary. The paper proves a sufficient condition for this alignment, shows it is cheap to compute, and uses it to recover the tangent bundle of the data manifold. If correct, support robustness is a first-order structural property of how these models transport probability, and it can be certified from the model's own dynamics.

Core claim

The central claim is that, for a wide class of stochastic and deterministic generative processes, infinitesimal errors in the learned score or drift cause the predicted density to differ from the target density only on the data manifold; off the manifold, the density perturbation vanishes at first order. The dynamical mechanism is that the top Lyapunov vectors, the directions in which infinitesimal perturbations grow fastest, align with the tangent spaces along the boundary of the data manifold. The paper gives a sufficient condition on the generating dynamics for this alignment, derived through a finite-time linear perturbation analysis of both sample paths and probability flows. In practic

What carries the argument

The central object is the top Lyapunov vector field of the generating flow, i.e., the directions along which infinitesimal perturbations grow fastest. The argument shows that when these vectors align with tangent spaces along the data manifold's boundary, first-order errors in the learned score or drift transport probability along the manifold rather than away from it. A sufficient condition on the flow's linearized dynamics guarantees this alignment, and the paper emphasizes that the condition is cheap to evaluate numerically.

Load-bearing premise

The load-bearing premise is that real learning errors are effectively infinitesimal and smooth enough that a first-order, finite-time linear perturbation analysis captures the full behavior; otherwise higher-order terms could move samples off the manifold.

What would settle it

Take a smooth 1D data manifold embedded in $\mathbb{R}^2$, and let the learned score equal the exact score plus a fixed normal perturbation of amplitude $\varepsilon$. Transport a test density to the final time. If the off-manifold density difference or the normal displacement of samples contains a term linear in $\varepsilon$ as $\varepsilon\to 0$, the paper's first-order support-robustness claim is false; if the normal effect is absent, the claim is confirmed.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Support robustness is a first-order generic property: for the analyzed class, any infinitesimal score or drift error displaces the predicted density along the data manifold, not away from it.
  • The sufficient alignment condition can be computed from the flow and used to check robustness; in robust models it simultaneously provides estimates of the data manifold's tangent bundle.
  • The perturbation analysis covers both deterministic probability-flow dynamics, such as conditional flow matching, and stochastic dynamics, such as score-based generative models.
  • The result does not require the manifold hypothesis: the robustness statement holds for target distributions with or without manifold structure.
  • The analysis complements existing theoretical guarantees obtained through stochastic analysis, statistical learning, and uncertainty quantification.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This suggests using alignment of top Lyapunov vectors with estimated tangent directions as a training-time robustness diagnostic: it could detect drift into normal directions before generated samples visibly degrade.
  • If the first-order picture extends to small but finite errors, one might expect normal leakage to be controlled by second-order terms; a testable extension is that normal displacement scales quadratically in the error amplitude for smooth flows.
  • Discretization error in numerical integrators is the same kind of perturbation, so the mechanism may also explain robustness to ODE or SDE solver step-size errors.
  • A regularizer that penalizes the normal component of the top Lyapunov spectrum could directly enforce support robustness during training.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The manuscript studies the phenomenon that certain dynamical generative models continue to produce samples that lie on the data manifold even when the score function or drift vector field is learned with error. The authors present a perturbation analysis of the probability flow and state that infinitesimal learning errors change the predicted density only on the data manifold. They identify alignment of top Lyapunov vectors with tangent spaces along the data-manifold boundary as the mechanism, prove a sufficient condition for this alignment, and argue that the condition is efficient to compute and yields tangent-bundle estimation. They claim applicability to conditional flow matching and score-based generative models, with or without the manifold hypothesis. The material is presented as a finite-time linear perturbation analysis of sample paths and probability flows.

Significance. If the central claim is established rigorously for finite (not merely infinitesimal) errors, this would be a significant bridge between dynamical systems theory and generative modeling, offering a principled explanation for a widely observed but poorly understood robustness. The computational alignment condition and its use for tangent-bundle estimation are concrete contributions. However, the advertised support robustness is exactly the kind of claim that can be true to first order while false for finite perturbations, and the current abstract does not draw that distinction. The paper's value accordingly hinges on the precise theorem statements and remainder estimates.

major comments (3)
  1. [Abstract, first sentence] The claim that infinitesimal learning errors cause the predicted density to differ from the target density 'only on the data manifold' is ambiguous and potentially overclaimed. A first-order perturbation analysis of a measure supported on M automatically yields a first-order density correction supported on M, because the unperturbed density is zero off M; it does not follow that the exact perturbed distribution has support M. Under a generic C^1 O(ε) perturbation of a flow that leaves M invariant, normal hyperbolicity gives a nearby invariant manifold M_ε at distance O(ε), not M itself, so p_ε is nonzero on M_ε\M. Lyapunov alignment of the top Lyapunov vectors controls tangential instability, not the normal component of the response. The abstract appears to prove at most an asymptotic statement. Please state the theorem as a first-order statement or prove an exact invariance condition; i
  2. [Abstract, 'finite-time linear perturbation analysis'] The analysis is explicitly finite-time and linear. Learned errors are finite, not infinitesimal, and practical sampling uses fixed finite integration times. The paper does not state how the remainder terms depend on the error size ε, the time horizon, or dimension. Without such bounds, the explanation remains asymptotic and may not account for the finite-error robustness observed in practice. Please provide explicit remainder estimates and discuss whether the main theorem covers the finite-error regime.
  3. [Abstract, 'sufficient condition'] The abstract mentions a sufficient condition on the dynamics to achieve Lyapunov alignment but does not state the condition. For evaluation, the manuscript must specify the class of dynamics (ODE/SDE, regularity, hyperbolicity), the precise definition of top Lyapunov vectors in the finite-time setting, and the notion of tangent spaces along the data-manifold boundary. If the manifold has a boundary, this requires a coordinate-free formulation. Please state the main theorem with all hypotheses exposed.
minor comments (3)
  1. [Abstract] The phrase 'target distributions that may or may not satisfy the manifold hypothesis' is vague; define what it means for a distribution to satisfy the manifold hypothesis. Similarly, 'wide class of generative models' should be made precise.
  2. [General notation] The term 'top Lyapunov vectors' needs a finite-time definition, since classical Lyapunov exponents are asymptotic limits. If finite-time estimates are used, provide confidence intervals or convergence statements for the alignment measure.
  3. [Related work] The relation to existing score-based generative model guarantees (e.g., score matching error bounds, Wasserstein/divergence estimates) should be stated more explicitly, so the new dynamical-systems contribution is clearly delineated.

Circularity Check

0 steps flagged

No significant circularity: the perturbation analysis and Lyapunov-alignment mechanism are independent of the support-robustness conclusion.

full rationale

The paper's central claim, as stated in the abstract, is a derived perturbation result: infinitesimal learning errors are shown to cause the predicted density to differ from the target density only on the data manifold, with the mechanism attributed to alignment of top Lyapunov vectors with the boundary tangent spaces. This is not a self-definitional or fitted-input-called-prediction step: the Lyapunov vectors are defined from the linearized dynamics, independently of the support-robustness conclusion, and no fitted constants are used to force the result. The alignment condition is presented as a sufficient condition, not as a restatement of the conclusion. The abstract and the readable portions of the paper do not exhibit a load-bearing self-citation chain or a uniqueness theorem imported from the authors' prior work. The skeptic's concern about finite perturbations moving an invariant manifold by O(epsilon) is a substantive correctness/assumption issue, not circularity. Because the full text is heavily corrupted and no specific equation or reduction can be quoted, no circular step can be established. The derivation is therefore, on the available evidence, self-contained with respect to the circularity criteria.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

Provisional ledger based on the abstract only. No fitted constants or invented entities are described; the assumptions above are the standard regularity and small-error premises for the claimed theorem.

axioms (3)
  • domain assumption Learning errors in the score/drift are infinitesimal, so first-order perturbation analysis is valid.
    Abstract: 'infinitesimal learning errors' and 'finite-time linear perturbation analysis'; central to the density perturbation claim.
  • domain assumption The generating dynamics admit well-defined Lyapunov exponents/vectors on finite time horizons.
    Abstract: 'top Lyapunov vectors' are the mechanism; their existence and differentiability is assumed.
  • domain assumption The data manifold boundary has well-defined tangent spaces along which alignment can be measured.
    Abstract: 'alignment of the top Lyapunov vectors with the tangent spaces along the boundary of the data manifold' requires such tangent spaces to exist.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of When and how can inexact generative models still sample from the data manifold?." pith.science (2026). https://pith.science/paper/RBJVYHNP

@misc{pith2026250807581,
  author       = {Pith},
  title        = {Pith review of: When and how can inexact generative models still sample from the data manifold?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RBJVYHNP}},
  note         = {Machine review of arXiv:2508.07581}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

A curious phenomenon observed in some dynamical generative models is the following: despite learning errors in the score function or the drift vector field, the generated samples appear to shift \emph{along} the support of the data distribution but not \emph{away} from it. In this work, we investigate this phenomenon of \emph{robustness of the support} by taking a dynamical systems approach on the generating stochastic/deterministic process. Our perturbation analysis of the probability flow reveals that infinitesimal learning errors cause the predicted density to be different from the target density only on the data manifold for a wide class of generative models. Further, what is the dynamical mechanism that leads to the robustness of the support? We show that the alignment of the top Lyapunov vectors (most sensitive infinitesimal perturbation directions) with the tangent spaces along the boundary of the data manifold leads to robustness and prove a sufficient condition on the dynamics of the generating process to achieve this alignment. Moreover, the alignment condition is efficient to compute and, in practice, for robust generative models, automatically leads to accurate estimates of the tangent bundle of the data manifold. Using a finite-time linear perturbation analysis on samples paths as well as probability flows, our work complements and extends existing works on obtaining theoretical guarantees for generative models from a stochastic analysis, statistical learning and uncertainty quantification points of view. Our results apply across different dynamical generative models, such as conditional flow-matching and score-based generative models, and for different target distributions that may or may not satisfy the manifold hypothesis.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

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

  1. Provably Learning Diffusion Models under the Manifold Hypothesis: Collapse and Refine

    cs.LG 2026-05 unverdicted novelty 6.0

    SiLD is a score-matching framework that learns both manifold projection and intrinsic density from a single objective, with proven sample complexity depending only on intrinsic dimension.

  2. From Platform Migration to Cultural Integration: the Ingress and Diffusion of #wlw from TikTok to RedNote in Queer Women Communities

    cs.SI 2025-08 unverdicted novelty 5.0

    The #wlw hashtag entered RedNote through TikTok immigrants' bold use and cross-group interpretation, then became a locally recognized queer tag that also carries feminist discussion.

Reference graph

Works this paper leans on

60 extracted references · 33 canonical work pages · cited by 2 Pith papers · 1 internal anchor

  1. [1]

    Losing dimensions: Geometric memorization in generative diffusion, 2024

    Beatrice Achilli, Enrico Ventura, Gianluigi Silvestri, Bao Pham, Gabriel Raya, Dmitry Krotov, Carlo Lucibello, and Luca Ambrogioni. Losing dimensions: Geometric memorization in generative diffusion, 2024. URL https://arxiv.org/abs/2410.08727

    arXiv 2024

  2. [2]

    Building normalizing flows with stochastic interpolants

    Michael S Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. arXiv preprint arXiv:2209.15571, 2022

    Pith/arXiv arXiv 2022

  3. [3]

    Stochastic interpolants: A unifying framework for flows and diffusions

    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

    Pith/arXiv arXiv 2023

  4. [4]

    Reverse-time diffusion equation models

    Brian DO Anderson. Reverse-time diffusion equation models. Stochastic Processes and their Applications, 12 0 (3): 0 313--326, 1982

    1982

  5. [5]

    Random dynamical systems

    Ludwig Arnold, Christopher KRT Jones, Konstantin Mischaikow, Genevi \`e ve Raugel, and Ludwig Arnold. Random dynamical systems. Springer, 1995

    work page 1995

  6. [6]

    Memorization and regularization in generative diffusion models

    Ricardo Baptista, Agnimitra Dasgupta, Nikola B Kovachki, Assad Oberai, and Andrew M Stuart. Memorization and regularization in generative diffusion models. arXiv preprint arXiv:2501.15785, 2025

    Pith/arXiv arXiv 2025

  7. [7]

    Lyapunov characteristic exponents for smooth dynamical systems and for hamiltonian systems; a method for computing all of them

    Giancarlo Benettin, Luigi Galgani, Antonio Giorgilli, and Jean-Marie Strelcyn. Lyapunov characteristic exponents for smooth dynamical systems and for hamiltonian systems; a method for computing all of them. part 1: Theory. Meccanica, 15: 0 9--20, 1980

    work page 1980

  8. [8]

    Dynamical regimes of diffusion models

    Giulio Biroli, Tony Bonnaire, Valentin De Bortoli, and Marc M \'e zard. Dynamical regimes of diffusion models. Nature Communications, 15 0 (1): 0 9957, 2024

    work page 2024

  9. [9]

    Align your latents: High-resolution video synthesis with latent diffusion models

    Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22563--22575, June 2023

    work page 2023

  10. [11]

    Riemannian score-based generative modelling, 2022

    Valentin De Bortoli, Emile Mathieu, Michael Hutchinson, James Thornton, Yee Whye Teh, and Arnaud Doucet. Riemannian score-based generative modelling, 2022. URL https://arxiv.org/abs/2202.02763

    Pith/arXiv arXiv 2022

  11. [12]

    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 Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors, Proceedings of the 40th International Conference on Machine Learning, volume 202 of P...

    work page 2023

  12. [13]

    Score approximation, estimation and distribution recovery of diffusion models on low-dimensional data

    Minshuo Chen, Kaixuan Huang, Tuo Zhao, and Mengdi Wang. Score approximation, estimation and distribution recovery of diffusion models on low-dimensional data. In International Conference on Machine Learning, pages 4672--4712. PMLR, 2023 b

    2023

  13. [14]

    Neural ordinary differential equations

    Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018

    2018

  14. [15]

    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. arXiv preprint arXiv:2209.11215, 2022

    Pith/arXiv arXiv 2022

  15. [16]

    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: 0 68552--68575, 2023 c

    work page 2023

  16. [17]

    Convergence of denoising diffusion models under the manifold hypothesis

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

    Pith/arXiv arXiv 2022

  17. [18]

    Diffusion schr \"o dinger bridge with applications to score-based generative modeling

    Valentin De Bortoli, James Thornton, Jeremy Heng, and Arnaud Doucet. Diffusion schr \"o dinger bridge with applications to score-based generative modeling. Advances in Neural Information Processing Systems, 34: 0 17695--17709, 2021

    2021

  18. [19]

    The mnist database of handwritten digit images for machine learning research

    Li Deng. The mnist database of handwritten digit images for machine learning research. IEEE Signal Processing Magazine, 29 0 (6): 0 141--142, 2012

    2012

  19. [20]

    Characterizing dynamics with covariant lyapunov vectors

    Francesco Ginelli, Pietro Poggi, Alessio Turchi, Hugues Chat \'e , Roberto Livi, and Antonio Politi. Characterizing dynamics with covariant lyapunov vectors. Physical review letters, 99 0 (13): 0 130601, 2007

    work page 2007

  20. [21]

    Lagrangian coherent structures and mixing in two-dimensional turbulence

    George Haller and Guocheng Yuan. Lagrangian coherent structures and mixing in two-dimensional turbulence. Physica D: Nonlinear Phenomena, 147 0 (3-4): 0 352--370, 2000

    work page 2000

  21. [22]

    Time reversal of diffusions

    Ulrich G Haussmann and Etienne Pardoux. Time reversal of diffusions. The Annals of Probability, pages 1188--1205, 1986

    work page 1986

  22. [23]

    Strong convergence of euler-type methods for nonlinear stochastic differential equations

    Desmond J Higham, Xuerong Mao, and Andrew M Stuart. Strong convergence of euler-type methods for nonlinear stochastic differential equations. SIAM journal on numerical analysis, 40 0 (3): 0 1041--1063, 2002

    work page 2002

  23. [24]

    Classifier-free diffusion guidance

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

    Pith/arXiv arXiv 2022

  24. [25]

    Denoising diffusion probabilistic models, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. URL https://arxiv.org/abs/2006.11239

    Pith/arXiv arXiv 2020

  25. [26]

    Simoncelli, and Stéphane Mallat

    Zahra Kadkhodaie, Florentin Guth, Eero P. Simoncelli, and Stéphane Mallat. Generalization in diffusion models arises from geometry-adaptive harmonic representations, 2024 a . URL https://arxiv.org/abs/2310.02557

    Pith/arXiv arXiv 2024

  26. [27]

    Feature-guided score diffusion for sampling conditional densities

    Zahra Kadkhodaie, Stéphane Mallat, and Eero P. Simoncelli. Feature-guided score diffusion for sampling conditional densities, 2024 b . URL https://arxiv.org/abs/2410.11646

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  27. [28]

    Ergodic theory of random transformations, volume 10

    Yuri Kifer. Ergodic theory of random transformations, volume 10. Springer Science & Business Media, 2012

    work page 2012

  28. [29]

    Stochastic differential equations based on l \'e vy processes and stochastic flows of diffeomorphisms

    Hiroshi Kunita. Stochastic differential equations based on l \'e vy processes and stochastic flows of diffeomorphisms. In Real and Stochastic Analysis: New Perspectives, pages 305--373. Springer, 2004

    work page 2004

  29. [30]

    Stochastic flows and stochastic differential equations, volume 24

    Hiroshi Kunita and Hiroshi Kunita. Stochastic flows and stochastic differential equations, volume 24. Cambridge university press, 1990

    work page 1990

  30. [31]

    Theory and computation of covariant lyapunov vectors

    Pavel V Kuptsov and Ulrich Parlitz. Theory and computation of covariant lyapunov vectors. Journal of nonlinear science, 22: 0 727--762, 2012

    work page 2012

  31. [32]

    Characterization of finite-time lyapunov exponents and vectors in two-dimensional turbulence

    Guillaume Lapeyre. Characterization of finite-time lyapunov exponents and vectors in two-dimensional turbulence. Chaos: An Interdisciplinary Journal of Nonlinear Science, 12 0 (3): 0 688--698, 2002

    work page 2002

  32. [33]

    Convergence of score-based generative modeling for general data distributions

    Holden Lee, Jianfeng Lu, and Yixin Tan. Convergence of score-based generative modeling for general data distributions. In International Conference on Algorithmic Learning Theory, pages 946--985. PMLR, 2023

    2023

  33. [34]

    A sharp convergence theory for the probability flow odes of diffusion models, 2024 a

    Gen Li, Yuting Wei, Yuejie Chi, and Yuxin Chen. A sharp convergence theory for the probability flow odes of diffusion models, 2024 a . URL https://arxiv.org/abs/2408.02320

    Pith/arXiv arXiv 2024

  34. [35]

    Understanding generalizability of diffusion models requires rethinking the hidden gaussian structure

    Xiang Li, Yixiang Dai, and Qing Qu. Understanding generalizability of diffusion models requires rethinking the hidden gaussian structure. 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 57499--57538. Curran Associates, Inc., 2024 b . URL http...

    work page 2024

  35. [36]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=PqvMRDCJT9t

    2023

  36. [37]

    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. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=XVjTT1nw5z

    work page 2023

  37. [38]

    Mathematical analysis of singularities in the diffusion model under the submanifold assumption, 2024

    Yubin Lu, Zhongjian Wang, and Guillaume Bal. Mathematical analysis of singularities in the diffusion model under the submanifold assumption, 2024. URL https://arxiv.org/abs/2301.07882

    Pith/arXiv arXiv 2024

  38. [39]

    Residual corrective diffusion modeling for km-scale atmospheric downscaling

    Morteza Mardani, Noah Brenowitz, Yair Cohen, Jaideep Pathak, Chieh-Yu Chen, Cheng-Chin Liu, Arash Vahdat, Mohammad Amin Nabian, Tao Ge, Akshay Subramaniam, et al. Residual corrective diffusion modeling for km-scale atmospheric downscaling. Communications Earth & Environment, 6 0 (1): 0 124, 2025

    work page 2025

  39. [40]

    Zhang, and Markos A

    Nikiforos Mimikos-Stamatopoulos, Benjamin J. Zhang, and Markos A. Katsoulakis. Score-based generative models are provably robust: an uncertainty quantification perspective. 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 63154--63183. Curran ...

    work page 2024

  40. [41]

    Improved denoising diffusion probabilistic models, 2021

    Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models, 2021. URL https://arxiv.org/abs/2102.09672

    Pith/arXiv arXiv 2021

  41. [42]

    Diffusion models are minimax optimal distribution estimators

    Kazusato Oko, Shunta Akiyama, and Taiji Suzuki. Diffusion models are minimax optimal distribution estimators. In International Conference on Machine Learning, pages 26517--26582. PMLR, 2023

    2023

  42. [43]

    Stochastic differential equations

    Bernt ksendal and Bernt ksendal. Stochastic differential equations. Springer, 2003

    work page 2003

  43. [44]

    Ot-flow: Fast and accurate continuous normalizing flows via optimal transport

    Derek Onken, Samy Wu Fung, Xingjian Li, and Lars Ruthotto. Ot-flow: Fast and accurate continuous normalizing flows via optimal transport. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 9223--9232, 2021

    work page 2021

  44. [45]

    Normalizing flows for probabilistic modeling and inference

    George Papamakarios, Eric Nalisnick, Danilo Jimenez Rezende, Shakir Mohamed, and Balaji Lakshminarayanan. Normalizing flows for probabilistic modeling and inference. Journal of Machine Learning Research, 22 0 (57): 0 1--64, 2021

    2021

  45. [46]

    Score-based generative models detect manifolds

    Jakiw Pidstrigach. Score-based generative models detect manifolds. Advances in Neural Information Processing Systems, 35: 0 35852--35865, 2022

    2022

  46. [47]

    The intrinsic dimension of images and its impact on learning

    Phil Pope, Chen Zhu, Ahmed Abdelkader, Micah Goldblum, and Tom Goldstein. The intrinsic dimension of images and its impact on learning. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=XJk19XzGq2J

    work page 2021

  47. [48]

    Estimating the support of a high-dimensional distribution

    Bernhard Sch \"o lkopf, John C Platt, John Shawe-Taylor, Alex J Smola, and Robert C Williamson. Estimating the support of a high-dimensional distribution. Neural computation, 13 0 (7): 0 1443--1471, 2001

    work page 2001

  48. [49]

    Definition and properties of lagrangian coherent structures from finite-time lyapunov exponents in two-dimensional aperiodic flows

    Shawn C Shadden, Francois Lekien, and Jerrold E Marsden. Definition and properties of lagrangian coherent structures from finite-time lyapunov exponents in two-dimensional aperiodic flows. Physica D: Nonlinear Phenomena, 212 0 (3-4): 0 271--304, 2005

    work page 2005

  49. [50]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256--2265. pmlr, 2015

    2015

  50. [51]

    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. arXiv preprint arXiv:2011.13456, 2020

    Pith/arXiv arXiv 2011

  51. [52]

    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, 2021. URL https://openreview.net/forum?id=PxTIG12RRHS

    2021

  52. [53]

    Diffusion models encode the intrinsic dimension of data manifolds

    Jan Pawel Stanczuk, Georgios Batzolis, Teo Deveney, and Carola-Bibiane Sch \"o nlieb. Diffusion models encode the intrinsic dimension of data manifolds. In Forty-first International Conference on Machine Learning, 2024. URL https://openreview.net/forum?id=a0XiA6v256

    work page 2024

  53. [54]

    Liouville flow importance sampler

    Yifeng Tian, Nishant Panda, and Yen Ting Lin. Liouville flow importance sampler. In Forty-first International Conference on Machine Learning, 2024. URL https://openreview.net/forum?id=OMKNBzf6HJ

    work page 2024

  54. [55]

    Improving and generalizing flow-based generative models with minibatch optimal transport

    Alexander Tong, Kilian FATRAS, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport. Transactions on Machine Learning Research, 2024 a . ISSN 2835-8856. URL https://openreview.net/forum?id=CD9Snc73AW. Expert Certification

    work page 2024

  55. [56]

    Tong, Nikolay Malkin, Kilian Fatras, Lazar Atanackovic, Yanlei Zhang, Guillaume Huguet, Guy Wolf, and Yoshua Bengio

    Alexander Y. Tong, Nikolay Malkin, Kilian Fatras, Lazar Atanackovic, Yanlei Zhang, Guillaume Huguet, Guy Wolf, and Yoshua Bengio. Simulation-free S chrödinger bridges via score and flow matching. In Sanjoy Dasgupta, Stephan Mandt, and Yingzhen Li, editors, Proceedings of The 27th International Conference on Artificial Intelligence and Statistics, volume 2...

    work page 2024

  56. [57]

    Consistency and convergence rates of one-class svms and related algorithms

    R \'e gis Vert, Jean-Philippe Vert, and Bernhard Sch \"o lkopf. Consistency and convergence rates of one-class svms and related algorithms. Journal of Machine Learning Research, 7 0 (5), 2006

    work page 2006

  57. [58]

    denoising-diffusion-pytorch

    Phil Wang. denoising-diffusion-pytorch. https://github.com/lucidrains/denoising-diffusion-pytorch, 2024. Accessed: 2025-05-16

    work page 2024

  58. [59]

    Diffusion models: A comprehensive survey of methods and applications

    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 0 (4): 0 1--39, 2023

    work page 2023

  59. [60]

    Ode2vae: Deep generative second order odes with bayesian neural networks

    Cagatay Yildiz, Markus Heinonen, and Harri Lahdesmaki. Ode2vae: Deep generative second order odes with bayesian neural networks. Advances in Neural Information Processing Systems, 32, 2019

    work page 2019

  60. [61]

    The emergence of reproducibility and generalizability in diffusion models

    Huijie Zhang, Jinfan Zhou, Yifu Lu, Minzhe Guo, Peng Wang, Liyue Shen, and Qing Qu. The emergence of reproducibility and generalizability in diffusion models. arXiv preprint arXiv:2310.05264, 2023

    Pith/arXiv arXiv 2023

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.