Pith. sign in

REVIEW 2 major objections 3 minor 1 cited by

The Effect of Stochasticity in Score-Based Diffusion Sampling: a KL Divergence Analysis

T0 review · 2 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper proves that adding stochasticity to the reverse-time SDE in score-based diffusion sampling contracts the KL divergence to the target distribution when the score is exact, and quantifies the trade-off when the score is…

desk verdict Solid new LSI-based KL decay for exact-score sampling; the approximate-score theorem has a proof gap on the score-error drift, and the abstract oversells an optimal-control analysis that isn't there. read the letter →

arxiv 2506.11378 v3 pith:W7O5F6WW submitted 2025-06-13 cs.LG

classification cs.LG MSC 60J6060H1068T0762B10
keywords score-baseddiffusionmodelsKullback-Leiblerdivergencestochasticityparameterreverse-timeSDEprobabilityflowODElog-Sobolevinequalityscoreapproximationerroroptimal
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 proves a quantitative statement about the free stochasticity parameter $\gamma$ in score-based diffusion sampling: when the score is exact, a positive $\gamma$ strictly contracts the Kullback-Leibler divergence between the generated distribution and the true marginal along the reversed-time trajectory, at an exponential rate proportional to $\gamma$, the squared noise scale, and a log-Sobolev constant of the diffused marginals. This formalizes the view of the stochastic reverse SDE as the probability-flow ODE plus a Langevin-like error-correcting term. When the score is learned and therefore approximate, the same analysis yields a trade-off: stochasticity corrects the initial prior error but amplifies the current score error, so its net effect depends on where along the path the score error is concentrated. The paper derives explicit KL bounds for both regimes, tests them on toy and benchmark datasets, and gives a fully analytic Gaussian example in which all quantities are explicit and the optimal stochasticity function can be characterized.

What carries the argument

The machinery is the reverse-time SDE family parameterized by $\gamma$, written as the probability-flow ODE plus a Langevin-like term, together with the log-Sobolev inequality (LSI) for the forward marginals: if $\mathrm{Ent}_p(f^2) \leq 2C \int |\nabla f|^2 p$, then $H(q|p) \leq (C/2) \int |\nabla \log(q/p)|^2 q$. The identity that carries the argument is the KL derivative formula from Lemma 2.2, which for exact scores gives $d/d\tau H(\tilde p_\tau|\bar p_\tau) = -(1/2)\gamma \bar g^2 \int |\nabla \log(\tilde p_\tau/\bar p_\tau)|^2 \tilde p_\tau\, dx$; the LSI turns this into a differential inequality, and Grönwall's inequality converts it into the exponential decay of Theorem 1. For approximate scores the same derivative acquires the score-error term $(1/2)\bar g^2(1+\gamma)\int \epsilon \cdot \nabla \log(\tilde p_\tau/\bar p_\tau)\, \tilde p_\tau\, dx$, and Young's plus LSI inequalities give Theorem 2. The paper also relies on Gaussian two-sided bounds and Lipschitz scores to justify differentiating the KL and to make the LSI constant finite; for compactly supported data Corollary 3.1 provides an explicit dimension-free constant.

What would settle it

Use the paper's fully analytic Gaussian example with the score error set to zero ($\alpha_\theta=1$, $\mu_\theta=\mu_0$) and a nonzero prior error; if the exact formula (40) for $H(\tilde p_\tau|\bar p_\tau)$ at some $\tau$ is not strictly smaller for a positive $\gamma$ than for $\gamma=0$, then the exponential contraction of Theorem 1 fails in the setting it claims.

Watch

Extended reading notes

Core claim

Under linear forward SDEs with additive noise, sub-Gaussian or compactly supported data, and Lipschitz scores, the paper establishes that for exact scores the KL divergence obeys $H(\tilde p_\tau|\bar p_\tau) \leq \exp\left(-\int_0^\tau C(s)\gamma(s)\bar g(s)^2\, ds\right) H(\tilde p_0|\bar p_0)$, so any $\gamma>0$ forces the relative entropy to decay and $\gamma\equiv 0$ leaves it constant in continuous time. For approximate scores with error $\epsilon$, it proves a bound with the same exponentially decaying initial-error term plus an integral of the score error weighted by the decay factor, with rate $\alpha(s)=C(s)\gamma(s)\bar g(s)^2$, and a second bound with explicit $L^2$ score error under strictly positive $\gamma$. The paper highlights that the component of the score error perpendicular to $\nabla \log(\tilde p_s/\bar p_s)$ does not affect the final KL, and that $H(\bar p_\tau|\tilde p_\tau)$, the mode-coverage direction, admits only a non-decaying bound in terms of initial and score error. The empirical message is that stochasticity helps when prior error dominates and hurts when score error dominates late in sampling.

Load-bearing premise

The load-bearing premise is that each diffused data distribution $\bar p_\tau$ satisfies a logarithmic Sobolev inequality with a finite constant $C(\tau)$; if that constant is infinite or cannot be estimated, the exponential decay factor in the KL bounds vanishes and the argument provides no provable benefit for stochastic sampling.

Editorial extensions

If this is right

  • With exact scores and compactly supported data, positive stochasticity makes the generated distribution provably closer to the target at every reversed time, while the probability-flow ODE keeps the KL gap constant.
  • The contraction rate is governed by the integrated product $C(s)\gamma(s)\bar g(s)^2$, so the design question becomes how to schedule $\gamma$ rather than whether to use it.
  • With a learned score, stochasticity is beneficial only when the initial prior error dominates the amplified score error; experimental grid searches show the best window often ends before the sampling horizon.
  • For the reverse KL direction measuring mode coverage, the paper can only bound the divergence by the initial error plus an integral of score error, with no exponential decay factor, regardless of stochasticity.

Reading between the lines

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

  • Theorem 2 suggests a concrete scheduler the paper leaves implicit: keep $\gamma$ positive early, when prior error is large, and cut it to zero late, when score error is amplified; the optimal-window experiments point in that direction and a principled schedule could be derived by minimizing the right-hand side of (25).
  • The principal direction $\nabla \log(\tilde p_s/\bar p_s)$ implies that stochastic sampling is insensitive to score-error components orthogonal to it; a training procedure that suppresses the parallel component could make the trade-off uniformly favorable.
  • When data are heavy-tailed so that no finite log-Sobolev constant exists, the paper's quantitative advantage disappears even though stochasticity may still help in practice; comparing exact-score SDE and ODE sampling on such data would separate the LSI mechanism from other sources of the correction.
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

2 major / 3 minor

Summary. This paper studies the effect of the stochasticity parameter γ in the reverse-time SDE family for score-based diffusion sampling, through KL divergence bounds. For linear forward SDEs with additive noise and Lipschitz scores, the authors prove a contraction estimate for H(p_generated | p_data) under a log-Sobolev inequality on the forward marginals, showing that exact-score stochastic sampling exponentially decays the initial prior error. For approximate scores with additive error ε, they derive a trade-off bound involving a correction term and an amplification of the score error, and they give a second bound depending only on the L2 norm of ε. The theoretical part is complemented by numerical experiments on 1D and 2D Gaussian mixtures and by a fully analytic Gaussian example in which the KL divergence and the bounds are computed explicitly.

Significance. If the technical gap identified below is repaired, the paper makes a useful conceptual contribution: it quantifies when and why positive stochasticity contracts the KL divergence H(p_generated | p_data), complementing the more common H(data | generated) bounds in the diffusion-modeling literature. The explicit dependence on the time profile of the score error, and the observation that only the component of ε along ∇ log(˜p/p̄) affects the first bound, are valuable. The analytic Gaussian example is a genuine strength: all quantities are computed explicitly, the bounds are compared with the exact KL evolution without fitted parameters, and the parameter study illustrates that stochasticity can help or hurt depending on the errors. The numerical experiments on 1D and 2D mixtures support the qualitative conclusions. However, the main theorem for approximate scores is not proved under the stated hypotheses, and the abstract promises an optimal-control characterization that does not appear in the body.

major comments (2)
  1. [Section 3.2, Proposition 3.2 and Theorem 2] The hypotheses on the score error ε are insufficient for the proof as written. Lemma 2.2 requires both densities and their gradients to admit two-sided Gaussian bounds, and in Appendix A.2 these bounds are obtained from [19, Theorem 1.2], whose hypotheses include a globally Lipschitz drift. For the approximate reverse SDE (19), the drift is −f̄ + ½ḡ²(1+γ)(∇log p̄ + ε); when ε is only C¹ with polynomial growth, this drift need not be Lipschitz. For instance, with f=0 and ε(x)=x³ the drift is superlinear, so the cited Gaussian estimates do not apply and the SDE may even fail to have a global solution. Consequently, the proof of Proposition 3.2, which simply says that the hypotheses of Proposition 3.1 are satisfied, is circular for the approximate process: the Gaussian estimates for ˜p and ∇˜p are exactly what needs to be established for the perturbed SDE. This gap propagates to formulas (21)–(22), Corollary 3.2, and Theorem 2, including the central rate α(s)=C(s)γ(s)ḡ(s)². The manuscript should either add a global Lipschitz condition on ε (and verify the resulting Gaussian estimates), or provide an alternative proof of the Gaussian bounds for the approximate reverse SDE under weaker drift assumptions.
  2. [Abstract and Section 5] The abstract states that 'the optimal stochasticity function can be characterized via an optimal control analysis,' but no such analysis appears in the body. Section 5 computes the KL divergence and the bounds for a Gaussian example and sweeps a grid of sampling parameters (µθ, σθ, µT, σT, γ), but it does not formulate a control objective, derive first-order optimality conditions, or characterize an optimal γ(τ). This is a mismatch between the advertised contribution and the delivered content. The authors should either add the optimal-control treatment or revise the abstract and Section 5 to describe the example accurately as a parameter study.
minor comments (3)
  1. [Section 3.1, Theorem 1] The decay estimate (15) is vacuous when C(τ)=0, i.e., when the forward marginal fails an LSI; for non-compact sub-Gaussian data this can happen for small τ. This is correctly stated via the definition (16), but the abstract's phrasing that stochasticity 'decreases KL divergence along the sampling trajectory' should make the LSI dependence explicit.
  2. [Section 4.3 and Appendix B.2] The text refers to 'Figures 15a and 6' to illustrate detrimental stochasticity, but Figure 15 first appears in Appendix B.2. The citation should read 'Figure 15a (Appendix B.2)' so the reader can locate it.
  3. [Equation (25) and Remark 3.2] The relation between the parameter δ in Theorem 2 and the Young-inequality parameter δ0 in the proof is easy to misread; writing δ=(1+γ)δ0 once more in the theorem statement, together with the constraint 0<δ≤γ, would improve clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the KL bounds are derived from stated external hypotheses (LSI, Lipschitz score) with no fitted parameter renamed as a prediction.

full rationale

The derivation chain runs from Lemma 2.2 (a differential identity proved in Appendix A.1 under stated regularity hypotheses) to Proposition 3.1/3.2 and then to Theorems 1-2 via Grönwall and log-Sobolev inequalities. The LSI constants are imported from external work ([6] for compactly supported data, [25] for Gaussian mixtures), not from the paper's own conclusions. The bound in Theorem 1 is not self-definitional: even though C(τ) is set to zero when no LSI is available, the nonzero case uses a genuine LSI and Grönwall to produce exponential decay. Theorem 2's error terms involve the actual score error ε and the density ratio h = p̃/p̄; the second bound (25) eliminates the unknown ∇ log h via Young's inequality, and no term is fitted to match the entropy curves. The analytic example (Section 5) solves the linear SDE explicitly and computes H(p̃τ|p̄τ) in closed form, so its agreement with the bounds is not manufactured. The only self-citation ([20] in the introduction's application list) is not load-bearing. The main risk in the paper is a correctness/generality gap, not circularity: Appendix A.2 invokes [19, Thm 1.2] which requires a Lipschitz drift, while Proposition 3.2 and Theorem 2 only assume ε is C¹ with polynomial growth, so a superlinear ε(x)=x³ would violate the condition under which the Gaussian estimates are proved. That is a missing hypothesis in a proof, not a reduction of a prediction to an input. The paper itself flags related limitations, e.g., inadequate estimation of p̃ for the first bound and the inability to ensure an LSI for p̃ in Section 4.2, so the claims are not dressed as stronger than their assumptions. Consequently, no circular step can be exhibited.

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

The theoretical results introduce no invented entities and fit no parameters. The example parameters listed are scanned for illustration, not fitted. The load-bearing inputs are the LSI assumption on the forward marginals, the Lipschitz score condition, and the external Gaussian estimates and LSI constants.

free parameters (5)
  • gamma(t) (stochasticity parameter)
    Control parameter of the sampling SDE family; studied across values (0,1,5,20) in the experiments and analytically, not fitted to data.
  • alpha_theta (score variance scaling) = scanned over a range in Figure 10
    Parameter of the analytic example's approximate score, chosen by hand to define the error model; not fitted to the target distribution.
  • beta_T (prior variance scaling) = scanned in Figure 10
    Parameter controlling the accuracy of the approximate prior in the analytic example; set by hand.
  • mu_theta (score mean shift) = scanned in Figure 10
    Translation error of the approximate score in the analytic example; chosen by hand.
  • mu_T (prior mean) = scanned in Figure 10
    Mean of the approximate prior in the analytic example; chosen by hand.
assumptions (5)
  • standard math Reverse-time formula (Lemma 2.1): the family of SDEs (2) has the same marginals as the forward SDE when started from the correct initial distribution.
    Used throughout to identify the target densities bar p_tau; proven in Appendix A.1.
  • standard math Derivative of the KL divergence formula (Lemma 2.2), adapted from [5, Lemma C.1] with stated regularity conditions.
    Central to all subsequent bounds; the paper provides a proof under Gaussian bounds and polynomial growth hypotheses.
  • domain assumption The forward marginals p_t satisfy a log-Sobolev inequality with constant C(t).
    Assumed in Theorem 1 and Theorem 2; for compactly supported data, guaranteed by Corollary 3.1 via [6, Corollary 1], and for general sub-Gaussian data only for large t.
  • domain assumption The score function grad log p_t is Lipschitz in x.
    Required in Proposition 3.1 and in the Gaussian estimates of Appendix A.2; shown for Gaussian mixtures in Lemma A.1 and for compact support in Corollary 3.1.
  • standard math Gaussian density estimates for fundamental solutions of the Fokker-Planck equation, from [19, Theorem 1.2].
    Used to justify differentiating the KL and to prove the required Gaussian bounds on tilde p in Appendix A.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Effect of Stochasticity in Score-Based Diffusion Sampling: a KL Divergence Analysis." pith.science (2026). https://pith.science/paper/W7O5F6WW

@misc{pith2026250611378,
  author       = {Pith},
  title        = {Pith review of: The Effect of Stochasticity in Score-Based Diffusion Sampling: a KL Divergence Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W7O5F6WW}},
  note         = {Machine review of arXiv:2506.11378}
}
read the original abstract

Sampling in score-based diffusion models can be performed by solving either a reverse-time stochastic differential equation (SDE) parameterized by an arbitrary stochasticity function or a probability flow ODE, corresponding to setting this stochasticity function to zero. In this work, we investigate the effect of this stochasticity on the generation process through the evolution of Kullback-Leibler (KL) divergences, obtaining general KL divergence bounds and a novel analysis of the impact of the time-profile of the score error on model performance. For exact score functions, stochasticity has a contractive effect, decreasing KL divergence along the sampling trajectory. For approximate scores, however, a trade-off arises between correcting accumulated errors and amplifying current score errors, meaning stochasticity can either improve or degrade generation performance. Theoretical considerations indicate that the gain from stochasticity depends on the time-localization of the trained model error. We test this in experiments on both toy and benchmark data sets, also comparing the KL divergence evolution with the obtained bounds. We also present a fully analytical example, where all the relevant quantities can be computed, and the optimal stochasticity function can be characterized via an optimal control analysis.

Figures

Figures reproduced from arXiv: 2506.11378 by the authors.

Figure 1
Figure 1. Data and generated distributions with exact score functions (left) and entropy evolution [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Prior error at a specific T (left) and the effect of initial time T on final entropy values (right). for large γ is a function of the ”effective step size” for the Langevin part of the reverse SDE, given by γ/nsteps ∝ γ∆t, as discussed in 3.1.1. Equation (18) gives two effective step sizes: ∆τi for the Flow ODE part, in the first line, and γ(τi)∆τi for the Langevin-like part in the second line, and for large γ the e… view at source ↗
Figure 3
Figure 3. The effect of constant γ and discretization step size on final entropy. The horizontal axis on the right is the ’effective step size’ described in 3.1.1. 4.2 Approximated score functions Here, we sample with learned score functions, in the setting of Section 3.2. The score is approximated by a multi-layer-perceptron (MLP) of size [100, 100, 100] with SiLU activations, trained by denoising score￾matching [PITH_FULL_… view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Distributions generated with approximate prior and learned score functions, entropy evolu [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Generated distributions and entropy evolution, with exact prior and learned score functions. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Generated distributions and entropy evolution, with exact prior and learned score functions, [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: The effect of γ(t) on relative entropies, sampling from approximate (top) and exact (bottom) prior with learned scores. Interestingly, comparing the cases of S = [0, T] with S = [0, Smax], we see that introducing stochas￾ticity before Smax initially reduces the entropy…
Figure 8
Figure 8. Figure 8: The evolution of the inverse entropy H(pt|p˜t), together with the bound of Corollary 3.2, for the EDM process with learned score functions. 4.3 The effect of non-conservativity In the ’Practical considerations’ of [13, Section 4], it is suggested that detrimental effec…
Figure 9
Figure 9. Figure 9: KL divergence and bounds evolution than the first one, even though its derivation uses one further inequality. This is possible since the extra inequality is applied before the LSI (see the proof of Theorem 2). Finally, [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: Final KL divergence H(˜p0|p0) for γ = 0.0, 1.0, 5.0, 20.0, for various choices of the sampling parameters µθ, σθ, µT , and σT , with µ0 = 2.0, σ0 = 1.0, σ = 6.0, and T = 1.0. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: ODE trajectories with exact scores (top) and approximate initial densities (bottom), for [PITH_FULL_IMAGE:figures/full_fig_p025_11.png]
Figure 12
Figure 12. Figure 12: Data and generated distributions for the three processes, with exact score functions, together [PITH_FULL_IMAGE:figures/full_fig_p025_12.png]
Figure 13
Figure 13. Figure 13: The effect of initial time T on final entropy values [PITH_FULL_IMAGE:figures/full_fig_p026_13.png]
Figure 14
Figure 14. Figure 14: Generated distributions (top), entropy evolution (middle) and [PITH_FULL_IMAGE:figures/full_fig_p026_14.png]
Figure 15
Figure 15. Figure 15: The effect of [Smin, Smax] on relative entropies, for MLPs with different activation functions. Jacobian J = ∇xsθ(·, t), given by A(x, t) := ∥J(x, t) − J T (x, t)∥F ∥J(x, t)∥F , (47) averaged from 1000 samples of x ∼ pt for each t, where ∥ · ∥F is the Frobenius norm …
Figure 16
Figure 16. Figure 16: Generated distributions and entropy evolution for the EDM process, with exact prior and [PITH_FULL_IMAGE:figures/full_fig_p027_16.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. A Unified Kullback--Leibler Divergence Analysis of Generative Diffusion Models via Entropy Production Rate

    math.NA 2026-08 conditional novelty 5.0 of 10

    Diffusion model generation error is decomposed through an entropy-production-rate identity that claims O(h²) Euler–Maruyama KL bounds and unifies score SDE, PF-ODE, flow matching, and stochastic interpolant analyses.

Reference graph

Works this paper leans on

33 extracted references · 22 canonical work pages · cited by 1 Pith paper

  1. [1]

    Abramson, J

    J. Abramson, J. Adler, J. Dunger, R. Evans, T. Green, A. Pritzel, O. Ronneberger, L. Willmore, A. Ballard, J. Bambrick, S. Bodenstein, D. Evans, C.-C. Hung, M. O’Neill, D. Reiman, K. Tun- yasuvunakool, C. Wu, A. ˇZemgulyt˙ e, E. Arvaniti, and J. Jumper. Accurate structure prediction of biomolecular interactions with alphafold 3. Nature, 630:493–500, 05 2024

  2. [2]

    M. S. Albergo, N. M. Boffi, and E. Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions, 2023

  3. [3]

    B. D. Anderson. Reverse-time diffusion equation models. Stochastic Process. Appl., 12(3): 313–326, 1982

  4. [4]

    Brooks, B

    T. Brooks, B. Peebles, C. Holmes, W. DePue, Y. Guo, L. Jing, D. Schnurr, J. Tay- lor, T. Luhman, E. Luhman, C. Ng, R. Wang, and A. Ramesh. Video genera- tion models as world simulators, 2024. Available at https://openai.com/research/ video-generation-models-as-world-simulators

  5. [5]

    H. Chen, H. Lee, and J. Lu. Improved analysis of score-based generative modeling: User-friendly bounds under minimal smoothness assumptions. In International Conference on Machine Learn- ing, 2022

  6. [6]

    H.-B. Chen, S. Chewi, and J. Niles-Weed. Dimension-free log-sobolev inequalities for mixture distributions. Journal of Functional Analysis , 281(11):109236, 2021

  7. [7]

    Conforti, A

    G. Conforti, A. Durmus, and M. G. Silveri. Kl convergence guarantees for score diffusion models under minimal data assumptions. SIAM Journal on Mathematics of Data Science , 7(1):86–109, 2025

  8. [8]

    Deveney, J

    T. Deveney, J. Stanczuk, L. M. Kreusser, C. Budd, and C.-B. Sch¨ onlieb. Closing the ODE-SDE gap in score-based diffusion models through the Fokker-Planck equation [preprint], 2023. Available at https://arxiv.org/abs/2311.15996

Show all 33 references
  1. [9]

    A. Figalli. Existence and uniqueness of martingale solutions for sdes with rough or degenerate coefficients. Journal of Functional Analysis , 254(1):109–153, 2008

  2. [10]

    Z. Guo, J. Liu, Y. Wang, M. Chen, D. Wang, D. Xu, and J. Cheng. Diffusion models in bioinfor- matics and computational biology. Nature Reviews Bioengineering, 2(2):136–154, 2024. 18

  3. [11]

    Jiralerspong, A

    M. Jiralerspong, A. J. Bose, I. Gemp, C. Qin, Y. Bachrach, and G. Gidel. Feature likelihood divergence: evaluating the generalization of generative models using samples. In Proceedings of the 37th International Conference on Neural Information Processing Systems , NIPS ’23, Re...

  4. [12]

    Karatzas and S

    I. Karatzas and S. E. Shreve. Brownian Motion and Stochastic Calculus . Springer-Verlag New York, 1988

  5. [13]

    Karras, M

    T. Karras, M. Aittala, T. Aila, and S. Laine. Elucidating the design space of diffusion-based generative models. NeurIPS, 2022. Available at https://arxiv.org/abs/2206.00364

  6. [14]

    Kazerouni, E

    A. Kazerouni, E. K. Aghdam, M. Heidari, R. Azad, M. Fayyaz, I. Hacihaliloglu, and D. Merhof. Diffusion models for medical image analysis: A comprehensive survey, 2023. Available at https: //arxiv.org/abs/2211.07804

  7. [15]

    H. Lee, J. Lu, and Y. Tan. Convergence for score-based generative modeling with polynomial complexity, 2023

  8. [16]

    W. Li, X. F. Cadet, D. Medina-Ortiz, M. D. Davari, R. Sowdhamini, C. Damour, Y. Li, A. Miranville, and F. Cadet. From thermodynamics to protein design: Diffusion models for biomolecule generation towards autonomous protein engineering, 2025. Available at https: //arxiv.org/abs...

  9. [17]

    L. Lin, Z. Li, R. Li, X. Li, and J. Gao. Diffusion models for time series applications: A survey,

  10. [18]

    P. A. Markowich and C. Villani. On the trend to equilibrium for the Fokker-Planck equation : An interplay between physics and functional analysis. 2004. Available at https://cedricvillani. org/sites/dev/files/old_images/2012/07/P01.MV-FPReview.pdf

  11. [19]

    Menozzi, A

    S. Menozzi, A. Pesce, and X. Zhang. Density and gradient estimates for non degenerate brownian sdes with unbounded measurable drift. Journal of Differential Equations , 272:330–369, 2021

  12. [20]

    Naiff, B

    D. Naiff, B. P. Schaeffer, G. Pires, D. Stojkovic, T. Rapstine, and F. Ramos. Controlled latent diffusion models for 3d porous media reconstruction, 2025. Available at https://arxiv.org/ abs/2503.24083

  13. [21]

    Price, A

    I. Price, A. Sanchez-Gonzalez, F. Alet, T. R. Andersson, A. El-Kadi, D. Masters, T. Ewalds, J. Stott, S. Mohamed, P. Battaglia, R. Lam, and M. Willson. Gencast: Diffusion-based ensemble forecasting for medium-range weather, 2024. Available at https://arxiv.org/abs/2312.15796

  14. [22]

    Ramesh, P

    A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, and M. Chen. Hierarchical text-conditional image generation with clip latents, 2022. Available at https://arxiv.org/abs/2204.06125

  15. [23]

    G. O. Roberts and R. L. Tweedie. Exponential convergence of langevin distributions and their discrete approximations. Bernoulli, Vol. 2, No. 4. (Dec., 1996), pp. 341-363 , 1996

  16. [24]

    M. S. M. Sajjadi, O. Bachem, M. Lucic, O. Bousquet, and S. Gelly. Assessing generative models via precision and recall. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems , volume 31....

  17. [25]

    Schlichting

    A. Schlichting. Poincar´ e and log–sobolev inequalities for mixtures.Entropy, 21(1), 2019. Available at https://arxiv.org/pdf/1812.06464

  18. [26]

    Y. Song, C. Durkan, I. Murray, and S. Ermon. Maximum likelihood training of score-based diffusion models, 2021

  19. [27]

    Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021. 19

  20. [28]

    P. Vincent. A connection between score matching and denoising autoencoders. Neural computa- tion, 23(7):1661–1674 , 2011

  21. [29]

    L. Yang, Z. Zhang, Y. Song, S. Hong, R. Xu, Y. Zhao, W. Zhang, B. Cui, and M.-H. Yang. Diffusion models: A comprehensive survey of methods and applications, 2024. Available at https: //arxiv.org/abs/2209.00796

  22. [30]

    Zhang, C

    C. Zhang, C. Zhang, S. Zheng, M. Zhang, M. Qamar, S.-H. Bae, and I. S. Kweon. A survey on audio diffusion models: Text to speech synthesis and enhancement in generative ai, 2023. Available at https://arxiv.org/abs/2303.13336

  23. [31]

    H. Zou, Z. M. Kim, and D. Kang. A survey of diffusion models in natural language processing,

  24. [33]

    A Additional proofs A.1 Section 2 Here we provide proofs of known results, which are stated in slightly different forms

    Available at hhttps://arxiv.org/abs/2305.14671. A Additional proofs A.1 Section 2 Here we provide proofs of known results, which are stated in slightly different forms. of Lemma 2.1. We show that the Fokker-Planck equations associated with both SDEs are the same. Then, the uni...

  25. [2023]

    Available at https://arxiv.org/abs/2305.00624

Pith tools

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