Information-Theoretic Proofs for Diffusion Sampling

Galen Reeves; Henry D. Pfister

Information-Theoretic Proofs for Diffusion Sampling

T0 review · 0 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read A discrete-time diffusion sampler is provably close to its target if each step is small and each learned function tracks the conditional mean of the target in Gaussian noise.

desk verdict A clean, honest discrete-time proof for diffusion sampling; the conditional-mean oracle gap is real but explicitly scoped, and the paper deserves a proper referee. read the letter →

arxiv 2502.02305 v2 pith:C6HPAKS6 submitted 2025-02-04 stat.ML cs.ITcs.LGmath.IT

Galen Reeves , Henry D. Pfister This is my paper · ORCID

classification stat.MLcs.ITcs.LGmath.IT MSC 94A1760J0562B10

keywords diffusionsamplingKLdivergenceI-MMSErelationconditionalmeanestimationmomentmatchingdiscrete-timeMarkovchainsstochasticlocalizationnon-asymptoticbounds

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proves a non-asymptotic guarantee for discrete-time diffusion sampling without passing through continuous-time stochastic differential equations. It couples the sampler to an idealized comparison process $Y_k = t_k X + W_k$, whose marginals are exactly the convolved target $\mu * \mathcal{N}(0,t_k^{-1}I)$, and bounds the Kullback-Leibler divergence between the two processes. The bound shows the error is controlled by the largest step size times the target covariance plus the cumulative squared error of each learned function $f_k$ against the conditional mean $E[X \mid Y_{k-1}]$. If the step sizes are small and the functions $f_k$ are accurate, the sampler output is close to the target distribution in total variation. The paper also derives faster rates from matching higher-order moments and from optimized non-uniform step schedules.

What carries the argument

The comparison process $Y_k = t_k X + W_k$, where $t_k = \delta_1+\dots+\delta_k$ and $W$ is a zero-mean Gaussian process with $\operatorname{Cov}(W_k,W_m)=\min\{t_k,t_m\}I$, has marginals $\operatorname{Law}(t_k^{-1}Y_k)=\mu * \mathcal{N}(0,t_k^{-1}I)$ and is itself Markov, because $Y_k$ is a sufficient statistic for $X$ from the observations. The I-MMSE relation, which states that the derivative of $I(X;\sqrt{s}X+N)$ equals half the minimum mean-square error function $M(s)=E[\|X-E[X\mid \sqrt{s}X+N]\|^2]$, supplies the sandwich inequality that converts the exact KL identity in Lemma 2 into the upper bound. In Theorem 3, a moment-matching lemma bounds $D(\mu * \mathcal{N}(s^{-1})\,\|\,\nu * \mathcal{N}(s^{-1}))$ by $c_{d,m}(s/\beta)^{m+1}$ times exponential-moment terms when $\mu$ and $\nu$ share moments up to order $m$.

What would settle it

Take a target $\mu$ that mixes two well-separated Gaussian components, train an estimator of the conditional mean on a finite sample, and run the discrete-time sampler with small steps. The high-signal-to-noise conditional mean is close to a point mass at one component, so the estimation error near the decision boundary is large; computing the empirical squared-error term in Theorem 1 will show whether the resulting bound stays below a useful threshold such as $1$. If the term dominates, the practical claim that diffusion sampling converges for finite data is not supported by the theorem.

Watch

Extended reading notes

Core claim

On its own terms, the central result is Theorem 1. For a sampler driven by $Z_k = Z_{k-1} + \delta_k f_k(Z_{k-1}) + \sqrt{\delta_k}\widetilde N_k$, the KL divergence from the joint law of the comparison process $(Y_1,\dots,Y_n)$ satisfies $$\Delta_n \leq \frac{\delta_{\max}}{2}\operatorname{tr}(\operatorname{Cov}(X)) + \sum_{k=1}^n \frac{\delta_k}{2} E\big[\|f_k(Y_{k-1}) - E[X \mid Y_{k-1}]\|^2\big].$$ This reduces the sampling problem to a sequence of conditional-mean estimation problems at increasing signal-to-noise levels. The proof is exact: Lemma 2 writes $\Delta_n$ as the integrated MMSE discretization error minus the mutual information, and the I-MMSE relation plus monotonicity of the MMSE function produce the upper bound. Extensions show that a geometric step schedule lets $T$ grow nearly exponentially in $n$ with only polylogarithmic penalty, and that matching conditional moments up to order $m$ makes the bound shrink like $n^{-m}$.

Load-bearing premise

The guarantee holds only if each learned function $f_k$ can accurately approximate the conditional mean $E[X \mid Y_{k-1}]$ of the unknown target, and the paper does not prove that such functions can be learned from finite data; Theorem 3 additionally assumes sub-Gaussian tails with a known bound on the exponential moments.

Editorial extensions

If this is right

With uniform step sizes $\delta_k = T/n$ and exact conditional means, $\Delta_n \leq T\operatorname{tr}(\operatorname{Cov}(X))/(2n)$, so the KL error decays at rate $1/n$.
By Pinsker's inequality, the total variation distance between the sampler and the comparison process is at most $\sqrt{\Delta_n/2}$.
With a geometric step schedule, the final noise level $T$ can grow nearly exponentially in $n$ while maintaining convergence up to logarithmic factors.
Matching conditional moments up to order $m$ yields a bound that decays like $n^{-m}$; matching second moments gives $1/n^2$ through a sampler that also adapts the covariance of the injected noise.
The bound measures divergence to the Gaussian-convolved target $\mu * \mathcal{N}(0,t_n^{-1}I)$, so the final step must also make $t_n$ large enough for this convolution to approximate $\mu$ itself.

Reading between the lines

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

The paper leaves implicit a finite-sample learning analysis: the error term $E\|f_k(Y_{k-1}) - E[X \mid Y_{k-1}]\|^2$ can be decomposed into approximation and estimation error, and its dependence on dimension and sample size would decide when the bound is non-vacuous for real data.
The same proof machinery should carry over to function-space diffusion models on Hilbert spaces, since all inequalities are dimension-free and only the definition of the conditional mean on function-valued targets needs care.
The moment-matching argument suggests a general acceleration recipe beyond second moments: modifying the sampler to match conditional cumulants of order $m$ reduces the KL rate by a power of $n$, with the noise covariance adjusted accordingly.
The I-MMSE sandwich recasts step-size selection as numerical integration of the MMSE curve: optimal schedules are those whose left Riemann sums track the integral of $M(t)$, which could in principle be estimated adaptively during sampling.

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.

Desk Editor's Note

read the letter

What's actually new here is the proof path, not the tools. The paper gives an elementary discrete-time analysis of diffusion sampling, replacing the usual SDE-plus-discretization story with a direct KL bound between the sampler and an idealized Gaussian-convolution process. The exact KL identity (Lemma 2) via the I-MMSE relationship is new, and so are the geometric step-size optimization (Theorem 2/Corollary 1) and the n^{-m} moment-matching rate (Theorem 3). The proof of Theorem 1 is clean and self-contained: Markov chains, conditional expectations, and the I-MMSE sandwich, nothing heavier. I checked the algebra; it works. The authors also say plainly that the novelty is the path, not each step, and that is accurate.

Credit where due: the result is dimension-free, the non-asymptotic bounds are explicit, and the connection between diffusion sampling and stochastic localization is made transparent. The paper would be genuinely useful for teaching or for a theory reading group.

The main soft spot is the conditional-mean oracle. The guarantee is only as strong as the assumption that each f_k approximates E[X | Y_{k-1}] for the unknown target mu. The paper states this reduction explicitly and doesn't oversell it, so it's a scope limitation rather than a flaw. But there is no learning-theoretic result showing how to obtain such f_k from finite data, and for large T the posterior concentrates near Y/t, making the high-SNR end easier, not harder; the moderate-SNR estimation is where the real difficulty lives. Readers should not mistake this for a finite-sample guarantee for trained diffusion models.

Two minor issues. First, in the proof of Theorem 2 the displayed 'delta_n' is actually alpha times delta_n; the theorem statement is consistent with the corrected expression, so it's a harmless typo. Second, Appendix A's Lemma 3 is the least machine-checked component; it borrows a remainder bound from Chen-Niles-Weed and a rescaling step. I don't see an error, but an independent verification would raise confidence.

For whom: theorists in ML and information theory, and grad students who want a digestible entry point into diffusion sampling analysis. I'd bring it to a reading group and would cite the KL identity and step-size results.

Recommendation: send it to review. It deserves a serious referee, and with the typo fixed and a brief discussion of the estimation gap, it should be accepted.

Referee Report

0 major / 5 minor

Summary. The paper develops non-asymptotic, information-theoretic convergence bounds for discrete-time diffusion samplers. Coupling the sampler (1) to an idealized comparison process (2) whose marginal is an explicit Gaussian convolution of the target, the authors derive an exact expression for the KL divergence between the two joint laws (Lemma 2). Theorem 1 uses the I-MMSE relation to bound this divergence by a term depending on the maximum step size and a term measuring how well each update function approximates the conditional mean of the target given the comparison process. Section 4 refines the analysis for geometrically spaced step sizes (Theorem 2) and extends the framework to stochastic updates with matched moments, obtaining n^{-m} rates under sub-Gaussian tail assumptions (Theorem 3). The paper is explicitly framed as an elementary presentation, with the reduction to conditional-mean estimation stated as the key assumption rather than an internal error.

Significance. If the results hold, the paper provides a clean, self-contained proof that discrete-time diffusion sampling is close in KL divergence to the Gaussian-convolution process whenever step sizes are small and the score/conditional-mean oracles are well approximated. The proof of Theorem 1 is elementary and the I-MMSE sandwich is a nice unifying idea, giving dimension-free bounds and a transparent route to higher-order acceleration. The paper is honest about its main limitation: the guarantee is conditional on access to approximations of the oracle E[X | Y_{k-1}], and no finite-sample learning-theoretic result is provided. This is a scope limitation rather than a flaw. The explicit, verifiable derivations and the clear reduction to well-studied estimation problems are strengths.

minor comments (5)

[Section 2, Lemma 2 proof] The displayed equality "E[log pk(Yk | Yk−1)] + d/2 log(2πeδk) = h(Yk | X, Yk−1) − h(Yk | Yk−1) = I(X; Yk | Yk−1)" has a sign error: since E[log pk] = -h(Yk | Yk−1) and h(Yk | X, Yk−1) = d/2 log(2πeδk), the correct identity is E[log pk] + d/2 log(2πeδk) = h(Yk | X, Yk−1) − h(Yk | Yk−1) = -I(X; Yk | Yk−1). The final statement of Lemma 2 (with the minus sign on I(X; Yn)) is correct, but the intermediate line should be corrected to avoid an apparent contradiction.
[Section 4.1, proof of Theorem 2] The formula "δn = (α − 1)tn + (α−1)/(αn−1) T" appears to contain a typo: the correct identity for geometric step sizes is δn = (α − 1)t_{n-1} + (α−1)/(αn−1) T (i.e., t_n should be t_{n-1}). As written, the formula is inconsistent with the definition δ_{k+1} = α δ_k and with the subsequent use of the bound.
[Section 4.1, discussion of I(s)] The sentence "If µ has finite entropy then I(s) is bounded. Otherwise, I(s) increases without bound" is incorrect: even for a target with finite entropy, I(s) = I(X; √s X + N) grows without bound as s → ∞ (at least like (d/2) log s for non-degenerate X). The subsequent upper bound I(s) ≤ (1/2) log det(I + s Cov(X)) is the relevant statement and should replace this sentence.
[Appendix A, proof of Lemma 3] In the integral remainder formula, the summation index "|α|=n+1" should read "|α|=m+1", consistent with the moment order m used throughout the lemma; the current notation is inconsistent with the rest of the proof.
[Section 1.1 and Remark 3] The abstract and overview state that the sampling distribution is close to the target distribution, but Theorem 1 directly bounds the divergence to the Gaussian-convolution process whose marginal is µ * N(0, t_n^{-1} I), not µ itself. This is fine because t_n can be taken large, but the paper could make the final two-step argument (KL to the blurred target, then TV to µ via t_n → ∞) more explicit for readers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the core KL bound is derived from standard external results (I-MMSE, Markov property) and explicitly reduces diffusion sampling to conditional-mean estimation rather than presupposing the conclusion.

full rationale

The paper's central derivation is self-contained and non-circular. Theorem 1 and Lemma 2 derive the KL divergence between the sampler (1) and the comparison process (2) using only the Markov property (Lemma 1), elementary Gaussian transition computations, the orthogonality principle, and the external I-MMSE identity [8] plus monotonicity of the MMSE function. The final bound contains a discretization term and an explicit approximation term depending on E||f_k(Y_{k-1}) − E[X|Y_{k-1}]||^2. This is an explicitly stated reduction to a conditional-mean oracle, not an assumption of the conclusion; the target distribution enters through the defined comparison process and the oracle error is a quantified input. Theorem 2 and Corollary 1 follow algebraically from Lemma 2 and the I-MMSE sandwich. Theorem 3 relies on Lemma 3, adapted from the external Chen–Niles-Weed result [25], not from the authors' own prior work. No load-bearing self-citation, no fitted parameter renamed as a prediction, and no ansatz smuggled via citation are present. The acknowledged limitation—the absence of a finite-sample learning guarantee for the conditional-mean estimators—is a scope limitation, not circularity, and is properly recorded as a correctness/scope risk rather than a circularity defect.

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

No data are fitted. The step sizes, T, alpha, and L are explicit design or assumption parameters chosen by the user; they are not hidden degrees of freedom tuned to make the bound work. The constant c_{d,m} is a universal constant inherited from the divergence lemma. The comparison process is a proof device, not a newly postulated physical entity.

assumptions (5)

standard math I-MMSE relation (Guo, Shamai and Verdu 2005)
Used in the proof of Theorem 1 and Theorem 2 to convert sums of MMSE terms into mutual information increments.
standard math Divergence bound for moment-matched measures (Lemma 3, adapted from Chen and Niles-Weed 2022)
Used in the proof of Theorem 3; proved in Appendix A using Hermite expansions and Holder's inequality.
standard math Orthogonality principle for conditional expectation
Used in Lemma 2 to decompose the squared error of f_k into conditional variance plus approximation error.
domain assumption Finite second moment of the target mu
Assumed in Theorem 1 so that tr(Cov(X)) is finite and conditional means are defined.
domain assumption Sub-Gaussian tail and moment-matching conditions (Condition 1)
Required for Theorem 3 to control exponential moments and to obtain the n^{-m} rate.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Information-Theoretic Proofs for Diffusion Sampling." pith.science (2026). https://pith.science/paper/C6HPAKS6

@misc{pith2026250202305,
  author       = {Pith},
  title        = {Pith review of: Information-Theoretic Proofs for Diffusion Sampling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C6HPAKS6}},
  note         = {Machine review of arXiv:2502.02305}
}

read the original abstract

This paper provides an elementary, self-contained analysis of diffusion-based sampling methods for generative modeling. In contrast to existing approaches that rely on continuous-time processes and then discretize, our treatment works directly with discrete-time stochastic processes and yields precise non-asymptotic convergence guarantees under broad assumptions. The key insight is to couple the sampling process of interest with an idealized comparison process that has an explicit Gaussian-convolution structure. We then leverage simple identities from information theory, including the I-MMSE relationship, to bound the discrepancy (in terms of the Kullback-Leibler divergence) between these two discrete-time processes. In particular, we show that, if the diffusion step sizes are chosen sufficiently small and one can approximate certain conditional mean estimators well, then the sampling distribution is provably close to the target distribution. Our results also provide a transparent view on how to accelerate convergence by using additional randomness in each step to match higher-order moments in the comparison process.

Figures

Figures reproduced from arXiv: 2502.02305 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 19 canonical work pages

[1]

Deep unsupervised learning using nonequilibrium thermodynamics,

J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” in International Conference on Machine Learning. PMLR, 2015, pp. 2256–2265. 10

work page 2015
[2]

Generative modeling by estimating gradients of the data distribution,

Y. Song and S. Ermon, “Generative modeling by estimating gradients of the data distribution,” Advances in neural information processing systems, vol. 32, 2019

2019
[3]

Denoising diffusion probabilistic models,

J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in Neural Information Processing Systems, vol. 33, pp. 6840–6851, 2020

work page 2020
[4]

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

A. Q. Nichol, P. Dhariwal, A. Ramesh, P. Shyam, P. Mishkin, B. Mcgrew, I. Sutskever, and M. Chen, “Glide: Towards photorealistic image generation and editing with text-guided diffusion models,” in International Conference on Machine Learning. PMLR, 2022, pp. 16 784–16 804

work page 2022
[5]

High-resolution image synthesis with latent diffusion models,

R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 10 684–10 695

2022
[6]

Photorealistic text-to-image diffusion models with deep language understanding,

C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. Denton, S. K. S. Ghasemipour, B. K. Ayan, S. S. Mahdavi, R. G. Lopes, T. Salimans, J. Ho, D. J. Fleet, and M. Norouzi, “Photorealistic text-to-image diffusion models with deep language understanding,” in Advances in Neural Information Processing Systems, 2022

work page 2022
[7]

Score-based gener- ative modeling through stochastic differential equations,

Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based gener- ative modeling through stochastic differential equations,” in International Conference on Learning Representations, 2021

work page 2021
[8]

Mutual information and minimum mean-square error in Gaussian channels,

D. Guo, S. Shamai, and S. Verd´ u, “Mutual information and minimum mean-square error in Gaussian channels,” IEEE Transactions on Information Theory, vol. 51, no. 4, pp. 1261–1282, Apr. 2005

work page 2005

Show all 25 references

[9]

Zero- shot text-to-image generation,

A. Ramesh, M. Pavlov, G. Goh, S. Gray, C. Voss, A. Radford, M. Chen, and I. Sutskever, “Zero- shot text-to-image generation,” in International Conference on Machine Learning. PMLR, 2021, pp. 8821–8831

2021
[10]

Hierarchical text-conditional image generation with clip latents,

A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, and M. Chen, “Hierarchical text-conditional image generation with clip latents,” arXiv preprint arXiv:2204.06125, 2022

2022 arXiv
[11]

Thin shell implies spectral gap up to polylog via a stochastic localization scheme,

R. Eldan, “Thin shell implies spectral gap up to polylog via a stochastic localization scheme,” Geometric and Functional Analysis, vol. 23, no. 2, pp. 532–569, 2013

2013
[12]

Localization schemes: A framework for proving mixing bounds for markov chains,

Y. Chen and R. Eldan, “Localization schemes: A framework for proving mixing bounds for markov chains,” in 2022 IEEE 63rd Annual Symposium on Foundations of Computer Science (FOCS), 2022, pp. 110–122

2022
[13]

An information-theoretic view of stochastic localization,

A. El Alaoui and A. Montanari, “An information-theoretic view of stochastic localization,” IEEE Transactions on Information Theory, vol. 68, no. 11, pp. 7423–7426, 2022

2022
[14]

Sampling, diffusions, and stochastic localization,

A. Montanari, “Sampling, diffusions, and stochastic localization,” arXiv preprint arXiv:2305.10690, 2023

2023 arXiv
[15]

Information-theoretic diffusion,

X. Kong, R. Brekelmans, and G. Ver Steeg, “Information-theoretic diffusion,” in The Eleventh Interna- tional Conference on Learning Representations, 2023

2023
[16]

Interpretable diffusion via information decomposition,

X. Kong, O. Liu, H. Li, D. Yogatama, and G. Ver Steeg, “Interpretable diffusion via information decomposition,” in The Twelfth International Conference on Learning Representations, 2024

2024
[17]

Convergence of score-based generative modeling for general data distributions,

H. Lee, J. Lu, and Y. Tan, “Convergence of score-based generative modeling for general data distributions,” in International Conference on Algorithmic Learning Theory. PMLR, 2023, pp. 946–985

2023
[18]

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

S. Chen, S. Chewi, J. Li, Y. Li, A. Salim, and A. R. Zhang, “Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions,” in International Conference on Learning Representations, 2023. 11

2023
[19]

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

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 Learning, vol. 202. PMLR, 2023, pp. 4735–4763

2023
[20]

Nearly d-linear convergence bounds for diffusion models via stochastic localization,

J. Benton, V. De Bortoli, A. Doucet, and G. Deligiannidis, “Nearly d-linear convergence bounds for diffusion models via stochastic localization,” in International Conference on Learning Representations, 2024

2024
[21]

Towards faster non-asymptotic convergence for diffusion-based generative models,

G. Li, Y. Wei, Y. Chen, and Y. Chi, “Towards faster non-asymptotic convergence for diffusion-based generative models,” in International Conference on Learning Representations, 2024

2024
[22]

Accelerating convergence of score-based diffusion models, provably,

G. Li, Y. Huang, T. Efimov, Y. Wei, Y. Chi, and Y. Chen, “Accelerating convergence of score-based diffusion models, provably,” in International Conference on Machine Learning, 2024

2024
[23]

Stochastic Runge-Kutta methods: Provable acceleration of diffusion models,

Y. Wu, Y. Chen, and Y. Wei, “Stochastic Runge-Kutta methods: Provable acceleration of diffusion models,” 2024. [Online]. Available: https://arxiv.org/abs/2410.04760

2024 arXiv
[24]

Provable acceleration for diffusion models under minimal assumptions,

G. Li and C. Cai, “Provable acceleration for diffusion models under minimal assumptions,” 2024. [Online]. Available: https://arxiv.org/abs/2410.23285

2024 arXiv
[25]

Asymptotics of smoothed Wasserstein distances,

H.-B. Chen and J. Niles-Weed, “Asymptotics of smoothed Wasserstein distances,” Potential Analysis, vol. 56, pp. 571–595, 2022. A Moment Matching Divergence Bound The following result provides a uniform upper bound on the divergence between distributions satisfying a moment mat...

2022

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

Deep unsupervised learning using nonequilibrium thermodynamics,

J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” in International Conference on Machine Learning. PMLR, 2015, pp. 2256–2265. 10

work page 2015

[2] [2]

Generative modeling by estimating gradients of the data distribution,

Y. Song and S. Ermon, “Generative modeling by estimating gradients of the data distribution,” Advances in neural information processing systems, vol. 32, 2019

2019

[3] [3]

Denoising diffusion probabilistic models,

J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in Neural Information Processing Systems, vol. 33, pp. 6840–6851, 2020

work page 2020

[4] [4]

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

A. Q. Nichol, P. Dhariwal, A. Ramesh, P. Shyam, P. Mishkin, B. Mcgrew, I. Sutskever, and M. Chen, “Glide: Towards photorealistic image generation and editing with text-guided diffusion models,” in International Conference on Machine Learning. PMLR, 2022, pp. 16 784–16 804

work page 2022

[5] [5]

High-resolution image synthesis with latent diffusion models,

R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 10 684–10 695

2022

[6] [6]

Photorealistic text-to-image diffusion models with deep language understanding,

C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. Denton, S. K. S. Ghasemipour, B. K. Ayan, S. S. Mahdavi, R. G. Lopes, T. Salimans, J. Ho, D. J. Fleet, and M. Norouzi, “Photorealistic text-to-image diffusion models with deep language understanding,” in Advances in Neural Information Processing Systems, 2022

work page 2022

[7] [7]

Score-based gener- ative modeling through stochastic differential equations,

Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based gener- ative modeling through stochastic differential equations,” in International Conference on Learning Representations, 2021

work page 2021

[8] [8]

Mutual information and minimum mean-square error in Gaussian channels,

D. Guo, S. Shamai, and S. Verd´ u, “Mutual information and minimum mean-square error in Gaussian channels,” IEEE Transactions on Information Theory, vol. 51, no. 4, pp. 1261–1282, Apr. 2005

work page 2005