Sch\"odinger Bridge Type Diffusion Models as an Extension of Variational Autoencoders

Kentaro Kaba; Masayuki Ohzeki; Reo Shimizu; Yuki Sughiyama

REVIEW 4 major objections 4 minor 27 references

Sch\"odinger Bridge Type Diffusion Models as an Extension of Variational Autoencoders

T0 review · 4 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read This paper claims that Schrödinger bridge diffusion models are variational autoencoders with infinitely many latent variables, making their training objective a prior loss plus a drift-matching term.

desk verdict A correct and instructive VAE-based derivation of a known Schrödinger-bridge objective; the missing regularity assumptions make the claimed bound formal, but it is a publishable conceptual contribution after revision. read the letter →

arxiv 2412.18237 v1 pith:V5RQ7UBU submitted 2024-12-24 cs.LG

Kentaro Kaba , Reo Shimizu , Masayuki Ohzeki , Yuki Sughiyama This is my paper · ORCID

classification cs.LG

keywords SchrödingerbridgediffusionmodelsvariationalautoencodersdataprocessinginequalityGirsanovtheoremdriftmatchingscore-basedgenerativepath-spaceKLdivergence

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

Schrödinger bridge diffusion models train the noising and denoising processes at once. This paper shows that their objective is really a variational autoencoder bound extended to infinitely many latent variables: one term pushes the encoded latent distribution toward the prior, and the other matches the decoder's drift to the time-reversal of the encoder. The argument rests on the data processing inequality, which lets the paper bound the data-model KL divergence by a KL divergence between path probabilities. The result gives an intuitive account of why SB-type models work and reduces them to a sum of a prior loss and a drift-matching term. Score-based models and the SB-FBSDE objective appear as special cases.

What carries the argument

The central object is the path-space KL divergence $D_{\mathrm{KL}}(P_\phi[x_{[0,T]}]\|Q_\theta[x_{[0,T]}])$ between the encoder's and decoder's path measures. The encoder conditions on the initial data point and runs forward in time; the decoder conditions on the terminal latent variable and runs backward. Three pieces of calculus turn that abstract divergence into the explicit formula above: the Onsager-Machlup path-integral representation of conditional path probabilities, the Fokker-Planck equation and its time reversal, which introduces the score $\nabla\log\rho_\phi$, and the Girsanov theorem for reverse-Itô SDEs, which converts the KL into the quadratic drift-matching expectation. The data processing inequality, applied to infinitely many latent variables, is what justifies minimizing the path-space objective in the first place.

What would settle it

For an exactly solvable case, such as linear encoder and decoder drifts with constant noise on a finite horizon, evaluate the main objective by Monte Carlo path simulation and compare it with the directly computed path-space KL divergence; agreement to numerical precision would support the derivation, while disagreement would locate a flaw in the Girsanov or path-integral step.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that any diffusion model that learns both forward and backward SDEs can be trained by minimizing $$D_{\mathrm{KL}}(p_\$\varphi$(z)\|\pi(z)) + \frac{1}{2}\int_0^T \frac{dt}{g(t)^2}\, \mathbb{E}_{\rho_\$\varphi$(t,x_t)}\left[\left\| u_\$\varphi$(t,x_t) - g(t)^2\nabla\log\rho_\$\varphi$(t,x_t) - s_\$\theta$(t,x_t) \right\|^2\right].$$ The first term is the prior loss: it trains the encoder to transport the data distribution to the prior $\pi$ by time $T$. The second term is drift matching: it trains the decoder's drift $s_\theta$ to equal the drift of the time-reversed encoder, which is the encoder drift $u_\phi$ minus the score term $g^2\nabla\log\rho_\phi$. Because the data processing inequality makes the path-space KL an upper bound on the data-space KL, training this objective is a variational way to bring the model's marginal $q_\theta$ close to the data $\mu$. The paper also claims this single derivation contains score-based models, as the special case where the encoder is fixed and the horizon is infinite, and the SB-FBSDE model, as the same objective after an implicit-score-matching rewrite.

Load-bearing premise

The main thing that has to be true is that the noising and denoising random paths can, in principle, follow the same trajectories, and that the probability density along the path is smooth enough for its logarithmic gradient and for integration by parts to be legitimate; if either fails, the KL divergence between the two path measures is infinite and the main objective is not defined.

Editorial extensions

If this is right

Score-based models are the special case where the encoder is not trained: with a fixed noising SDE whose stationary distribution is the prior, the prior loss vanishes as the horizon goes to infinity and the objective reduces to explicit score matching.
The SB-FBSDE objective is the same objective after an implicit-score-matching rewrite, so the paper's scheme implicitly solves the Schrödinger bridge problem in finite time without computing a score function.
When training succeeds, the decoder alone can generate new samples, because the decoder's path measure coincides with the time-reversed encoder path measure and no score function is needed at sampling time.
The probability-flow ODE for Schrödinger-bridge models uses the average of the encoder and decoder drifts, and it is exact only when the objective reaches zero.

Reading between the lines

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

Editorial inference: because the objective is an upper bound, the gap between the learned generative distribution and the data distribution is controlled by the sum of the prior loss and the drift-matching error; comparing those two terms during training could diagnose whether the encoder or the decoder is the bottleneck.
Editorial inference: the derivation's structure suggests the two-term decomposition should survive for other stochastic bridge constructions, including non-Markov encoders; a direct test would be to train such an encoder and check numerically whether the data-processing-inequality bound still holds.
Editorial inference: if the claim is correct, score-free finite-horizon training of SB-type models is a direct consequence of the variational view rather than an extra trick; one could test this by training a simple SB model with the score-free objective and comparing generation quality against the original objective with the same networks.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Short version: this is a clean conceptual reframing of the Schrödinger-bridge diffusion objective as a VAE with path latents, and the derivation of Eq. (20) is correct on the page. The catch is that the paper never states the regularity conditions under which the path KL and the score are well-defined, and with the empirical data distribution it uses throughout, the 'upper bound' in Eq. (12) is actually infinite, so Eq. (20) is only formal.

What is new: the VAE/DPI route to the objective. Eq. (20) is not new — the paper itself says it agrees with Theorem 4 of the SB-FBSDE model [7] — and the prior-loss plus drift-matching decomposition is standard in that literature. The genuinely new piece is showing that the existing objective can be obtained by applying the data-processing inequality to an encoder-decoder pair on path space, which makes the structure of the training loss interpretable. That is a real pedagogical and conceptual contribution, and the paper is honest about the equivalence.

What is sound: Eq. (18) is just the chain rule for KL, Eq. (19) follows from Girsanov in the standard way (modulo the usual Novikov-type conditions), and the implicit score matching in Appendix E is the textbook integration-by-parts argument. The derivations are self-contained and correctly assembled.

Soft spots: (1) The absolute-continuity and smoothness hypotheses are never stated. The paper defines mu empirically, atomic at data points. The encoder path measure then has an atomic law at t=0 while the decoder path measure has a continuous one, so the path KL is infinite at the initial layer. That makes Eq. (12) a trivial bound and Eq. (20) formally infinite at t=0; the score nabla log rho_phi is undefined as a function. This is not merely cosmetic: the claimed upper bound is not finite in the actual training setup. Fixing it requires either working with a smoothed absolutely continuous mu or truncating the integral at epsilon and passing to a limit, and the paper should say which. (2) Appendix E's integration by parts silently assumes vanishing boundary terms and integrability of div s_theta; for unbounded NN drifts and non-compact support that is extra structure. And (3) novelty: the objective itself is already in [7]; the paper should be explicit that its contribution is the derivation route, not the training loss.

Verdict: I think this paper is worth publishing as a conceptual/interpretative contribution after revision. The central derivation is right, the VAE framing is helpful, and the authors cite predecessors properly. It does not need new experiments. It does need a paragraph of assumptions and a tempering of the 'upper bound' claim.

I would send it to a serious referee, but I would expect the referee to ask for those assumptions. My own verdict is conditional, not reject.

Referee Report

4 major / 4 minor

Summary. This letter proposes a unified variational-autoencoder (VAE) interpretation of Schrödinger-bridge (SB) diffusion models. The authors model the encoder and decoder as forward and backward stochastic differential equations with neural-network drifts u_phi and s_theta, apply the data-processing inequality to path measures, and derive the main objective in Eq. (20): the KL divergence between encoder and decoder path measures decomposes into a prior loss D_KL(p_phi(z)||pi(z)) plus a drift-matching integral involving u_phi, the score g(t)^2 nabla log rho_phi, and s_theta. The paper then shows that score-based models and the SB-FBSDE objective of Chen et al. are recovered as special cases, and it rederives the probability-flow ODE. The derivation is self-contained and proceeds through the Onsager-Machlup path integral, Girsanov theorem, and an implicit-score-matching reformulation.

Significance. If the missing regularity hypotheses are supplied, the paper is a valuable conceptual contribution: it makes the SB training objective interpretable as a VAE-style decomposition, clarifies the role of the learned encoder, and unifies the derivation of score-based models, SB-FBSDE models, and probability-flow ODEs. The manuscript is careful to connect its formulas to previously published objectives, which is a useful external consistency check, and the formal derivation is reproducible from the appendices. The main weakness is that the central statements are made without the absolute-continuity and smoothness conditions needed for the KL divergences and score functions to be well defined; this is a missing-hypotheses issue rather than an algebraic error. There are no experiments, but the paper is framed as a theory/interpretation letter, so the absence of empirical evaluation is acceptable.

major comments (4)

[Diffusion models in the framework of VAEs, Eqs. (8)-(20)] The derivation requires the encoder path measure P_phi and decoder path measure Q_theta to be mutually absolutely continuous, and requires the time-marginal density rho_phi(t, x) to be smooth enough for the score function and the Fokker-Planck equation to be well defined. These conditions are never stated. With the empirical data distribution mu(x) = (1/n) sum_i delta(x - x_i) defined at the start of the paper, P_phi has an atomic marginal at t=0 while Q_theta has a smooth marginal, so the two path measures are singular and D_KL(P_phi||Q_theta) = infinity. Eq. (12) then reduces to a trivial bound, the score nabla log rho_phi(t, .) is undefined near t=0, and the objective in Eq. (20) is not a finite quantity. The authors should either assume an absolutely continuous data density with sufficient regularity, or explicitly introduce a smoothed/dequantized data distribution and discuss the limiting procedure. This is load-bearing because the central claim that Eq. (20) provides a tight upper bound on D_KL(mu||q_theta) is only formal without such hypotheses.
[Appendix D, Eq. (19)] The application of the Girsanov theorem is not justified by the hypotheses stated in the paper. To obtain Eq. (19), one needs the two conditional path measures P_phi[x_[0,T)|x_T=z] and Q_theta[x_[0,T)|x_T=z] to be mutually absolutely continuous, the noise coefficient g(t) to be strictly positive, and a Novikov-type integrability condition on the drift difference (u_phi - g^2 nabla log rho_phi - s_theta)/g(t). The discretized derivation in Eqs. (D2)-(D6) is a formal heuristic that manipulates Delta w_i / Delta t as if it were a well-defined continuous-time object. The final formula is correct under standard regularity assumptions, but as written Eq. (19) is not a theorem. The authors should state the required conditions or cite a precise Girsanov theorem and verify its hypotheses.
[Appendix E, Eqs. (E2)-(E4) and (E8)] The reformulation leading to Eq. (24) silently assumes that integration-by-parts boundary terms vanish and that the divergence of the neural-network drift s_theta is integrable against rho_phi. For unbounded NN drifts and densities with unbounded support, these assumptions are not guaranteed. In addition, Eq. (E2) uses E_mu[log mu], which is only meaningful if mu is an absolutely continuous density; for the empirical atomic mu defined in the main text this term is not well defined in the usual sense. This is the same missing-regularity issue as in the main derivation, but it affects the implicit-score-matching step and hence the practical objective Eq. (24). The authors should either add explicit conditions under which the surface terms vanish and the score identity holds, or present Eq. (24) as a formal objective requiring regularization in practice.
[Score-based models, Eq. (22)] The recovery of the score-based model objective requires taking T to infinity and using lim_{T->infty} D_KL(rho_T||pi) = 0. The paper does not state the conditions on the noising drift f and noise g under which this limit holds, nor does it address the fact that the score-matching integral is then over an infinite time horizon. This is a secondary issue, but it affects the claim that the framework exactly reproduces the standard SBM training objective rather than a truncated approximation.

minor comments (4)

[Throughout] The manuscript contains several typographical errors: 'acceletates' in the Introduction, 'culculate' in Appendix C, 'annlyzed' in Appendix C, 'descrtized represatation' in Appendix D, 'reformed' for 'reformulated' in Appendix E, 'pluging' in Appendix C, and 'forcussing' in Appendix F. A careful proofreading pass is recommended.
[Appendix C, Eqs. (C12)-(C15)] The Taylor-expansion argument used to derive the Onsager-Machlup Lagrangian for the reverse-Itô SDE is hard to follow because the notation mixes ordinary function arguments with stochastic increments. In particular, the term involving (x_{t+Delta t} - x_t)^2 is not clearly defined. Clarifying the order of the terms in Delta t would improve readability.
[Appendix E, title] The title 'Reformation of the objective function with implicit score matching' should be 'Reformulation of the objective function with implicit score matching'.
[References] Reference [20] is cited as an arXiv preprint; if a later published version exists, it should be cited instead. Also, the Girsanov theorem reference [25] could be supplemented with a more specific pointer to the version used for path measures with reverse-time SDEs.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: Eq. (20) is a self-contained algebraic decomposition of the path-space KL, not a fitted or self-citational result.

full rationale

The paper's central derivation is an identity, not a prediction. It defines encoder and decoder path measures via the SDEs (8) and (9), applies the chain rule for KL divergences to obtain Eq. (18), and then uses Girsanov's theorem to rewrite the conditional KL as the drift-matching integral in Eq. (19), yielding the main objective Eq. (20). Each step is an algebraic or measure-theoretic manipulation under the stated path-measure setup; no parameter is fitted to data and then renamed a prediction, and no external benchmark is needed to make the decomposition go through. The later agreement with score-based models, Eq. (23), and with the SB-FBSDE objective, Eq. (24), is an external consistency check that the authors explicitly acknowledge, not a circular step. The paper also contains no load-bearing self-citation: the cited prior work [7], [20], [26], etc. is not by the present authors, and the derivation does not rely on any uniqueness theorem imported from the authors' own prior work. The main caveat is technical: absolute continuity of the path measures and regularity of the score and drifts are not stated, so with the empirical atomic data distribution the KL in Eq. (12) may be infinite and the Girsanov step in Appendix D may be only formal. But missing hypotheses are a correctness/rigor concern, not circularity: the result is not assumed as an input. Under the smooth and absolutely continuous setting implicitly required, Eq. (20) follows from the stated definitions. Therefore no circularity is found.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper's derivation is built on standard stochastic calculus identities and information-theoretic inequalities. The only non-standard content is the modelling choice that encoder and decoder are SDEs with a common, non-degenerate noise coefficient, and the implicit smoothness and decay conditions on the time-marginal density. No free parameters are fitted; g(t), T, and the prior pi(z) are given inputs.

assumptions (4)

standard math Girsanov theorem for reverse-Itô SDEs and the path-integral form of conditional path probabilities
Used in Appendix D to turn the conditional KL divergence into the drift-matching term (Eq. 19). Assumes the two path measures are mutually absolutely continuous.
standard math Data processing inequality for KL divergence (chain rule for KL on product/path spaces)
Gives the upper bound (Eq. 12) that motivates minimizing DKL(P_phi||Q_theta).
domain assumption The time-marginal density rho_phi exists, is smooth, and satisfies lim_{||x||→∞} rho_phi = 0 for integration by parts
Needed for the implicit score matching step (Appendix E, Eq. E8) and for the time-reversal SDE (Eq. 16). Not stated in the paper.
domain assumption Encoder and decoder share the same non-degenerate noise coefficient g(t)>0, so their path measures are mutually absolutely continuous
Without this, the KL divergence between path measures is infinite and Eq. (20) is not a finite objective. The paper fixes the same g(t) in Eqs. (8)-(9) but never states the absolute-continuity condition explicitly.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Sch\"odinger Bridge Type Diffusion Models as an Extension of Variational Autoencoders." pith.science (2026). https://pith.science/paper/V5RQ7UBU

@misc{pith2026241218237,
  author       = {Pith},
  title        = {Pith review of: Sch\"odinger Bridge Type Diffusion Models as an Extension of Variational Autoencoders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V5RQ7UBU}},
  note         = {Machine review of arXiv:2412.18237}
}

read the original abstract

Generative diffusion models use time-forward and backward stochastic differential equations to connect the data and prior distributions. While conventional diffusion models (e.g., score-based models) only learn the backward process, more flexible frameworks have been proposed to also learn the forward process by employing the Schr\"odinger bridge (SB). However, due to the complexity of the mathematical structure behind SB-type models, we can not easily give an intuitive understanding of their objective function. In this work, we propose a unified framework to construct diffusion models by reinterpreting the SB-type models as an extension of variational autoencoders. In this context, the data processing inequality plays a crucial role. As a result, we find that the objective function consists of the prior loss and drift matching parts.

Figures

Figures reproduced from arXiv: 2412.18237 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 23 canonical work pages

[7]

Chen, G.-H

T. Chen, G.-H. Liu, and E. A. Theodorou, Likelihood training of schr¨ odinger bridge using forward-backward sdes theory, in International Conference on Learning Representations (2022)

work page 2022
[1]

Sohl-Dickstein, E

J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, Deep unsupervised learning using nonequi- librium thermodynamics, in International Conference on Machine Learning (PMLR, 2015) pp. 2256–2265

work page 2015
[2]

J. Ho, A. Jain, and P. Abbeel, Denoising diffusion proba- bilistic models, Advances in Neural Information Process- ing Systems 33, 6840 (2020)

work page 2020
[3]

Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, Score-based generative modeling through stochastic differential equations (2021)

work page 2021
[4]

G. Wang, Y. Jiao, Q. Xu, Y. Wang, and C. Yang, Deep generative learning via schr¨ odinger bridge, inProceedings of the 38th International Conference on Machine Learn- ing, Proceedings of Machine Learning Research, Vol. 139 (PMLR, 2021) pp. 10794–10804

work page 2021
[5]

Vargas, P

F. Vargas, P. Thodoroff, A. Lamacraft, and N. Lawrence, Solving schr¨ odinger bridges via maximum likelihood, En- tropy 23 (2021)

work page 2021
[6]

De Bortoli, J

V. De Bortoli, J. Thornton, J. Heng, and A. Doucet, Diffusion schr¨ odinger bridge with applications to score- based generative modeling, Advances in Neural Informa- tion Processing Systems 34, 17695 (2021)

work page 2021
[8]

A. Tong, K. F ATRAS, N. Malkin, G. Huguet, Y. Zhang, J. Rector-Brooks, G. Wolf, and Y. Bengio, Improving and generalizing flow-based generative models with mini- batch optimal transport, Transactions on Machine Learn- 10 ing Research (2024)

work page 2024

Show all 27 references

[9]

Dhariwal and A

P. Dhariwal and A. Nichol, Diffusion models beat gans on image synthesis, Advances in Neural Information Pro- cessing Systems 34, 8780 (2021)

2021
[10]

Batzolis, J

G. Batzolis, J. Stanczuk, C.-B. Sch¨ onlieb, and C. Etmann, Conditional image generation with score- based diffusion models, arXiv preprint arXiv:2111.13606 (2021)

2021 arXiv
[11]

N. Chen, Y. Zhang, H. Zen, R. J. Weiss, M. Norouzi, and W. Chan, Wavegrad: Estimating gradients for waveform generation, arXiv preprint arXiv:2009.00713 (2020)

2020 arXiv
[12]

Popov, I

V. Popov, I. Vovk, V. Gogoryan, T. Sadekova, and M. Kudinov, Grad-tts: A diffusion probabilistic model for text-to-speech, in International Conference on Ma- chine Learning (PMLR, 2021) pp. 8599–8608

2021
[13]

Y. Song, C. Durkan, I. Murray, and S. Ermon, Maxi- mum likelihood training of score-based diffusion models, Advances in Neural Information Processing Systems 34, 1415 (2021)

2021
[14]

H. Cao, C. Tan, Z. Gao, Y. Xu, G. Chen, P.-A. Heng, and S. Z. Li, A Survey on Generative Diffusion Models , IEEE Transactions on Knowledge & Data Engineering 36, 2814 (2024)

2024
[15]

Schr¨ odinger, Sur la th´ eorie relativiste de l’´ electron et l’interpr´ etation de la m´ ecanique quantique, Annales de l’institut Henri Poincar´ e2, 269 (1932)

E. Schr¨ odinger, Sur la th´ eorie relativiste de l’´ electron et l’interpr´ etation de la m´ ecanique quantique, Annales de l’institut Henri Poincar´ e2, 269 (1932)

1932
[16]

K. F. Caluya and A. Halder, Wasserstein proximal al- gorithms for the schr¨ odinger bridge problem: Density control with nonlinear drift, IEEE Transactions on Au- tomatic Control 67, 1163 (2021)

2021
[17]

Exarchos and E

I. Exarchos and E. A. Theodorou, Stochastic optimal control via forward and backward stochastic differential equations and importance sampling, Automatica 87, 159 (2018)

2018
[18]

D. P. Kingma and M. Welling, Auto-encoding variational bayes, in Advances in Neural Information Processing Sys- tems (2014)

2014
[19]

D. P. Kingma and M. Welling, An Introduction to Vari- ational Autoencoders (2019)

2019
[20]

Luo, Understanding diffusion models: A unified per- spective, arXiv preprint arXiv:2208.11970 (2022)

C. Luo, Understanding diffusion models: A unified per- spective, arXiv preprint arXiv:2208.11970 (2022)

2022 arXiv
[21]

T. M. Mitchell, Machine Learning (The McGraw-Hill, 1997)

1997
[22]

Gardiner, Stochastic Methods: A Handbook for the Natural and Social Sciences , 4th ed., Springer Series in Synergetics, Vol

C. Gardiner, Stochastic Methods: A Handbook for the Natural and Social Sciences , 4th ed., Springer Series in Synergetics, Vol. 13 (Springer, Berlin, Heidelberg, 2009)

2009
[23]

H. H. Risken, The Fokker-Planck equation : methods of solution and applications, 2nd ed., Springer series in syn- ergetics No. v. 18 (Springer-Verlag, 1996)

1996
[24]

Hirono, A

Y. Hirono, A. Tanaka, and K. Fukushima, Understand- ing diffusion models by feynman’s path integral, in Pro- ceedings of the 41st International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 235 (PMLR, 2024) pp. 18324–18351

2024
[25]

Oksendal, Stochastic differential equations: an intro- duction with applications (Springer Science & Business Media, 2013)

B. Oksendal, Stochastic differential equations: an intro- duction with applications (Springer Science & Business Media, 2013)

2013
[26]

Hyv¨ arinen and P

A. Hyv¨ arinen and P. Dayan, Estimation of non- normalized statistical models by score matching., Journal of Machine Learning Research 6 (2005)

2005
[27]

Vincent, A connection between score matching and denoising autoencoders, Neural computation 23, 1661 (2011)

P. Vincent, A connection between score matching and denoising autoencoders, Neural computation 23, 1661 (2011)

2011

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [7]

Chen, G.-H

T. Chen, G.-H. Liu, and E. A. Theodorou, Likelihood training of schr¨ odinger bridge using forward-backward sdes theory, in International Conference on Learning Representations (2022)

work page 2022

[2] [1]

Sohl-Dickstein, E

J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, Deep unsupervised learning using nonequi- librium thermodynamics, in International Conference on Machine Learning (PMLR, 2015) pp. 2256–2265

work page 2015

[3] [2]

J. Ho, A. Jain, and P. Abbeel, Denoising diffusion proba- bilistic models, Advances in Neural Information Process- ing Systems 33, 6840 (2020)

work page 2020

[4] [3]

Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, Score-based generative modeling through stochastic differential equations (2021)

work page 2021

[5] [4]

G. Wang, Y. Jiao, Q. Xu, Y. Wang, and C. Yang, Deep generative learning via schr¨ odinger bridge, inProceedings of the 38th International Conference on Machine Learn- ing, Proceedings of Machine Learning Research, Vol. 139 (PMLR, 2021) pp. 10794–10804

work page 2021

[6] [5]

Vargas, P

F. Vargas, P. Thodoroff, A. Lamacraft, and N. Lawrence, Solving schr¨ odinger bridges via maximum likelihood, En- tropy 23 (2021)

work page 2021

[7] [6]

De Bortoli, J

V. De Bortoli, J. Thornton, J. Heng, and A. Doucet, Diffusion schr¨ odinger bridge with applications to score- based generative modeling, Advances in Neural Informa- tion Processing Systems 34, 17695 (2021)

work page 2021

[8] [8]

A. Tong, K. F ATRAS, N. Malkin, G. Huguet, Y. Zhang, J. Rector-Brooks, G. Wolf, and Y. Bengio, Improving and generalizing flow-based generative models with mini- batch optimal transport, Transactions on Machine Learn- 10 ing Research (2024)

work page 2024