Pith. sign in

REVIEW 1 major objections 5 minor 53 references

Bidirectional Diffusion Bridge Models

T0 review · 1 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read One network trained on a masked noise-prediction loss parameterizes both directions of a diffusion bridge, so paired image-to-image translation becomes bidirectional at the cost of a single model.

desk verdict A simple shared-noise trick that works in practice, but the theoretical justification for parameter sharing is incomplete and the SOTA claim only holds on the harder direction. read the letter →

arxiv 2502.09655 v2 pith:EAQAC674 submitted 2025-02-12 cs.CV cs.AI

Duc Kieu , Kien Do , Toan Nguyen , Dang Nguyen , Thin Nguyen This is my paper · ORCID
classification cs.CVcs.AI
keywords bidirectionaldiffusionbridgeimage-to-imagetranslationChapman-Kolmogorovequationpairedimagenoise-predictionobjectiveBrowniansingle-networkgenerationmaskedconditional
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 claims that paired image-to-image translation can be made bidirectional with a single network, whereas current bridge-based methods such as I2SB, BBDM, and DDBM need a separately trained model for each direction. The construction treats a data pair $(x_0, x_T)$ as the two ends of a stochastic bridge whose intermediate states are Gaussian mixtures of the endpoints plus noise. In that setting, the forward and backward transition kernels both have closed Gaussian forms, and both are written in terms of the same noise variable $z$, so one network $z_\varphi(t, x_t, (1-m) \ast x_0, m \ast x_T)$ can predict that noise for either direction, depending on a binary mask $m$. The paper reports that this single network matches or beats dedicated unidirectional models on sketch-to-shoes, edges-to-handbags, normal-map-to-outdoor, and day-to-night translation, while also performing the reverse translation with the same checkpoint. If the claim is right, bidirectional translation costs roughly the same as unidirectional translation, and the shared training signal even improves quality.

What carries the argument

The load-bearing object is the Chapman-Kolmogorov equation for bridges, $q(x_s|\hat{x}_T,\hat{x}_0) = \int p_\theta(x_s|x_t,\hat{x}_0) q(x_t|\hat{x}_T,\hat{x}_0) dx_t$ (Eq. 14), which recasts the problem of building a bridge from $x_0$ to $x_T$ as the problem of matching, with a learned transition kernel, the marginals of a double conditional Markov process pinned at both endpoints. When those marginals are Gaussian, $q(x_t|x_0,x_T) = \mathcal{N}(\alpha_t x_0 + \beta_t x_T, \sigma_t^2 I)$, both transition directions have closed-form Gaussian expressions whose means are parameterized by the single noise variable $z$ in $x_t = \alpha_t x_0 + \beta_t x_T + \sigma_t z$; the binary mask $m$ in the network input $z_\varphi(t, x_t, (1-m) \ast x_0, m \ast x_T)$ decides which endpoint is visible, letting one model predict $z$ for both directions under the shared masked noise-prediction loss (Eq. 23). The experiments specialize the general schedule to a Brownian bridge with $\alpha_t = 1 - t/T$, $\beta_t = t/T$, $\sigma_t^2 = k(t/T)(1 - t/T)$, and the transition-kernel variance $\delta_{s,t}^2$ chosen so the kernels take the explicit forms in Eqs. 17 and 21.

What would settle it

Train BDBM on a synthetic paired dataset whose coupling is deliberately non-Gaussian or multimodal, for instance $x_T$ drawn from a two-mode Gaussian mixture whose mode assignment depends nonlinearly on $x_0$, and compare forward and reverse translation quality against a unidirectional bridge model trained with the same iteration budget; if the reverse direction degrades substantially while the forward direction does not, the Gaussian-bridge assumption is load-bearing rather than cosmetic. A more direct check of the shared-noise premise: fix $(t, x_0, x_T, z)$, form $x_t = \alpha_t x_0 + \beta_t x_T + \sigma_t z$, and compare $z_\varphi(t, x_t, x_0, 0)$ with $z_\varphi(t, x_t, 0, x_T)$; if the two masked predictions diverge by more than the training error of the loss, the two directions are not in fact predicting the same noise variable.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the forward and backward transition kernels of a diffusion bridge over paired data are the same object seen from opposite ends, so one masked network can learn both. For a bridge whose endpoint-conditioned marginals are $q(x_t|x_0,x_T) = \mathcal{N}(\alpha_t x_0 + \beta_t x_T, \sigma_t^2 I)$, with $\alpha_0 = \beta_T = 1$ and $\alpha_T = \beta_0 = \sigma_0 = \sigma_T = 0$, the forward kernel $q(x_s|x_t,x_0,x_T)$ and the backward kernel $q(x_t|x_s,x_0,x_T)$ obtained from it through Bayes' rule are both Gaussian, and their reparameterized means (Eqs. 17 and 21) are both linear in the same Gaussian noise variable $z$ that was added to form $x_t$. The paper's loss $\mathcal{L}_{\mathrm{BDBM}} = \mathbb{E}[\lVert z_\varphi(t, x_t, (1-m) \ast x_0, m \ast x_T) - z \rVert^2]$ (Eq. 23) therefore trains one network, switched by a Bernoulli mask $m$, to predict $z$ for forward steps conditioned on $x_0$ and for backward steps conditioned on $x_T$. Because the objective is a conditional score-matching loss for bridges, the paper also shows how this unifies Doob's h-transform, variational bridge learning, and conditional score matching, and it instantiates the construction as a Brownian bridge in the experiments.

Load-bearing premise

The construction assumes that each data pair $(x_0, x_T)$ can be joined by a diffusion bridge whose intermediate states are exactly Gaussian mixtures of the two endpoints plus Gaussian noise, realized in the experiments as a Brownian bridge with variance schedule $\sigma_t^2 = k(t/T)(1 - t/T)$; if the true statistical relationship between the two domains is not of this Gaussian-bridge form, the closed-form transition kernels and the shared-noise objective are approximations, and reverse-translation quality can degrade.

Editorial extensions

If this is right

  • One trained checkpoint performs translation in both directions; the reverse direction (color-to-sketch, color-to-normal) is obtained by setting $m = 1$ without any retraining (Table 7).
  • Training cost for bidirectional translation is essentially the cost of one unidirectional bridge model, since the forward and backward directions share every parameter.
  • Even for the forward direction, the paper reports that bidirectional training improves FID and LPIPS over unidirectional baselines I2SB, BBDM, and DDBM at the same number of training iterations (Table 1).
  • The Chapman-Kolmogorov formulation recovers Doob's h-transform bridges, variational bridge learning, and conditional score matching as special cases, so results obtained for this objective speak to those methods as well (Appdx A.4).
  • Discrete-time training on fixed noise levels ($T = 1000$, $\Delta t = 1$) is reported to outperform the continuous-time variant given the same iteration budget (Appdx B.3).

Reading between the lines

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

  • If the shared-noise premise holds beyond images, the same masked single-network recipe should transfer to any paired-domain problem such as text-to-image, speech-to-text, or molecule-to-property translation, because the derivation never uses image-specific structure; the paper lists text as future work but does not test it.
  • A natural stress test the paper does not run is cycle consistency: composing the forward and backward samplers of the same checkpoint and measuring $\lVert \mathrm{backward}(\mathrm{forward}(x_0)) - x_0 \rVert$ should be governed by the transition-kernel variances $\delta_{s,t}^2$; a round-trip error far larger than that would indicate the two masked parameterizations are not predicting the same n
  • The hypothesis in Section 4.2.1, that the endpoint closer in time to $x_t$ provides better noise supervision, suggests a testable training variant: weighting the mask toward the farther endpoint early in training, or annealing $m$, could reveal whether the quality gain comes from shared parameters or from the nearer-endpoint signal.
  • Because the method assumes a Gaussian (Brownian) bridge between endpoints, a dataset with many-to-many pairings (one sketch matching many shoes) should be the hardest case; the paper measures diversity but does not report how well the reverse direction preserves the one-to-many structure.
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

1 major / 5 minor

Summary. The paper introduces the Bidirectional Diffusion Bridge Model (BDBM), a diffusion-bridge framework for paired image-to-image translation that aims to learn forward and reverse translations with a single network. The method constructs a double conditional Markov process between paired endpoints x0 and xT, assumes Gaussian conditional marginals q(xt|x0,xT) = N(αt x0 + βt xT, σt² I), and derives the forward and backward transition kernels in closed form. The core proposal is to train one noise-prediction network zφ(t, xt, (1−m)x0, m xT) with a masked objective (Eq. 23), where m=0 corresponds to forward translation and m=1 to backward translation. The paper provides derivations connecting the framework to variational inference, score matching, and Doob's h-transform, and reports experiments on Edges→Shoes, Edges→Handbags, DIODE Outdoor, and Night→Day, comparing against I2SB, BBDM, DDBM, DDIB, and Rectified Flow.

Significance. If the central claim were exactly correct, this would be a practically valuable contribution: bidirectional translation with a single model halves parameters and training cost relative to training two separate bridge models. The paper's derivations of the Gaussian forward and backward transition kernels (Eqs. 17 and 21) from the Chapman-Kolmogorov equation are correct, and the connection to existing bridge frameworks is a useful conceptual synthesis. The experiments are reasonably extensive and the code is made publicly available. However, the theoretical justification for sharing a single network between the two directions is incomplete under stochastic pairings, and the empirical claim of outperforming state-of-the-art is not consistent across both translation directions. These issues affect the strength of the central claims but do not invalidate the method as a practical multi-task approach.

major comments (1)
  1. [Sec. 3.3, Eq. (23)] The paper's abstract and conclusion claim that BDBM 'outperforms state-of-the-art bridge models' and 'consistently outperforms existing I2I translation methods across diverse datasets.' This is only consistently true for the sketch-to-color direction. In the reverse direction (Appendix B.2, Table 7), BDBM underperforms BBDM on Shoes→Edges (FID 0.98 vs 0.66) and on Handbags→Edges (FID 1.87 vs 1.54), and underperforms I2SB on Outdoor→Normal (FID 11.69 vs 11.54). The claims should be qualified to the direction for which the evidence holds, or the paper should provide additional analysis explaining why the reverse direction is not consistently state-of-the-art.
minor comments (5)
  1. [Sec. 2.1] In the definition of the Markov property, the sentence 'where xu, xt, xt denote random states at times u, t, s' contains a typo: the second 'xt' should be 'xs'.
  2. [Sec. 3.1] The sentence 'we denotes ˆx0 := yA' should be 'we denote ˆx0 := yA'.
  3. [Sec. 4.2.1] The hypothesis explaining why bidirectional training outperforms unidirectional training (namely, that the optimizer leverages the nearer endpoint for more effective updates) is plausible but is not directly tested. An ablation that varies the mask probability p(m=1) or that trains BDBM-1 with the same number of per-direction updates would strengthen this explanation.
  4. [Table 7 / Appendix B.2] The sentence 'BDBM performs comparably to most baselines and even surpasses some on specific datasets, despite using only half of the training resources' is slightly imprecise: BDBM uses the same total number of training iterations as the baselines, but each individual direction receives roughly half of the updates. Please consider clarifying this.
  5. [Reference [8]] Reference [8] is a work from the same group; it is not used to support the central claim, but if a published version exists, citing it would be preferable.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the bridge derivation is self-contained; the only same-group self-citation is not load-bearing, and the shared-noise unification is an approximation rather than a circular reduction.

full rationale

Walked the derivation chain. The bridge marginals q(xt|x0,xT)=N(αt x0+βt xT, σt^2 I), the forward transition kernel (Eq. 17), and the backward transition kernel (Eq. 21) are derived from the Gaussian Chapman-Kolmogorov equations and Bayes' rule (Appendix A.5), not from the target translation outputs; they hold for any chosen schedule. The losses in Eq. 19 and Eq. 23 are conditional score-matching objectives, with the Tweedie link established in Appendix A.3; no parameter is fitted to the reported FID/LPIPS numbers and then renamed a prediction. Hyperparameters k and η are model choices studied in ablations, not circular inputs. The only same-group self-citation is Ref. [8] in a related-work list of flow models (Sec. 5.2); it is not used to justify the bridge equations or the single-network construction. The central unification claim—that one network zφ(t,xt,(1−m)x0,m xT) can serve both directions because both parameterizations use the same z—is an identification of the generative noise variable, but the two masked conditional expectations E[z|xt,x0] and E[z|xt,xT] are generally different under stochastic couplings; that is a correctness/robustness gap rather than a circular reduction. The paper is self-contained against external baselines (I2SB, BBDM, DDBM, DDIB, RF) and reports code, so no circularity score beyond the minor self-citation is warranted.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

No new physical or mathematical entities are introduced; the binary mask m is an input encoding, not an entity with independent existence.

free parameters (3)
  • k = 2 (selected by ablation, Table 4)
    Scales the bridge noise variance σt^2 = k t/T (1−t/T); larger k raises diversity but worsens FID/LPIPS; k=2 chosen as best trade-off.
  • η = 1 (selected by ablation, Table 5)
    Scales the transition variance δs,t^2 = η(σs^2 − σt^2 αs^2/αt^2); η=1 gives the fully stochastic bridge and best FID.
  • wt (loss weight) = 1
    Loss weight in Eqs. 19 and 23; set to 1 without tuning.
assumptions (6)
  • standard math Markov property holds for the underlying stochastic process.
    Used in Section 2.1 and throughout to factor transition densities.
  • standard math Chapman-Kolmogorov equation is valid for the conditional Markov process.
    Basis of Eq. 8 and the bridge CKE in Section 3.1.
  • domain assumption A double conditional Markov process with Dirac boundary conditions at x0 and xT exists for the paired data distribution.
    Section 3.1, Eq. 14; assumes the joint distribution p(yA,yB) can be realized as endpoints of a bridge.
  • domain assumption The marginal distribution q(xt|x0,xT) is Gaussian with the form Eq. 16.
    Eq. 16; the analytical transition kernels and the shared-noise reparameterization depend on this Gaussian form.
  • domain assumption The Brownian bridge schedule (αt=1−t/T, βt=t/T, σt^2=k t/T(1−t/T)) is an adequate model for the data bridge.
    Section 4.1.2; all experiments use this schedule, so the reported results are conditional on it.
  • standard math The reverse process of the bridge is Markov with transition q(xt|xs,x0,xT) given by Bayes' rule (Eq. 20).
    Used in Section 3.3 and Appendix A.5.2 to derive the backward kernel.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bidirectional Diffusion Bridge Models." pith.science (2026). https://pith.science/paper/EAQAC674

@misc{pith2026250209655,
  author       = {Pith},
  title        = {Pith review of: Bidirectional Diffusion Bridge Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EAQAC674}},
  note         = {Machine review of arXiv:2502.09655}
}
read the original abstract

Diffusion bridges have shown potential in paired image-to-image (I2I) translation tasks. However, existing methods are limited by their unidirectional nature, requiring separate models for forward and reverse translations. This not only doubles the computational cost but also restricts their practicality. In this work, we introduce the Bidirectional Diffusion Bridge Model (BDBM), a scalable approach that facilitates bidirectional translation between two coupled distributions using a single network. BDBM leverages the Chapman-Kolmogorov Equation for bridges, enabling it to model data distribution shifts across timesteps in both forward and backward directions by exploiting the interchangeability of the initial and target timesteps within this framework. Notably, when the marginal distribution given endpoints is Gaussian, BDBM's transition kernels in both directions possess analytical forms, allowing for efficient learning with a single network. We demonstrate the connection between BDBM and existing bridge methods, such as Doob's h-transform and variational approaches, and highlight its advantages. Extensive experiments on high-resolution I2I translation tasks demonstrate that BDBM not only enables bidirectional translation with minimal additional cost but also outperforms state-of-the-art bridge models. Our source code is available at [https://github.com/kvmduc/BDBM||https://github.com/kvmduc/BDBM].

Figures

Figures reproduced from arXiv: 2502.09655 by the authors.

Figure 1
Figure 1. An illustration of Bidirectional Diffusion Bridge Models (BDBM). Instead of learning two separate models zθ (t, xt, x0) and zϕ (s, xs, xT ) for the forward and backward transitions, we learn a single model zφ (t, xt,(1 − m) ∗ x0, m ∗ xT ) with a binary mask m that enables transition in both directions. Grey and white nodes denote initial and generated samples, respectively. simplicity, we assume that the process is … view at source ↗
Figure 2
Figure 2. Images generated by BDBM and unidirectional baselines in the Edges [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. LPIPS curves of BDBM and unidirectional baselines on Edges [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Images generated by BDBM and bidirectional baselines on Edges [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Samples generated by BDBM when translating from sketches to shoes using NFE=20 and [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results of BDBM on a test set of Edges [PITH_FULL_IMAGE:figures/full_fig_p025_6.png]
Figure 7
Figure 7. Figure 7: Qualitative results of BDBM on a test set of Edges [PITH_FULL_IMAGE:figures/full_fig_p025_7.png]
Figure 8
Figure 8. Figure 8: Qualitative results of BDBM on a test set of Night [PITH_FULL_IMAGE:figures/full_fig_p025_8.png]
Figure 9
Figure 9. Figure 9: Samples generated by BDBM and unidirectional baselines for color-to-sketch/normal map [PITH_FULL_IMAGE:figures/full_fig_p026_9.png]
Figure 10
Figure 10. Figure 10: Visualization of discrete-time BDBM and continuous-time BDBM accross [PITH_FULL_IMAGE:figures/full_fig_p027_10.png]
Figure 11
Figure 11. Figure 11: Additional qualitative results for Edges [PITH_FULL_IMAGE:figures/full_fig_p028_11.png]
Figure 12
Figure 12. Figure 12: Additional qualitative results for Edges [PITH_FULL_IMAGE:figures/full_fig_p028_12.png]
Figure 13
Figure 13. Figure 13: Additional qualitative results for DIODE Outdoor [PITH_FULL_IMAGE:figures/full_fig_p029_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 41 canonical work pages

  1. [1]

    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

    arXiv 2023

  2. [2]

    Conditional image generation with score-based diffusion models

    Georgios Batzolis, Jan Stanczuk, Carola-Bibiane Schönlieb, and Christian Etmann. Conditional image generation with score-based diffusion models. arXiv preprint arXiv:2111.13606, 2021

    arXiv 2021

  3. [3]

    Pattern recognition and machine learning

    C Bishop. Pattern recognition and machine learning. Springer google schola, 2:531–537, 2006

    work page 2006

  4. [4]

    Diffusion schrödinger bridge with applications to score-based generative modeling

    Valentin De Bortoli, James Thornton, Jeremy Heng, and Arnaud Doucet. Diffusion schrödinger bridge with applications to score-based generative modeling. In NeurIPS, pages 17695–17709, 2021

    work page 2021

  5. [5]

    Theodorou

    Tianrong Chen, Guan-Horng Liu, and Evangelos A. Theodorou. Likelihood training of schrödinger bridge using forward-backward sdes theory. In ICLR, 2022

    work page 2022

  6. [6]

    ILVR: conditioning method for denoising diffusion probabilistic models

    Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: conditioning method for denoising diffusion probabilistic models. InICCV, pages 14347–14356, 2021

    work page 2021

  7. [7]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 34:8780–8794, 2021

    2021

  8. [8]

    Variational Flow Models: Flowing in Your Style

    Kien Do, Duc Kieu, Toan Nguyen, Dang Nguyen, Hung Le, Dung Nguyen, and Thin Nguyen. Variational flow models: Flowing in your style. arXiv preprint arXiv:2402.02977, 2024

    work page Pith review arXiv 2024

Show all 53 references
  1. [9]

    Joseph L Doob and JI Doob.Classical potential theory and its probabilistic counterpart, volume

  2. [10]

    Résolution d’un système d’équations de m

    Robert Fortet. Résolution d’un système d’équations de m. schrödinger. Journal de Mathéma- tiques Pures et Appliquées, 19(1-4):83–105, 1940

    1940

  3. [11]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NIPS, 27:2672–2680, 2014

    2014

  4. [12]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NIPS, 30:6626–6637, 2017

    2017

  5. [13]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840–6851, 2020

    2020

  6. [14]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications , 2021

    2021

  7. [15]

    Estimation of non-normalized statistical models by score matching

    Aapo Hyvärinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6(4), 2005

    2005

  8. [16]

    Image-to-image translation with conditional adversarial networks

    Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. CVPR, pages 1125–1134, 2017

    2017

  9. [17]

    On the chapman-kolmogorov equation

    Jack Karush. On the chapman-kolmogorov equation. The Annals of Mathematical Statistics , 32(4):1333–1337, 1961

    1961

  10. [18]

    Unpaired image-to-image translation via neural schrödinger bridge

    Beomsu Kim, Gihyun Kwon, Kwanyoung Kim, and Jong Chul Ye. Unpaired image-to-image translation via neural schrödinger bridge. In ICLR, 2024

    2024

  11. [19]

    Variational diffusion models

    Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. NeurIPS, 34:21696–21707, 2021

    2021

  12. [20]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 13

    2015

  13. [21]

    Auto-encoding variational bayes

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes. ICLR, 2014

    2014

  14. [22]

    Minimizing trajectory curvature of ode-based generative models

    Sangyun Lee, Beomsu Kim, and Jong Chul Ye. Minimizing trajectory curvature of ode-based generative models. ICML, 202:18957–18973, 2023

    2023

  15. [23]

    A survey of the schrödinger problem and some of its connections with optimal transport

    Christian Léonard. A survey of the schrödinger problem and some of its connections with optimal transport. Discrete & Continuous Dynamical Systems-A , 34(4):1533–1574, 2014

    2014

  16. [24]

    BBDM: image-to-image translation with brownian bridge diffusion models

    Bo Li, Kaitao Xue, Bin Liu, and Yu-Kun Lai. BBDM: image-to-image translation with brownian bridge diffusion models. In CVPR, pages 1952–1961, 2023

    1952

  17. [25]

    Flow matching for generative modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In ICLR, 2022

    2022

  18. [26]

    Theodorou

    Guan-Horng Liu, Tianrong Chen, Oswin So, and Evangelos A. Theodorou. Deep generalized schrödinger bridge. In NeurIPS, 2022

    2022

  19. [27]

    Theodorou, Weili Nie, and Anima Anandkumar

    Guan-Horng Liu, Arash Vahdat, De-An Huang, Evangelos A. Theodorou, Weili Nie, and Anima Anandkumar. I2sb: Image-to-image schrödinger bridge. In ICML, volume 202, pages 22042–22062, 2023

    2023

  20. [28]

    Rectified flow: A marginal preserving approach to optimal transport

    Qiang Liu. Rectified flow: A marginal preserving approach to optimal transport. arXiv preprint arXiv:2209.14577, 2022

    2022 arXiv

  21. [29]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, et al. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2022

    2022

  22. [30]

    Let us build bridges: Understanding and extending diffusion generative models

    Xingchao Liu, Lemeng Wu, Mao Ye, and Qiang Liu. Let us build bridges: Understanding and extending diffusion generative models. In NeurIPS 2022 Workshop on Score-Based Methods, 2022

    2022

  23. [31]

    Learning diffusion bridges on constrained domains

    Xingchao Liu, Lemeng Wu, Mao Ye, and Qiang Liu. Learning diffusion bridges on constrained domains. In ICLR, 2023

    2023

  24. [32]

    On free energy, stochastic control, and schrödinger processes

    Michele Pavon and Anton Wakolbinger. On free energy, stochastic control, and schrödinger processes. In Modeling, Estimation and Control of Systems with Uncertainty: Proceedings of a Conference held in Sopron, Hungary, September 1990 , pages 334–348. Springer, 1991

    1990

  25. [33]

    Diffusion bridge mixture transports, schrödinger bridge problems and generative modeling

    Stefano Peluchetti. Diffusion bridge mixture transports, schrödinger bridge problems and generative modeling. Journal of Machine Learning Research, 24(374):1–51, 2023

    2023

  26. [34]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. In CVPR, pages 10684–10695, 2022

    2022

  27. [35]

    Palette: Image-to-image diffusion models

    Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In SIGGRAPH, pages 1–10, 2022

    2022

  28. [36]

    Improved techniques for training gans

    Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. NeurIPS, 29:2226–2234, 2016

    2016

  29. [37]

    Unit-ddpm: Unpaired image translation with denoising diffusion probabilistic models

    Hiroshi Sasaki, Chris G Willcocks, and Toby P Breckon. Unit-ddpm: Unpaired image translation with denoising diffusion probabilistic models. arXiv preprint arXiv:2104.05358, 2021

    2021 arXiv

  30. [38]

    Schrödinger

    E. Schrödinger. Über die Umkehrung der Naturgesetze . Sitzungsberichte der Preussischen Akademie der Wissenschaften. Physikalisch-mathematische Klasse. Verlag der Akademie der Wissenschaften in Kommission bei Walter De Gruyter u. Company, 1931

    1931

  31. [39]

    Diffusion schrödinger bridge matching

    Yuyang Shi, Valentin De Bortoli, Andrew Campbell, and Arnaud Doucet. Diffusion schrödinger bridge matching. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 62183–62223, 2023

    2023

  32. [40]

    Deep unsuper- vised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsuper- vised learning using nonequilibrium thermodynamics. In ICML, volume 37, pages 2256–2265, 2015. 14

    2015

  33. [41]

    Aligned diffusion schrödinger bridges

    Vignesh Ram Somnath, Matteo Pariset, Ya-Ping Hsieh, María Rodríguez Martínez, Andreas Krause, and Charlotte Bunne. Aligned diffusion schrödinger bridges. In UAI, pages 1985–1995, 2023

    1985

  34. [42]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021

    2021

  35. [43]

    Generative modeling by estimating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. NeurIPS, pages 11895–11907, 2019

    2019

  36. [44]

    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 ICLR, 2021

    2021

  37. [45]

    Dual diffusion implicit bridges for image-to-image translation

    Xuan Su, Jiaming Song, Chenlin Meng, and Stefano Ermon. Dual diffusion implicit bridges for image-to-image translation. In ICLR, 2023

    2023

  38. [46]

    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. Trans. Mach. Learn. Res., 2024

    2024

  39. [47]

    Simulation-free schr\" odinger bridges via score and flow matching

    Alexander Tong, Nikolay Malkin, Kilian Fatras, Lazar Atanackovic, Yanlei Zhang, Guillaume Huguet, Guy Wolf, and Yoshua Bengio. Simulation-free schr\" odinger bridges via score and flow matching. arXiv preprint arXiv:2307.03672, 2023

    2023 arXiv

  40. [48]

    Solving schrödinger bridges via maximum likelihood

    Francisco Vargas, Pierre Thodoroff, Austen Lamacraft, and Neil Lawrence. Solving schrödinger bridges via maximum likelihood. Entropy, 23(9):1134, 2021

    2021

  41. [49]

    Dai, Andrea F

    Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z. Dai, Andrea F. Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R. Walter, and Gre- gory Shakhnarovich. DIODE: A Dense Indoor and Outdoor DEpth Dataset. arXiv preprint arXiv:1908.00463, 2019

    1908 arXiv

  42. [50]

    The swiss army knife for image-to-image translation: Multi-task diffusion models

    Julia Wolleb, Robin Sandkühler, Florentin Bieder, and Philippe C Cattin. The swiss army knife for image-to-image translation: Multi-task diffusion models. arXiv preprint arXiv:2204.02641, 2022

    2022 arXiv

  43. [51]

    The un- reasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The un- reasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586–595, 2018

    2018

  44. [52]

    Egsde: Unpaired image-to-image translation via energy-guided stochastic differential equations

    Min Zhao, Fan Bao, Chongxuan Li, and Jun Zhu. Egsde: Unpaired image-to-image translation via energy-guided stochastic differential equations. Advances in Neural Information Processing Systems, 35:3609–3623, 2022

    2022

  45. [53]

    Edges” domain and its translation in the “Shoes

    Linqi Zhou, Aaron Lou, Samar Khanna, and Stefano Ermon. Denoising diffusion bridge models. In ICLR, 2024. 15 Table of Content for Appendix A Theoretical Results 16 A.1 Derivation of the backward CKE in Eq. 7 from Eq. 6 . . . . . . . . . . . . . . . . 16 A.2 Chapman-Kolmogorov ...

    2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.