Pith. sign in

REVIEW 4 major objections 6 minor 52 references

Discrete Markov Bridge

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper claims that discrete diffusion models can learn their rate transition matrix instead of keeping it fixed, and that alternating matrix learning with score learning converges in KL divergence to the data distribution.

desk verdict The learned triangular rate matrix with permutation is a genuinely new and useful design for discrete diffusion, but the paper's advertised convergence guarantee is much weaker than the abstract claims and rests on optimality assumptions the actual algorithm does not satisfy. read the letter →

arxiv 2505.19752 v1 pith:RTO6WQSE submitted 2025-05-26 cs.LG cs.AIcs.CL

classification cs.LGcs.AIcs.CL
keywords discretediffusionMarkovbridgeratetransitionmatrixlearningscorevariationalinferenceconvergencerepresentation
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 argues that discrete diffusion models do not have to be stuck with fixed noise transition matrices such as Uniform or Absorb. It proposes Discrete Markov Bridge (DMB), which alternates Matrix-learning, where an adaptive rate matrix $Q_\alpha$ maps data to a latent distribution, with Score-learning, where a neural network estimates the probability ratios needed to reverse the process. The central theoretical claim is that this alternating scheme converges: if each subproblem is solved optimally in every epoch, then $D_{\mathrm{KL}}(\mu\,\|\,\hat{p}_0^{(k)})$ is non-increasing and has a limit. If true, this would give discrete diffusion the learnable latent representation that variational methods enjoy, while keeping exact discrete-state generation. The paper also reports a 1.38 bits-per-character ELBO on Text8, better than SEDD, and competitive CIFAR-10 image generation when paired with a VQ-VAE.

What carries the argument

The load-bearing object is the parameterized rate matrix $Q_\alpha = A H A^{-1}$, where $H$ is strictly upper triangular with zero row sums and $A$ is a fixed permutation matrix. It is carried by the identity $Q_\alpha = (A U)\Lambda(A U)^{-1}$, with $U$ the all-one upper triangular matrix and $\Lambda$ diagonal, which turns the matrix exponential $\exp\{Q_\alpha\}$ into diagonal exponentiation and makes the forward transition computable in $O(nd)$ space. The proof machinery also uses the reverse CTDMC formula of Theorem 3.1, where the backward rate matrix is $(p_t(y)/p_t(x))Q_{y,x}\sigma(t)$, and the KL data-processing inequality, which shows that passing distributions through a common transition kernel can only shrink their KL divergence and is used to establish monotone convergence of the alternating algorithm.

What would settle it

Train DMB on Text8 with the proposed $Q_\alpha$, and after each epoch check whether $D_{\mathrm{KL}}(\mu\,\|\,\hat{p}_0^{(k)})$ is non-increasing; a single epoch with a KL increase, or a converged score model that disagrees with Proposition 4.6's closed form for the score optimum, would show that the advertised convergence guarantee does not apply to the model actually trained.

Watch

Extended reading notes

Core claim

The central discovery is a constructive, learnable parameterization of the forward rate matrix. The paper defines $Q_\alpha = A H A^{-1}$, a permutation-conjugated strictly upper triangular matrix with zero row sums, and proves that it is diagonalizable, $Q_\alpha = (A U)\Lambda(A U)^{-1}$, so its matrix exponential is cheap to evaluate. It then proves that any two discrete distributions $p$ and $q$ can be connected by $p = q\exp\{Q\}$ for a matrix of this class, after a permutation that sorts cumulative probability ratios, so the forward process is both valid and accessible. On the learning side, the paper proves that the score model's optimum is determined by the true data distribution $\mu$, not by the current estimate $\hat{p}_0^{(k)}$, and uses this to prove that, when both learning subprocesses achieve optimality each epoch, the KL divergence between $\mu$ and the reconstructed $\hat{p}_0^{(k)}$ decreases monotonically and converges.

Load-bearing premise

The proof of Theorem 4.7 in Section D.2 depends on both subprocesses reaching their exact optima in every epoch, and Section C's Proposition 4.6 assumes every off-diagonal entry of $Q(t)$ is non-zero, while the proposed $Q_\alpha$ (Equation 3) is strictly upper triangular with many zero entries; the paper's Section H also notes that no theorem on optimality is provided.

Editorial extensions

If this is right

  • If the convergence theorem holds, alternating matrix and score updates become a principled training loop for discrete representation learning, with the learned latent distribution replacing a hand-chosen prior.
  • The class of upper-triangular diagonalizable rate matrices makes learning the noise schedule practical: matrix exponentials reduce to diagonal exponentials and storage is $O(nd)$, removing the main obstacle that restricted earlier work to Uniform or Absorb matrices.
  • Any two discrete distributions are reachable by the learned forward process after a permutation, so the representation space is not constrained by the transition structure in the way fixed matrices are.
  • The score learner optimizes against the true data distribution $\mu$ rather than the current model iterate; the convergence proof says this mismatch is still compatible with a monotone KL decrease when the subproblems are solved exactly.
  • On Text8 the framework attains a 1.38 BPC bound, below the SEDD Uniform/Absorb baselines, and on CIFAR-10 with a VQ-VAE codebook it reaches FID 11.63, comparable to image-specific generative models.

Reading between the lines

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

  • Beyond the paper: the convergence guarantee is conditional on exact per-epoch optimality, which real training never achieves; a testable practical claim is that DMB's stability depends on how close the inner loops get to their optima, so recording $J_Q$ and $J_{\mathrm{score}}$ at each epoch and comparing them with the predicted score optimum would reveal how far the proof's assumptions are from t
  • Beyond the paper: because the proposed $Q_\alpha$ is strictly upper triangular, many off-diagonal entries are zero, which violates the non-zero condition in Proposition 4.6; a direct check would be to train the score model with $Q_\alpha$ and see whether its converged values match the formula in Proposition 4.6, and if not, to design a $Q$ with positive off-diagonal entries that keeps the diagonal
  • Beyond the paper: the permutation $A$ is fixed by sorting marginal histograms; a natural extension is to learn $A$ jointly or choose it per batch, which could change which transitions are allowed and potentially improve expressiveness.
  • Beyond the paper: since the paper notes that Theorem 4.7 does not depend on the ELBO loss, the same bridge scheme may be trainable with other objectives, such as higher-order bounds or spectral losses, and would then inherit the same convergence structure under exact inner-loop optimality.
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

4 major / 6 minor

Summary. The paper introduces Discrete Markov Bridge (DMB), a discrete-state diffusion framework with two learned components: Matrix-learning, which fits a parameterized rate matrix Q_alpha of upper-triangular form, and Score-learning, which estimates the probability ratios needed for the reverse process. The authors derive an ELBO-style objective, prove conservation of total probability and accessibility between any two discrete distributions, claim a convergence guarantee for the alternating training algorithm, propose a diagonalizable rate matrix family that makes matrix exponentials cheap to compute, and report an ELBO/BPC of 1.38 on Text8 and an FID of 11.63 on CIFAR-10.

Significance. If the claims are properly scoped, the paper makes a useful contribution: the upper-triangular diagonalizable rate matrix family is parameter-efficient, has closed-form exponentials, and the accessibility construction is a genuine theoretical result for this family. The conservation-of-mass argument is correct, and the reported empirical results are competitive with existing discrete diffusion baselines. The code release is a further positive feature. The main weakness is that the headline convergence guarantee is weaker than advertised and rests on an exact-optimality premise that the actual training algorithm does not satisfy; the theoretical contribution therefore needs re-scoping or additional proof to support the claims made in the abstract and conclusion.

major comments (4)
  1. [Section 4.2, Theorem 4.7, Appendix D.2] The theorem as stated and proved only establishes that the nonnegative scalar sequence D_KL(µ || p0^(k)) is non-increasing, hence converges as a real sequence. It does not show that p0^(k) converges to µ or that the KL limit is zero. The abstract and Section 4.2 say the result 'proving the convergence of the overall framework,' which overstates what is actually shown. Please either prove convergence to the data distribution in a stronger sense or explicitly restate the result as monotone decrease and convergence of the scalar sequence.
  2. [Algorithm 2, Theorem 4.7, Appendix H] Theorem 4.7 assumes that both the Matrix-learning and Score-learning subprocesses achieve exact optimality in every epoch, and that p0^(k+1) is obtained by the exact composition p0^(k) P_{T|0}^{(k)} P_{0|T}^{(k);←}. Algorithm 2 does not run this exact operation: the inner loops stop at max_step or when L_Q < ε_Q or J_score < ε_score, and Eq. (10) updates p0 by a finite-sample Monte Carlo average. Appendix H explicitly states that no optimality theorem is provided. Thus the theorem applies to an idealized process, and no argument connects the actual training trajectory to its premise. The paper should either justify why the stopping conditions approximate exact optimality or clearly label Theorem 4.7 as a statement about an idealized exact-optimization procedure.
  3. [Proposition 4.6 and Appendix C] Proposition 4.6 assumes that Q(t)'s elements are non-zero, but the proposed Q_α in Eq. (3) is strictly upper triangular with many zero off-diagonal entries. For the trained model, the score needs to be optimal only on pairs where the forward rate Q_{y,x_t} is positive; as written, however, the proposition's global non-zero assumption does not cover the model actually trained. The support condition should be stated explicitly, and the convergence proof in Appendix D.2 should be rechecked under that condition.
  4. [Section 3.2, Eq. (10), and Section 4.2] The update of p0 in Eq. (10) uses a finite Euler-step approximation of the reverse process, while Theorem 4.7 assumes the exact reverse operator P_{0|T}^{(k);←}. The discrepancy between the discretized, Monte Carlo estimate used in training and the exact kernel used in the proof is not addressed. At minimum, the theorem statement should specify that it concerns an idealized continuous-time, exact-expectation update, separate from the implementation in Algorithm 2.
minor comments (6)
  1. [Section 2.1] The state space is introduced as X={1,2,...,n} with 'where n∈R'; this should be n∈N (or n∈Z_{>0}).
  2. [Section 5.2 and Appendix E] The main text refers to 'Proposition 5.1' while the appendix labels the same result as 'Proposition E.1' and later calls it 'Theorem 5.1'; unify the numbering and labels.
  3. [Algorithm 2, line 6] The instruction 'Update Q_α, J_Q according to Eqn. (5) and predict p_T using Eqn. (4)' is ambiguous because p_T depends on p0, which is itself updated only later in the epoch; clarify the order of updates and which p0 is used.
  4. [Equation (8)] The displayed formula for J_score is hard to parse because of the nested expectations and integral limits; please rewrite it with clear parentheses and proper placement of the dt term.
  5. [Table 1 and Section 5.4] The reported value 1.38 on Text8 is an ELBO/BPC upper bound; the text should state consistently that all comparable entries are also upper bounds, and clarify whether the '1,000 independent trials' refers to evaluation runs and whether error bars were computed.
  6. [Section 5.4 and Table 2] The sentence saying the method 'surpasses' DDPM and SNGAN is accurate only for DDPM (fixed isotropic) and SNGAN, not for DDPM (simple), which has a much lower FID; make this distinction explicit to avoid overstatement.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; the formal results follow from external CTMC facts and a data-processing inequality, not from fitted parameters or self-citation.

full rationale

The paper's theoretical chain is self-contained. Theorem 3.1 is a standard CTMC time-reversal result cited to external literature. The Matrix-learning objective J_Q is a variational prior-KL term; its validity and accessibility (Theorem 4.2) are established by an explicit construction of an upper-triangular rate matrix, not by assuming the conclusion. Proposition 4.6 is a direct first-order optimality computation for the score objective. Theorem 4.7's proof uses the data-processing inequality (Lemma D.1) to show that D_KL(mu || p0^(k)) is non-increasing under the assumed exact-optimality update; convergence of a bounded monotone scalar sequence is a standard consequence, and the conclusion is not equivalent to the premise by construction. There are no load-bearing self-citations: the author list does not appear in the references, and the reversibility/ELBO facts are cited to external prior work. The main limitations are that the optimality premise is unproven (Appendix H concedes this) and the conclusion is only scalar convergence, not convergence to mu; these are correctness/completeness concerns, not circularity. The mild self-reference in defining p_T via the model's own p0 is a design choice, not a derivation step that reduces a prediction to its fitted input.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim of a learnable bridge is built on standard CTMC and ELBO results, plus two strong domain assumptions (independent dimensions) and one ad hoc optimality assumption. The only free numerical parameters are the rate-matrix entries and the unspecified noise schedule.

free parameters (2)
  • a_1,...,a_{n-1} = not reported (learned during training)
    The entries of the learned rate matrix Q_α, optimized by minimizing J_Q in Eq. (5). These are the free parameters of the Matrix-learning process.
  • noise schedule σ(t) = not specified
    Eq. (3) defines Q(t) = σ(t) Q_α, but the paper never states the form of σ(t) or its hyperparameters. It is a free choice affecting the forward process.
assumptions (4)
  • standard math Forward Kolmogorov equation and reversibility of CTMCs (Theorem 3.1)
    Cited from Campbell et al. [3] and SEDD [4]. Standard result used without proof.
  • domain assumption Continuous-time ELBO decomposition for discrete diffusion (Eq. 9)
    The bound E[-log p0] ≤ J_score + J_Q is taken from the discrete diffusion literature [3,4,8]; the paper does not re-derive it.
  • domain assumption Independent Evolution and Independent Terminal assumptions (Section 5.1)
    Needed for the closed-form KL computation in Prop 5.2; standard in discrete diffusion but a real restriction for data with inter-dimension dependencies.
  • ad hoc to paper Optimality of both subprocesses in every epoch (Theorem 4.7)
    The convergence proof depends on this assumption, which is not enforced by Algorithm 2 and is unjustified for gradient-based training.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Discrete Markov Bridge." pith.science (2026). https://pith.science/paper/RTO6WQSE

@misc{pith2026250519752,
  author       = {Pith},
  title        = {Pith review of: Discrete Markov Bridge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RTO6WQSE}},
  note         = {Machine review of arXiv:2505.19752}
}
read the original abstract

Discrete diffusion has recently emerged as a promising paradigm in discrete data modeling. However, existing methods typically rely on a fixed rate transition matrix during training, which not only limits the expressiveness of latent representations, a fundamental strength of variational methods, but also constrains the overall design space. To address these limitations, we propose Discrete Markov Bridge, a novel framework specifically designed for discrete representation learning. Our approach is built upon two key components: Matrix Learning and Score Learning. We conduct a rigorous theoretical analysis, establishing formal performance guarantees for Matrix Learning and proving the convergence of the overall framework. Furthermore, we analyze the space complexity of our method, addressing practical constraints identified in prior studies. Extensive empirical evaluations validate the effectiveness of the proposed Discrete Markov Bridge, which achieves an Evidence Lower Bound (ELBO) of 1.38 on the Text8 dataset, outperforming established baselines. Moreover, the proposed model demonstrates competitive performance on the CIFAR-10 dataset, achieving results comparable to those obtained by image-specific generation approaches.

Figures

Figures reproduced from arXiv: 2505.19752 by the authors.

Figure 1
Figure 1. Overview of the DMB framework. DMB consists of two component: the Matrix-learning and the Score-learning. The Matrix-learning process is designed to learn an adaptive transition rate matrix, which facilitates the estimation of an adapted latent distribution. Concurrently, the score-learning process focuses on estimating the probability ratio necessary for constructing the inverse transition rate matrix, thereby enab… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 21 canonical work pages

  1. [1]

    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. InInternational Conference on Learning Representations, 2021. URLhttps://openreview.net/forum?id=PxTIG12RRHS

  2. [2]

    Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

  3. [3]

    A continuous time framework for discrete denoising models, 2022

    Andrew Campbell, Joe Benton, Valentin De Bortoli, Tom Rainforth, George Deligiannidis, and Arnaud Doucet. A continuous time framework for discrete denoising models, 2022

  4. [4]

    Discrete diffusion modeling by estimating the ratios of the data distribution.CoRR, 2023

    Aaron Lou, Chenlin Meng, and Stefano Ermon. Discrete diffusion modeling by estimating the ratios of the data distribution.CoRR, 2023

  5. [5]

    Auto-encoding variational bayes.International Conference on Learning Representations (ICLR), 2014

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes.International Conference on Learning Representations (ICLR), 2014

  6. [6]

    Neural discrete representation learning,

    Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning,

  7. [7]

    Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg

    Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. InAdvances in Neural Information Processing Systems (NeurIPS), pages 17981–17993, 2021. URL https://proceedings.neurips.cc/paper/2021/hash/ 958c530554f78bcd8e97125b70e6973d-Abstract.html

  8. [8]

    Concrete score matching: Generalized score matching for discrete data, 2023

    Chenlin Meng, Kristy Choi, Jiaming Song, and Stefano Ermon. Concrete score matching: Generalized score matching for discrete data, 2023. URLhttps://arxiv.org/abs/2211.00802

Show all 52 references
  1. [9]

    Non-autoregressive sequence generation

    Jiatao Gu and Xu Tan. Non-autoregressive sequence generation. In Luciana Benotti, Naoaki Okazaki, Yves Scherrer, and Marcos Zampieri, editors,Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: Tutorial Abstracts, pages 21–27, Dublin, Irela...

  2. [10]

    Roberta: A robustly optimized bert pretraining approach.CoRR, 2019

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach.CoRR, 2019

  3. [11]

    Bert: Pre-training of deep bidi- rectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidi- rectional transformers for language understanding. InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human lang...

  4. [12]

    Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023

  5. [13]

    Language models are unsupervised multitask learners.OpenAI blog, 1(8):9, 2019

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners.OpenAI blog, 1(8):9, 2019

  6. [14]

    Albert: A lite bert for self-supervised learning of language representations, 2020

    Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations, 2020

  7. [15]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021

  8. [16]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need, 2023

  9. [17]

    Deep residual learning for image recognition, 2015

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. 10

  10. [18]

    Denoising diffusion implicit models, 2022

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022

  11. [19]

    Weiss, Niru Maheswaranathan, and Surya Ganguli

    Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics, 2015

  12. [20]

    Diffusion-lm improves controllable text generation.Advances in neural information processing systems, 35:4328–4343, 2022

    Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation.Advances in neural information processing systems, 35:4328–4343, 2022

  13. [21]

    Hashimoto

    Ishaan Gulrajani and Tatsunori B. Hashimoto. Likelihood-based diffusion language models, 2023

  14. [22]

    Score-based continuous-time discrete diffusion models, 2023

    Haoran Sun, Lijun Yu, Bo Dai, Dale Schuurmans, and Hanjun Dai. Score-based continuous-time discrete diffusion models, 2023

  15. [23]

    Richemond, Arnaud Doucet, Robin Strudel, Chris Dyer, Conor Durkan, Curtis Hawthorne, Rémi Leblond, Will Grathwohl, and Jonas Adler

    Sander Dieleman, Laurent Sartran, Arman Roshannai, Nikolay Savinov, Yaroslav Ganin, Pierre H. Richemond, Arnaud Doucet, Robin Strudel, Chris Dyer, Conor Durkan, Curtis Hawthorne, Rémi Leblond, Will Grathwohl, and Jonas Adler. Continuous diffusion for categorical data, 2022

  16. [24]

    Large language diffusion models, 2025

    Shen Nie, Fengqi Zhu, Zebin You, Xiaolu Zhang, Jingyang Ou, Jun Hu, Jun Zhou, Yankai Lin, Ji-Rong Wen, and Chongxuan Li. Large language diffusion models, 2025. URL https://arxiv.org/abs/2502. 09992

  17. [25]

    Kingma, Tim Salimans, Ben Poole, and Jonathan Ho

    Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models, 2023. URL https://arxiv.org/abs/2107.00630

  18. [26]

    Masked diffusion models are secretly time-agnostic masked models and exploit inaccurate categorical sampling, 2024

    Kaiwen Zheng, Yongxin Chen, Hanzi Mao, Ming-Yu Liu, Jun Zhu, and Qinsheng Zhang. Masked diffusion models are secretly time-agnostic masked models and exploit inaccurate categorical sampling, 2024. URL https://arxiv.org/abs/2409.02908

  19. [27]

    Variational inference with normalizing flows, 2016

    Danilo Jimenez Rezende and Shakir Mohamed. Variational inference with normalizing flows, 2016

  20. [28]

    Kingma and Prafulla Dhariwal

    Diederik P. Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions, 2018

  21. [29]

    Flow straight and fast: Learning to generate and transfer data with rectified flow, 2022

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

  22. [30]

    Fuchs, Ingmar Posner, and Max Welling

    Victor Garcia Satorras, Emiel Hoogeboom, Fabian B. Fuchs, Ingmar Posner, and Max Welling. E(n) equivariant normalizing flows, 2022

  23. [31]

    Albergo, Nicholas M

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

  24. [32]

    Zico Kolter

    Asher Trockman and J. Zico Kolter. Orthogonalizing convolutional layers with the cayley transform, 2021

  25. [33]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks, 2014. URL https://arxiv.org/abs/ 1406.2661

  26. [34]

    Argmax flows and multinomial diffusion: Learning categorical distributions, 2021

    Emiel Hoogeboom, Didrik Nielsen, Priyank Jaini, Patrick Forré, and Max Welling. Argmax flows and multinomial diffusion: Learning categorical distributions, 2021. URLhttps://arxiv.org/abs/2102. 05379

  27. [35]

    Image style transfer based on improved convolutional neural

    Kun Sun, Mingli Jing, Yuliag Hu, and Yao Jiao. Image style transfer based on improved convolutional neural. In2021 2nd International Conference on Artificial Intelligence and Computer Engineering (ICAICE), pages 575–579, 2021. doi: 10.1109/ICAICE54393.2021.00114

  28. [36]

    Ziegler and Alexander M

    Zachary M. Ziegler and Alexander M. Rush. Latent normalizing flows for discrete sequences, 2019. URL https://arxiv.org/abs/1901.10548

  29. [37]

    Discrete flows: Invertible generative models of discrete data

    Dustin Tran, Keyon Vafa, Kumar Agrawal, Laurent Dinh, and Ben Poole. Discrete flows: Invertible generative models of discrete data. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors,Advances in Neural Information Processing Systems, ...

  30. [38]

    Training and inference on any-order autoregressive models the right way, 2022

    Andy Shih, Dorsa Sadigh, and Stefano Ermon. Training and inference on any-order autoregressive models the right way, 2022. URLhttps://arxiv.org/abs/2205.13554

  31. [39]

    Bayesian flow networks,

    Alex Graves, Rupesh Kumar Srivastava, Timothy Atkinson, and Faustino Gomez. Bayesian flow networks,

  32. [40]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors,Advances in Neural Information Processing Systems, volume 33, pages 6840–6851. Curran Asso- ciates, Inc., 2020. URL h...

  33. [41]

    Implicit generation and modeling with energy based models

    Yilun Du and Igor Mordatch. Implicit generation and modeling with energy based models. InAdvances in Neural Information Processing Systems, pages 3603–3613, 2019

  34. [42]

    Your classifier is secretly an energy based model and you should treat it like one

    Will Grathwohl, Kuan-Chieh Wang, Joern-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. In International Conference on Learning Representations, 2020

  35. [43]

    Large scale GAN training for high fidelity natural image synthesis

    Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. InInternational Conference on Learning Representations, 2019

  36. [44]

    Training generative adversarial networks with limited data.arXiv preprint arXiv:2006.06676v1, 2020

    Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data.arXiv preprint arXiv:2006.06676v1, 2020

  37. [45]

    Conditional image generation with PixelCNN decoders

    Aaron van den Oord, Nal Kalchbrenner, Oriol Vinyals, Lasse Espeholt, Alex Graves, and Koray Kavukcuoglu. Conditional image generation with PixelCNN decoders. InAdvances in Neural Information Processing Systems, pages 4790–4798, 2016

  38. [46]

    Autoregressive quantile networks for generative modeling

    Georg Ostrovski, Will Dabney, and Remi Munos. Autoregressive quantile networks for generative modeling. InInternational Conference on Machine Learning, pages 3936–3945, 2018

  39. [47]

    Generative modeling by estimating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Advances in Neural Information Processing Systems, pages 11895–11907, 2019

  40. [48]

    Spectral normalization for generative adversarial networks

    Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. InInternational Conference on Learning Representations, 2018

  41. [49]

    Your GAN is secretly an energy-based model and you should use discriminator driven latent sampling.arXiv preprint arXiv:2003.06060, 2020

    Tong Che, Ruixiang Zhang, Jascha Sohl-Dickstein, Hugo Larochelle, Liam Paull, Yuan Cao, and Yoshua Bengio. Your GAN is secretly an energy-based model and you should use discriminator driven latent sampling.arXiv preprint arXiv:2003.06060, 2020

  42. [50]

    Decoupled weight decay regularization, 2019

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. URLhttps://arxiv. org/abs/1711.05101. 12 Appendices Contents A Proof of Conservation of the Sum 14 B Proof of Accessibility 14 B.1 Proof of Lemmas . . . . . . . . . . . . . . . . . . . . . . . . . ....

  43. [2018]

    URLhttps://arxiv.org/abs/1711.00937

  44. [2024]

    URLhttps://arxiv.org/abs/2308.07037. 11

Pith tools

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