Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Transition Matching: Scalable and Flexible Generative Modeling

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A generative framework that learns full transition probabilities, not only expected steps, outperforms flow matching on text-to-image and enables fully causal image generation.

desk verdict Solid framework paper; DTM is well-supported, but the causal-AR variants rest on a poorly-characterized design choice and the SOTA claims need external validation. read the letter →

arxiv 2506.23589 v1 pith:KJWETYAN submitted 2025-06-30 cs.LG cs.AI

Neta Shaul , Uriel Singer , Itai Gat , Yaron Lipman This is my paper
classification cs.LGcs.AI
keywords transitionmatchingflowautoregressivegenerationtext-to-imagediscrete-timeMarkovprocesscontinuoustokensdiffusionmodels
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

Transition Matching recasts generative modeling as learning the transition kernels of a discrete-time Markov chain, rather than learning only the expected step, as flow matching does. The paper shows that this one change, augmented by a choice of supervision process and kernel parameterization, produces three working models: DTM, a discrete-time generalization of flow matching, and ARTM and FHTM, partially and fully causal autoregressive variants. On a fixed 1.7B-parameter DiT architecture and dataset, DTM reports higher image quality and text adherence than flow matching, and does so with roughly 16 backbone evaluations per image instead of 128, a near-7x sampling speedup. ARTM and FHTM are reported to match non-causal baselines in image quality, and FHTM is claimed to be the first fully causal continuous model to match or surpass flow-based text-to-image performance.

What carries the argument

The load-bearing object is the transition kernel $p^\theta_{t+1|t}(x_{t+1}|x_t)$ of a Markov chain, trained to match a supervising conditional $q_{t+1|t}$ through an empirical loss over sampled pairs $(X_t, X_{t+1})$. A latent-variable parameterization $q_{t+1|t}(x_{t+1}|x_t) = \int q_{t+1|t,Y}(x_{t+1}|x_t,y) q_{Y|t}(y|x_t) dy$ turns learning into predicting a latent $Y$, and the model is an FM velocity head trained with the conditional flow-matching loss. Two named objects carry the results: the difference latent $Y = X_T - X_0$, which makes the transition $X_{t+1} = X_t + \frac{1}{T}Y$ and links DTM to flow matching; and the independent linear process (equation 15), which draws a new $X_{0,t} \sim \mathcal{N}(0,I)$ at every $t$ and is what makes the autoregressive kernels work.

What would settle it

Train ARTM-3 and FHTM-3 with the dependent linear process (equation 10) at full scale and compare CLIPScore and PickScore on PartiPrompts; if the gap over the independent process (equation 15) disappears or reverses, the paper's claim that independent noise enables causal AR generation is refuted. A second check on the convergence theorem: measure the conditional variance $\mathrm{Var}(X_{t+1}|X_t)$ under the learned DTM kernel at decreasing step sizes; it should shrink toward the deterministic FM update as $T \to \infty$.

Watch

Extended reading notes

Core claim

Transition Matching treats generation as a finite-horizon Markov chain and trains each step by matching the probability kernel of a supervising stochastic process, rather than only its conditional mean. DTM chooses the linear (Cond-OT) process with difference latent $Y = X_T - X_0$; its expected step is exactly the flow-matching velocity, and as $T \to \infty$ its samples converge to Euler-step flow matching (Theorem 1). ARTM and FHTM instead use the independent linear process $X_t = (1 - t/T)X_{0,t} + (t/T)X_T$ with fresh noise at each time, which widens the conditional support of the next state and lets an autoregressive kernel be trained by teacher forcing. On text-to-image, the paper reports DTM bests flow matching on quality and prompt alignment while requiring roughly 16 backbone evaluations instead of 128, and reports FHTM as the first fully causal model to match or surpass flow-based continuous text-to-image generation.

Load-bearing premise

The load-bearing assumption is that the independent linear process (equation 15) is the right supervising process for causal autoregressive kernels; the paper offers no theory for this, only an empirical advantage over the dependent process and a hypothesis that the dependent process lets the AR kernel take a linear shortcut.

Editorial extensions

If this is right

  • DTM can replace a flow-matching model with no change to architecture or training data and immediately improve text-to-image quality and prompt alignment while cutting sampling cost by about 7x (16 backbone forwards vs 128).
  • ARTM and FHTM establish that causal, continuous-token autoregressive generation can reach the image quality of non-causal flow and masked-autoregressive models, removing a main architectural obstacle to fusing image generation with language-model-style next-token training.
  • Because DTM's expected transition coincides with the flow-matching velocity, the method interpolates between stochastic discrete-time generation and deterministic flow; Theorem 1 says the two coincide as the step size goes to zero.
  • The independent linear process is a new, reusable supervision choice; the paper reports that it is essential for ARTM and FHTM, with the dependent linear process failing on AR kernels.
  • FHTM implemented in an LLM-style architecture with 1D positions and prefix text matches or improves on a same-size DiT model, suggesting the causal variant can be slotted into existing multimodal autoregressive systems.

Reading between the lines

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

  • A testable extension beyond the paper: train FHTM on video or audio, where temporal coherence is stronger, to see whether the independent-linear-process advantage persists outside images.
  • The paper's DTM architecture generates each image patch independently, so its reported gains may understate what a correlated multi-token kernel could do at low transition-step counts, since a larger patch head helps most in the 1-4 step regime.
  • Because FHTM is fully causal, a natural next move, not explored here, is prefix-conditioned or in-context image generation, where the model observes some image tokens and completes the rest.
  • The convergence theorem implies a stochasticity knob controlled by step size; tuning the residual noise between DTM and FM limits could trade sample diversity for fidelity.
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 / 4 minor

Summary. The paper introduces Transition Matching (TM), a discrete-time, continuous-state generative framework in which a Markov chain is trained by matching the transition kernels of a supervising process. The framework is instantiated in three variants: Difference Transition Matching (DTM), which learns the distribution of differences X_T - X_0 rather than only its expectation; Autoregressive Transition Matching (ARTM) and Full History Transition Matching (FHTM), which use causal autoregressive kernels with a newly introduced independent linear supervising process. The central empirical claims are that DTM outperforms re-implemented flow matching and diffusion baselines in image quality and text adherence while sampling about seven times faster, and that ARTM/FHTM achieve generation quality comparable to non-causal baselines, with FHTM described as the first fully causal continuous-domain model to match flow-based text-to-image performance. The paper also proves that DTM converges to Euler-step flow matching as the number of steps tends to infinity.

Significance. If the claims hold, the TM framework is a genuinely useful unification of flow/diffusion models and continuous autoregressive models, and the paper makes a credible theoretical contribution in the convergence proof of DTM to flow matching. The strengths of the paper are its clean general formulation, the explicit training and sampling algorithms, the controlled experimental setup with fixed architecture, data, and hyperparameters, and the provision of Python code in the appendix. However, the significance is currently qualified by three issues: the causal ARTM/FHTM results depend on a supervising process whose advantage is supported only by a single empirical comparison on one proprietary dataset; none of the main tables report error bars or repeated runs; and the 'state-of-the-art' and 'first' claims are made only against re-implemented baselines, not against published results from other systems. These issues are addressable but should be fixed before the claims can be taken at face value.

major comments (4)
  1. [Section 2.2, Eq. (15); Appendix A.4] The ARTM and FHTM causal models rely on the independent linear process in Eq. (15), but the paper provides no theoretical characterization of when this process yields a valid or learnable supervising process. The only evidence for its advantage is Figure 9, which shows 100K-iteration runs on the proprietary Shutterstock dataset, and the mechanism is explicitly left as a hypothesis: 'We hypothesize that the AR kernels exploit the linear relationship between Xt and Xt+1 during training...' Without a transfer test, sensitivity analysis with respect to T, backbone capacity, or data distribution, the conclusion that ARTM/FHTM 'achieve generation quality comparable to non-causal methods' is not robust. Please either derive conditions under which Eq. (15) is appropriate, or add experiments that vary these factors.
  2. [Tables 1, 2, and 7] No error bars, confidence intervals, or multiple-seed runs are reported for any of the main results. Several of the differences that support the paper's headline claims are small: for example, DTM vs. FM on PartiPrompts gives CLIPScore 26.8 vs. 26.0 and PickScore 21.2 vs. 21.0, and on GenEval DTM scores 0.54 vs. MAR 0.52 and FM 0.47. Without an estimate of variability, 'state-of-the-art', 'surpasses', and 'outperforms' are not statistically supported. Please add repeated runs, bootstrapped confidence intervals, or at least a variance report for the key comparisons.
  3. [Section 4.1 and Table 1] The claim that 'FHTM is the first fully causal model to match or surpass the performance of flow-based methods on text-to-image task in continuous domains' is not established by the evidence presented. All baselines were re-implemented by the authors, and no comparison is made with published state-of-the-art numbers on the same benchmarks. The word 'first' is an absolute claim that requires a survey of existing published causal models; as written, it should be qualified to 'first among the configurations tested here' or supported by a direct comparison with prior published causal text-to-image systems.
  4. [Section 4.2, Table 3] The 'almost 7-fold speedup' claim compares DTM at 16 backbone forwards plus 64 head forwards against FM at 128 backbone forwards with zero head forwards. The reported DTM time of 1.6s includes the cost of the head, while the FM time of 10.8s does not, because FM has no head. This is not an apples-to-apples comparison of total inference cost. Please report the full inference cost for both methods at matched quality, including the cost of the head and any activation caching, or clearly separate backbone and head costs in the speedup claim.
minor comments (4)
  1. [References] The reference to MaskGIT contains a corrupted author string: 'gledhhnddinerbdilnulnfjWilliam T. Freeman' should be corrected.
  2. [Appendix C] There are several typos in the proof, including 'Lipshcitz' for 'Lipschitz' and inconsistent notation for the random variable Y (defined as X_T - X_0 in the main text and as the normalized increment in Appendix C). Please align the notation.
  3. [Section 2.2, DTM paragraph] The sentence 'Note that in this case one can also learn a continuous time t in [0,T] which allows more flexible sampling' is not elaborated anywhere in the paper; either provide details or remove the claim.
  4. [Algorithms 5 and 6] In ARTM training and sampling, the loop over tokens is described with 'parallel for' but sampling in Algorithm 6 is inherently sequential over i; please clarify which loops are parallel and which are sequential.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TM's definitions and convergence proof are self-contained, and empirical claims rest on re-implemented external baselines.

full rationale

Transition Matching's core derivation is definitional but not circular: the supervising process (Eq. 2) and loss (Eq. 5) define a regression objective independent of the model; DTM's update (Eq. 12) follows algebraically from Eqs. 10-11, and its connection to FM (Eq. 14) is a conditional-expectation identity, not an assumed conclusion. Theorem 1's proof (Appendix C) assumes explicit Lipschitz and quadratic-variation bounds and verifies them for the Gaussian linear process (C.1); it does not assume the convergence it proves. ARTM/FHTM use a newly introduced independent linear process (Eq. 15); its advantage over Eq. 10 is established empirically on re-implemented baselines, with the mechanism explicitly labeled a hypothesis (Appendix A.4). That is a robustness/transfer concern, not a circular reduction. Self-citations (Lipman et al. 2022; Shaul et al. 2023; Holderrieth et al. 2025) appear, but they are background/related-work citations and are not load-bearing: the paper's SOTA claims are compared against re-implemented external baselines (FM, MAR, MAR-Fluid, discrete AR) under fixed data and hyperparameters. No fitted parameter is renamed as a prediction, and no uniqueness theorem is imported from the authors' prior work.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central empirical results depend on a handful of hyperparameters that were tuned by validation (transition steps, head ODE steps, guidance scale, patch size), and on the independent linear process whose benefit is not derived. The convergence theorem assumes bounded data and Lipschitz velocity fields.

free parameters (5)
  • DTM transition steps T = 32
    Chosen via validation; Table 8 shows saturation around 16-32 TM steps.
  • ARTM/FHTM transition steps T = 3
    Evaluated at T=2 and T=3; T=3 gives the best metrics in Tables 1 and 2.
  • Flow head ODE steps (head NFE) = 64
    Chosen from saturation in Figure 8; performance is roughly flat above 16 head NFE.
  • Classifier-free guidance scale = 6.5
    Following Li et al. 2024; sensitivity shown in Figure 20, but 6.5 is used for all main results.
  • Flow head patch size = 2x2x4
    Architecture choice; a 2x8x4 patch variant is ablated in Figure 10 but the main results use 2x2x4.
assumptions (4)
  • standard math Law of total probability is used to introduce latent variable Y in eq. 6 and to factor autoregressive kernels via the probability chain rule.
    Section 2.1 relies on this to decompose transition kernels; it is a standard probability identity.
  • domain assumption Data latent codes are bounded (||X1|| <= r) and the conditional kernel is Gaussian, used to prove Lipschitz and moment bounds in Appendix C.
    Appendix C, equations 60 and 66 assume bounded data to establish the convergence theorem for DTM.
  • domain assumption The flow matching loss (eq. 9) with a Gaussian prior is a valid objective for learning the conditional samplers p^theta_{Y|t}.
    This is a known result from Lipman et al. 2022 and Liu et al. 2022, invoked without new proof.
  • ad hoc to paper The independent linear process (eq. 15) is a suitable supervising process for autoregressive kernels.
    The paper shows it works empirically (Figure 9) but provides no theoretical justification; it is a design choice introduced for ARTM/FHTM.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Transition Matching: Scalable and Flexible Generative Modeling." pith.science (2026). https://pith.science/paper/KJWETYAN

@misc{pith2026250623589,
  author       = {Pith},
  title        = {Pith review of: Transition Matching: Scalable and Flexible Generative Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KJWETYAN}},
  note         = {Machine review of arXiv:2506.23589}
}
read the original abstract

Diffusion and flow matching models have significantly advanced media generation, yet their design space is well-explored, somewhat limiting further improvements. Concurrently, autoregressive (AR) models, particularly those generating continuous tokens, have emerged as a promising direction for unifying text and media generation. This paper introduces Transition Matching (TM), a novel discrete-time, continuous-state generative paradigm that unifies and advances both diffusion/flow models and continuous AR generation. TM decomposes complex generation tasks into simpler Markov transitions, allowing for expressive non-deterministic probability transition kernels and arbitrary non-continuous supervision processes, thereby unlocking new flexible design avenues. We explore these choices through three TM variants: (i) Difference Transition Matching (DTM), which generalizes flow matching to discrete-time by directly learning transition probabilities, yielding state-of-the-art image quality and text adherence as well as improved sampling efficiency. (ii) Autoregressive Transition Matching (ARTM) and (iii) Full History Transition Matching (FHTM) are partially and fully causal models, respectively, that generalize continuous AR methods. They achieve continuous causal AR generation quality comparable to non-causal approaches and potentially enable seamless integration with existing AR text generation techniques. Notably, FHTM is the first fully causal model to match or surpass the performance of flow-based methods on text-to-image task in continuous domains. We demonstrate these contributions through a rigorous large-scale comparison of TM variants and relevant baselines, maintaining a fixed architecture, training data, and hyperparameters.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Transition Matching Distillation for Fast Video Generation

    cs.CV 2026-01 conditional novelty 6.0 of 10

    Splitting a video diffusion model into a fixed feature extractor and a small recurrent flow head lets TMD generate videos in one to two effective steps with better VBench scores than prior distilled models.

Reference graph

Works this paper leans on

54 extracted references · 3 canonical work pages · cited by 1 Pith paper

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Building normalizing flows with stochastic interpolants

    Michael S Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. arXiv preprint arXiv:2209.15571, 2022

    arXiv 2022

  3. [3]

    Chameleon: Mixed-modal early-fusion foundation models, 2025

    Chameleon-Team. Chameleon: Mixed-modal early-fusion foundation models, 2025. https://arxiv.org/abs/2405.09818

    arXiv 2025

  4. [4]

    Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and gledhhnddinerbdcilnulnfjWilliam T. Freeman. Maskgit: Masked generative image transformer, 2022. https://arxiv.org/abs/2202.04200

    arXiv 2022

  5. [5]

    F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching

    Yushen Chen, Zhikang Niu, Ziyang Ma, Keqi Deng, Chunhui Wang, Jian Zhao, Kai Yu, and Xie Chen. F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching. arXiv preprint arXiv:2410.06885, 2024

    arXiv 2024

  6. [6]

    Patch n' pack: Navit, a vision transformer for any aspect ratio and resolution, 2023

    Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim Alabdulmohsin, Avital Oliver, Piotr Padlewski, Alexey Gritsenko, Mario Lučić, and Neil Houlsby. Patch n' pack: Navit, a vision transformer for any aspect ratio and resolution, 2023. https://arxiv.org/...

    arXiv 2023

  7. [7]

    Diffusion models beat gans on image synthesis, 2021

    Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis, 2021. https://arxiv.org/abs/2105.05233

    arXiv 2021

  8. [8]

    Jukebox: A generative model for music, 2020

    Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. Jukebox: A generative model for music, 2020. https://arxiv.org/abs/2005.00341

    arXiv 2020

Show all 54 references
  1. [9]

    Cogview: Mastering text-to-image generation via transformers, 2021

    Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang. Cogview: Mastering text-to-image generation via transformers, 2021. https://arxiv.org/abs/2105.13290

    2021 arXiv

  2. [10]

    Scaling rectified flow transformers for high-resolution image synthesis, 2024

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transf...

    2024 arXiv

  3. [11]

    Fluid: Scaling autoregressive text-to-image generative models with continuous tokens, 2024

    Lijie Fan, Tianhong Li, Siyang Qin, Yuanzhen Li, Chen Sun, Michael Rubinstein, Deqing Sun, Kaiming He, and Yonglong Tian. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens, 2024. https://arxiv.org/abs/2410.13863

    2024 arXiv

  4. [12]

    Geneval: An object-focused framework for evaluating text-to-image alignment, 2023

    Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment, 2023. https://arxiv.org/abs/2310.11513

    2023 arXiv

  5. [13]

    Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis, 2024

    Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis, 2024. https://arxiv.org/abs/2412.04431

    2024 arXiv

  6. [14]

    Clipscore: A reference-free evaluation metric for image captioning, 2022

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning, 2022. https://arxiv.org/abs/2104.08718

    2022 arXiv

  7. [15]

    Denoising diffusion probabilistic models, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. https://arxiv.org/abs/2006.11239

    2020 arXiv

  8. [16]

    Peter Holderrieth, Marton Havasi, Jason Yim, Neta Shaul, Itai Gat, Tommi Jaakkola, Brian Karrer, Ricky T. Q. Chen, and Yaron Lipman. Generator matching: Generative modeling with arbitrary markov processes, 2025. https://arxiv.org/abs/2410.20587

    2025 arXiv

  9. [17]

    Pyramidal flow matching for efficient video generative modeling, 2025

    Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling, 2025. https://arxiv.org/abs/2410.05954

    2025

  10. [18]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35: 0 26565--26577, 2022

    2022

  11. [19]

    Auto-encoding variational bayes, 2013

    Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013

    2013

  12. [20]

    Kingma, Tim Salimans, Ben Poole, and Jonathan Ho

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

    2023 arXiv

  13. [21]

    Pick-a-pic: An open dataset of user preferences for text-to-image generation, 2023

    Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation, 2023. https://arxiv.org/abs/2305.01569

    2023 arXiv

  14. [22]

    Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024

    2024

  15. [23]

    Autoregressive image generation without vector quantization

    Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. Advances in Neural Information Processing Systems, 37: 0 56424--56445, 2024

    2024

  16. [24]

    Lawrence Zitnick, and Piotr Dollár

    Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dollár. Microsoft coco: Common objects in context, 2015. https://arxiv.org/abs/1405.0312

    2015 arXiv

  17. [25]

    Flow matching for generative modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022

    2022 arXiv

  18. [26]

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

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022

    2022 arXiv

  19. [27]

    The llama 3 herd of models, 2024

    Llama 3 Team Meta. The llama 3 herd of models, 2024. https://arxiv.org/abs/2407.21783

    2024 arXiv

  20. [28]

    Improved denoising diffusion probabilistic models, 2021

    Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models, 2021. https://arxiv.org/abs/2102.09672

    2021 arXiv

  21. [29]

    Taming transformers for high-resolution image synthesis, 2021

    Patrick, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis, 2021. https://arxiv.org/abs/2012.09841

    2021 arXiv

  22. [30]

    Scalable diffusion models with transformers, 2023

    William Peebles and Saining Xie. Scalable diffusion models with transformers, 2023. https://arxiv.org/abs/2212.09748

    2023 arXiv

  23. [31]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. https://arxiv.org/abs/2307.01952

    2023 arXiv

  24. [32]

    Movie gen: A cast of media foundation models, 2025

    A Polyak, A Zohar, A Brown, A Tjandra, A Sinha, A Lee, A Vyas, B Shi, CY Ma, CY Chuang, et al. Movie gen: A cast of media foundation models, 2025. URL https://arxiv. org/abs/2410.13720, page 51

    2025 arXiv

  25. [33]

    Learning transferable visual models from natural language supervision, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. https://arxiv....

    2021 arXiv

  26. [34]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer, 2023. https://arxiv.org/abs/1910.10683

    2023 arXiv

  27. [35]

    Zero-shot text-to-image generation, 2021

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation, 2021. https://arxiv.org/abs/2102.12092

    2021 arXiv

  28. [36]

    High-resolution image synthesis with latent diffusion models, 2022

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. https://arxiv.org/abs/2112.10752

    2022 arXiv

  29. [37]

    Laion-5b: An open large-scale dataset for training next generation image-text models, 2022

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion...

    2022 arXiv

  30. [38]

    Neta Shaul, Ricky T. Q. Chen, Maximilian Nickel, Matt Le, and Yaron Lipman. On kinetic optimal probability paths for generative models, 2023. https://arxiv.org/abs/2306.06626

    2023 arXiv

  31. [39]

    Weiss, Niru Maheswaranathan, and Surya Ganguli

    Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics, 2015. https://arxiv.org/abs/1503.03585

    2015 arXiv

  32. [40]

    Denoising diffusion implicit models, 2022

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022. https://arxiv.org/abs/2010.02502

    2022 arXiv

  33. [41]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations, 2021. https://arxiv.org/abs/2011.13456

    2021 arXiv

  34. [42]

    Autoregressive model beats diffusion: Llama for scalable image generation, 2024

    Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation, 2024. https://arxiv.org/abs/2406.06525

    2024 arXiv

  35. [43]

    Ul2: Unifying language learning paradigms

    Yi Tay, Mostafa Dehghani, Vinh Q Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Siamak Shakeri, Dara Bahri, Tal Schuster, et al. Ul2: Unifying language learning paradigms. arXiv preprint arXiv:2205.05131, 2022

    2022 arXiv

  36. [44]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction, 2024

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction, 2024. https://arxiv.org/abs/2404.02905

    2024 arXiv

  37. [45]

    Givt: Generative infinite-vocabulary transformers, 2024

    Michael Tschannen, Cian Eastwood, and Fabian Mentzer. Givt: Generative infinite-vocabulary transformers, 2024. https://arxiv.org/abs/2312.02116

    2024 arXiv

  38. [46]

    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. https://arxiv.org/abs/1706.03762

    2023 arXiv

  39. [47]

    Unified reward model for multimodal understanding and generation, 2025

    Yibin Wang, Yuhang Zang, Hao Li, Cheng Jin, and Jiaqi Wang. Unified reward model for multimodal understanding and generation, 2025. https://arxiv.org/abs/2503.05236

    2025 arXiv

  40. [48]

    Imagereward: Learning and evaluating human preferences for text-to-image generation, 2023 a

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation, 2023 a . https://arxiv.org/abs/2304.05977

    2023 arXiv

  41. [49]

    Restart sampling for improving generative processes

    Yilun Xu, Mingyang Deng, Xiang Cheng, Yonglong Tian, Ziming Liu, and Tommi Jaakkola. Restart sampling for improving generative processes. Advances in Neural Information Processing Systems, 36: 0 76806--76838, 2023 b

    2023

  42. [50]

    Teaching large language models to regress accurate image quality scores using score distribution, 2025

    Zhiyuan You, Xin Cai, Jinjin Gu, Tianfan Xue, and Chao Dong. Teaching large language models to regress accurate image quality scores using score distribution, 2025. https://arxiv.org/abs/2501.11561

    2025

  43. [51]

    Scaling autoregressive models for content-rich text-to-image generation, 2022

    Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-ric...

    2022 arXiv

  44. [52]

    Scaling autoregressive multi-modal models: Pretraining and instruction tuning

    Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multi-modal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2023

    2023 arXiv

  45. [53]

    E-car: Efficient continuous autoregressive image generation via multistage modeling, 2024

    Zhihang Yuan, Yuzhang Shang, Hanling Zhang, Tongcheng Fang, Rui Xie, Bingxin Xu, Yan Yan, Shengen Yan, Guohao Dai, and Yu Wang. E-car: Efficient continuous autoregressive image generation via multistage modeling, 2024. https://arxiv.org/abs/2412.14170

    2024 arXiv

  46. [54]

    Transfusion: Predict the next token and diffuse images with one multi-modal model, 2024

    Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model, 2024. https://arxiv.org/abs/2408.11039

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.