Pith. sign in

REVIEW 4 major objections 3 minor 41 references

Diffusion models trained over as few as 32 latent states—and even a single state, when several are combined—can match models trained over about 1,000 states, given a carefully chosen noise schedule.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

With a carefully selected noise schedule, diffusion models trained with as few as 32 latent states, or composed from single-state models, match 1,000-state training and converge 4-6x faster.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A bold few-step training claim (T=32 ~ T=1000) that I cannot vet from the abstract alone; the noise schedule is the elephant in the room. the 4 major comments →

arxiv 2508.14413 v1 pith:TKER44W6 submitted 2025-08-20 cs.LG cs.CV

Disentanglement in T-space for Faster and Distributed Training of Diffusion Models with Fewer Latent-states

Samarth Gupta , Raghudeep Gadde , Rui Chen , Aleix M. Martinez This is my paper
classification cs.LG cs.CV
keywords diffusion modelsnoise schedulelatent statesT-space disentanglementfew-step generationdistributed trainingsingle latent-state modeltraining convergence
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 challenges a central assumption behind diffusion models: that a large number of time steps (latent states) is needed so the generative reverse process is close to Gaussian. It claims that with a carefully chosen noise schedule, models trained over roughly 32 latent states match the quality of models trained over roughly 1,000, and that even a single latent state suffices when several independently trained single-state models are combined. This 'complete disentanglement in T-space' turns each noise level into its own small denoising problem, so the pieces can be trained separately and composed at generation time. The payoff, if correct, is 4–6× faster convergence on the two datasets tested, with a training scheme that distributes naturally across workers. The load-bearing premise is the schedule itself: the paper's result stands only if that schedule is a general design choice, not a fitted knob.

Core claim

The paper's core claim is that the need for T≈1,000 latent states is an artifact of the noise schedule and discretization, not a fundamental property of diffusion. By choosing the schedule carefully, T≈32 reaches parity with T≈1,000; going further, a single latent-state model can be trained to invert one noise level, and multiple such models can be combined to generate high-quality samples. The authors call this complete disentanglement in T-space. Their experiments on two datasets report 4–6× faster convergence relative to standard large-T training across several metrics.

What carries the argument

The central object is the 'disentangled T-space' decomposition: instead of one model spanning all T diffusion steps, independent models are trained for individual latent states (single noise levels) and then combined. A carefully selected noise schedule is the enabling mechanism; it controls how much information each one-step transition must recover, making each single-state subproblem learnable in isolation.

Load-bearing premise

The result depends on the 'careful selection' of the noise schedule; if that schedule only works because it was tuned on the same two datasets used for evaluation, the speedup will not transfer.

What would settle it

Train the disentangled model on a third dataset, such as ImageNet-64, using the paper's noise schedule chosen before training and without per-dataset tuning. If the single-latent-state combined model cannot reach the T≈1,000 baseline's FID within six times the compute budget, the parity claim is falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Training cost drops by a factor of 4–6× in convergence time because each latent-state model is simpler and can be trained independently.
  • T≈32 becomes a practical training regime, so few-step generation no longer requires post-hoc distillation or auxiliary samplers.
  • The model partitions naturally across workers: one worker per latent state enables distributed training without complex communication.
  • Large-T discretization is a modeling choice, not a requirement; research attention shifts to noise-schedule design rather than increasing step count.
  • The independently trained single-state models can serve as building blocks for hierarchical or parallel generative pipelines.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, the actual contribution is the schedule itself; a principled rule for constructing it from data statistics would extend the result beyond the two datasets tested.
  • The combined one-state models likely trade off sample diversity or high-frequency detail, so testing on large natural-image benchmarks where T≈1,000 is the strong baseline would clarify how far the parity claim generalizes.
  • If the result holds at scale, few-step samplers and consistency-style models may be understood as rediscovering what a well-chosen noise schedule already provides.
  • A testable extension is applying the same schedule to a text-to-image diffusion backbone and measuring quality and speed at T≈32; the two datasets in this paper do not establish scale transfer.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 3 minor

Summary. The paper challenges the assumption that diffusion models require a large number of latent states / time steps (T ~ 1000). It claims that, with careful selection of a noise schedule, T ~ 32 matches the performance of T ~ 1000; it further claims that T can be pushed to a single latent state by independently training several single-latent-state models and combining them into a disentangled model, yielding 4-6x faster convergence. The abstract reports extensive experiments on two datasets. However, the supplied full text is unreadable because of character-encoding corruption, so no derivation, schedule specification, composition rule, dataset names, baseline details, or error bars can be inspected. The central claims are therefore not verifiable from the provided material.

Significance. If substantiated, the paper would constitute a notable conceptual result: it would show that the large-T requirement in diffusion training is not fundamental and that small-T models can be trained independently and composed, with practical implications for faster and distributed training. The claims are also falsifiable and would be useful to the community. However, the paper as supplied contains no inspectable evidence: there are no visible equations, experimental tables, or algorithm descriptions. The significance is conditional on a readable manuscript that supports the abstract's claims.

major comments (4)
  1. [Full text] The entire full text is corrupted and unreadable; no section, equation, table, or algorithm can be inspected. This is load-bearing for all central claims: the T=32 parity claim, the single-latent-state composition claim, and the 4-6x speedup claim all require supporting derivations and experiments that are absent from the provided text. The manuscript must be resubmitted in a readable form before any substantive review can occur.
  2. [Abstract, first claim] The parity claim T=32 vs T=1000 is explicitly conditional on "careful selection of a noise schedule." The schedule itself, its parameterization, and the selection rule are not specified anywhere in the accessible text. This is exactly the fitting-type circularity concern: if the schedule was chosen on the same two evaluation datasets until parity was reached, the result does not establish that small T suffices in general. The authors need to state the schedule formula, the selection protocol, and ideally demonstrate transfer to held-out datasets or datasets not used in schedule development.
  3. [Abstract, second claim] The claim that "several independently trained single latent-state models" can be combined into a valid disentangled model is stated without any algorithmic or mathematical construction. There is no description of the composition rule, no sampling procedure, and no argument that the composed model approximates the target reverse process. Without this construction, the central T=1 result cannot be assessed. A precise definition of "complete disentanglement in T-space" and a formal composition rule are required.
  4. [Abstract, speedup claim] The claim of "4-6x faster convergence measured across a variety of metrics on two different datasets" is not backed by any visible table, figure, or baseline definition. No dataset names, metrics, error bars, or comparison methods are given. The speedup claim is central to the paper's practical contribution, and it must be supported by reproducible experimental details, including baselines, hyperparameters, and variance estimates.
minor comments (3)
  1. [Abstract] Typo: "much large number" should be "much larger number".
  2. [Abstract / terminology] The terms "T-space disentanglement" and "complete disentanglement in T-space" are used without definitions. They should be formally introduced, since they carry the paper's conceptual novelty.
  3. [General] No references or comparison to prior work are visible in the accessible text. A proper introduction and related-work section are needed once the manuscript is readable.

Circularity Check

0 steps flagged

No demonstrated circularity; the noise-schedule condition is an experimental design variable, not a fitted input renamed as a prediction.

full rationale

The only clearly readable portion of the manuscript is the abstract, which makes an explicitly conditional empirical claim: with careful selection of a noise schedule, T~32 matches T~1000, and a combination of single latent-state models can generate high-quality samples. A noise schedule is a standard hyperparameter/design choice in diffusion models; the claim is stated as conditional on that choice rather than as a schedule-free necessity, so it does not by construction reduce to a fitted parameter. The T=1 'complete disentanglement' is described as obtained by combining several independently trained single latent-state models; this is a compositional construction claim, not an equation-level reduction of the conclusion to the premise. No self-citation chain, imported uniqueness theorem, or ansatz-smuggling citation is legible in the provided text. The full text is heavily corrupted, so no equations or derivations can be inspected to exhibit a specific reduction. The unspecified schedule selection rule and missing dataset details are transparency and transferability concerns, but under the hard rules they do not by themselves demonstrate circularity. Therefore the appropriate finding is no significant circularity, with a minor score reflecting the unresolved dependence on an undisclosed schedule choice.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 1 invented entities

Listed items are extracted from the abstract because the full text was unreadable. The noise schedule is the main undisclosed free parameter. The T-space disentanglement framing is a postulated idea without independent evidence visible in the abstract.

free parameters (2)
  • noise schedule hyperparameters = not disclosed in abstract
    The abstract credits "careful selection of a noise schedule" for the T=32 and T=1 results. The schedule's shape, noise levels, and discretization are degrees of freedom that must be fixed to reproduce the reported 4-6x speedup, but they are not specified. If the schedule was tuned on the evaluation benchmarks, it is a fitted quantity.
  • number of latent states T = 32 and 1
    T=32 (parity claim) and T=1 (complete disentanglement) are values chosen empirically rather than derived from a bound. The abstract gives no criterion for selecting them.
axioms (3)
  • domain assumption Diffusion training operates over a discrete set of latent states (time steps) forming a Markov chain over those states.
    The entire framing of the paper (T latent states, T-space, single latent-state models, composing models across T) presumes the standard discrete Markov-chain diffusion formulation.
  • ad hoc to paper A well-chosen noise schedule can compensate for the discretization induced by very few latent states.
    This is the core enabling premise of the paper: schedule selection substitutes for step count. If this fails, the T=32 parity claim and the T=1 composition claim both fail. It is asserted in the abstract and not derived there.
  • ad hoc to paper Several independently trained single latent-state models can be composed into a valid reverse sampler without any joint training.
    The abstract asserts the disentangled model is obtained by combining several independently trained single latent-state models. Composability is assumed, not demonstrated at the abstract level, and no combination rule is given.
invented entities (1)
  • T-space disentanglement / complete disentanglement in T-space no independent evidence
    purpose: Conceptual framing asserting that each latent state (time step) can be learned independently as a separate single-step model, enabling parallel and distributed training with few latent states.
    This is a new postulated organizing concept introduced by the paper. The abstract provides no falsifiable handle outside the paper: no formal definition, no prediction, and no metric for measuring disentanglement are visible at the abstract level.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Disentanglement in T-space for Faster and Distributed Training of Diffusion Models with Fewer Latent-states." pith.science (2026). https://pith.science/paper/TKER44W6

@misc{pith2026250814413,
  author       = {Pith},
  title        = {Pith review of: Disentanglement in T-space for Faster and Distributed Training of Diffusion Models with Fewer Latent-states},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TKER44W6}},
  note         = {Machine review of arXiv:2508.14413}
}
Share X Bluesky LinkedIn Reddit HN
abstract

We challenge a fundamental assumption of diffusion models, namely, that a large number of latent-states or time-steps is required for training so that the reverse generative process is close to a Gaussian. We first show that with careful selection of a noise schedule, diffusion models trained over a small number of latent states (i.e. $T \sim 32$) match the performance of models trained over a much large number of latent states ($T \sim 1,000$). Second, we push this limit (on the minimum number of latent states required) to a single latent-state, which we refer to as complete disentanglement in T-space. We show that high quality samples can be easily generated by the disentangled model obtained by combining several independently trained single latent-state models. We provide extensive experiments to show that the proposed disentangled model provides 4-6$\times$ faster convergence measured across a variety of metrics on two different datasets.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 27 canonical work pages

  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]

    ediff-i: Text-to-image diffusion models with ensemble of expert denoisers

    Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022

    Pith/arXiv arXiv 2022

  3. [3]

    Reproducible scaling laws for contrastive language-image learning

    Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818--2829, 2023

    2023

  4. [4]

    Perception prioritized training of diffusion models

    Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11462--11471, 2022

    work page 2022

  5. [5]

    NVIDIA NVLink Interconnect

    NVIDIA Corporation. NVIDIA NVLink Interconnect . https://www.nvidia.com/en-us/data-center/nvlink/, 2023

    work page 2023

  6. [6]

    Socher, Li Fei-Fei, Wei Dong, Kai Li, and Li-Jia Li

    Jia Deng, R. Socher, Li Fei-Fei, Wei Dong, Kai Li, and Li-Jia Li. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition(CVPR), pages 248--255, 2009

    work page 2009

  7. [7]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In Proceedings of the 35th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2021. Curran Associates Inc

    work page 2021

  8. [8]

    Taming transformers for high-resolution image synthesis, 2020

    Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis, 2020

    work page 2020

  9. [9]

    Patrick Esser, Sumith Kulal, A. Blattmann, Rahim Entezari, Jonas Muller, 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 transformers for high-resolution image synthesis. ArXiv, abs/2403.03206, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    W. Feller . On the Theory of Stochastic Processes, with Particular Reference to Applications . In First Berkeley Symposium on Mathematical Statistics and Probability, pages 403--432, 1949

    work page 1949

  11. [11]

    Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts

    Zhida Feng, Zhenyu Zhang, Xintong Yu, Yewei Fang, Lanxin Li, Xuyi Chen, Yuxiang Lu, Jiaxiang Liu, Weichong Yin, Shi Feng, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages...

    work page 2023

  12. [12]

    Masked diffusion transformer is a strong image synthesizer

    Shanghua Gao, Pan Zhou, Ming - Ming Cheng, and Shuicheng Yan. Masked diffusion transformer is a strong image synthesizer. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023 , pages 23107--23116. IEEE , 2023

    work page 2023

  13. [13]

    Efficient diffusion training via min-snr weighting strategy

    Tiankai Hang, Shuyang Gu, Chen Li, Jianmin Bao, Dong Chen, Han Hu, Xin Geng, and Baining Guo. Efficient diffusion training via min-snr weighting strategy. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7441--7451, 2023

    work page 2023

  14. [14]

    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. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017

    work page 2017

  15. [15]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239, 2020

    Pith/arXiv arXiv 2006

  16. [16]

    Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering

    Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. arXiv preprint arXiv:2303.11897, 2023

    Pith/arXiv arXiv 2023

  17. [17]

    Rethinking fid: Towards a better evaluation metric for image generation

    Sadeep Jayasumana, Srikumar Ramalingam, Andreas Veit, Daniel Glasner, Ayan Chakrabarti, and Sanjiv Kumar. Rethinking fid: Towards a better evaluation metric for image generation. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9307--9315, 2024

    work page 2024

  18. [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. In Proc. NeurIPS, 2022

    2022

  19. [19]

    Manmatha, Ashwin Swaminathan, Zhuowen Tu, Stefano Ermon, and Stefano Soatto

    Hao Li, Yang Zou, Ying Wang, Orchid Majumder, Yusheng Xie, R. Manmatha, Ashwin Swaminathan, Zhuowen Tu, Stefano Ermon, and Stefano Soatto. On the scalability of diffusion-based text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9400--9409, 2024

    work page 2024

  20. [20]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning, 2023

    2023

  21. [21]

    Belongie, Lubomir D

    Tsung - Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll ' a r, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. CoRR, abs/1405.0312, 2014

    Pith/arXiv arXiv 2014

  22. [22]

    Learning in implicit generative models

    Shakir Mohamed and Balaji Lakshminarayanan. Learning in implicit generative models. ArXiv, abs/1610.03483, 2016

    Pith/arXiv arXiv 2016

  23. [23]

    Switch diffusion transformer: Synergizing denoising tasks with sparse mixture-of-experts

    Byeongjun Park, Hyojun Go, Jin-Young Kim, Sangmin Woo, Seokil Ham, and Changick Kim. Switch diffusion transformer: Synergizing denoising tasks with sparse mixture-of-experts. In Computer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Proceedings, Part LIII, page 461–477, Berlin, Heidelberg, 2024. Springer-Verlag

    work page 2024

  24. [24]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4195--4205, 2023

    work page 2023

  25. [25]

    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

    2023

  26. [26]

    Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever

    Alec Radford, Jong Wook Kim, Chris Hallacy, A. 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. In ICML, 2021

    work page 2021

  27. [27]

    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 Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    work page 2022

  28. [28]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2015, pages 234--241. Springer International Publishing, 2015

    work page 2015

  29. [29]

    Pyramidal denoising diffusion probabilistic models

    Dohoon Ryu and Jong-Chul Ye. Pyramidal denoising diffusion probabilistic models. ArXiv, abs/2208.01864, 2022

    Pith/arXiv arXiv 2022

  30. [30]

    Improved techniques for training gans

    Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016

    work page 2016

  31. [31]

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

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION -5b: An open large-scale dataset for training next generation image-text mo...

    work page 2022

  32. [32]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the 32nd International Conference on Machine Learning, pages 2256--2265, Lille, France, 2015. PMLR

    work page 2015

  33. [33]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020

    Pith/arXiv arXiv 2010

  34. [34]

    A closer look at time steps is worthy of triple speed-up for diffusion model training, 2024

    Kai Wang, Mingjia Shi, Yukun Zhou, Zekai Li, Zhihang Yuan, Yuzhang Shang, Xiaojiang Peng, Hanwang Zhang, and Yang You. A closer look at time steps is worthy of triple speed-up for diffusion model training, 2024

    work page 2024

  35. [35]

    Patch diffusion: Faster and more data-efficient training of diffusion models

    Zhendong Wang, Yifan Jiang, Huangjie Zheng, Peihao Wang, Pengcheng He, Zhangyang "Atlas" Wang, Weizhu Chen, and Mingyuan Zhou. Patch diffusion: Faster and more data-efficient training of diffusion models. In Advances in Neural Information Processing Systems, pages 72137--72154. Curran Associates, Inc., 2023

    work page 2023

  36. [36]

    Tackling the generative learning trilemma with denoising diffusion gans

    Zhisheng Xiao, Karsten Kreis, and Arash Vahdat. Tackling the generative learning trilemma with denoising diffusion gans. In International Conference on Learning Representations, 2022

    work page 2022

  37. [37]

    Towards faster training of diffusion models: An inspiration of a consistency phenomenon

    Tianshuo Xu, Peng Mi, Ruilin Wang, and Yingcong Chen. Towards faster training of diffusion models: An inspiration of a consistency phenomenon. ArXiv, abs/2404.07946, 2024

    Pith/arXiv arXiv 2024

  38. [38]

    Truncated diffusion probabilistic models and diffusion-based adversarial auto-encoders

    Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Truncated diffusion probabilistic models and diffusion-based adversarial auto-encoders. In The Eleventh International Conference on Learning Representations, 2023

    work page 2023

  39. [39]

    Fast training of diffusion models with masked transformers

    Hongkai Zheng, Weili Nie, Arash Vahdat, and Anima Anandkumar. Fast training of diffusion models with masked transformers. In Transactions on Machine Learning Research (TMLR), 2024 a

    work page 2024

  40. [40]

    Non-uniform timestep sampling: Towards faster diffusion model training

    Tianyi Zheng, Cong Geng, Peng-Tao Jiang, Ben Wan, Hao Zhang, Jinwei Chen, Jia Wang, and Bo Li. Non-uniform timestep sampling: Towards faster diffusion model training. In Proceedings of the 32nd ACM International Conference on Multimedia, page 7036–7045, New York, NY, USA, 2024 b . Association for Computing Machinery

    work page 2024

  41. [41]

    Beta-tuned timestep diffusion model

    Tianyi Zheng, Peng-Tao Jiang, Ben Wan, Hao Zhang, Jinwei Chen, Jia Wang, and Bo Li. Beta-tuned timestep diffusion model. In Computer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Proceedings, Part III, page 114–130, Berlin, Heidelberg, 2024 c . Springer-Verlag

    work page 2024

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.