Pith. sign in

REVIEW 3 major objections 3 minor 1 cited by

OmniCache accelerates diffusion transformer sampling by reusing cached activations across the full denoising trajectory, estimating and filtering out the noise at each reuse step.

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 →

A training-free cache-reuse scheme that spreads computation across the full diffusion trajectory and subtracts estimated noise, accelerating DiT sampling with claimed competitive quality.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Plausible new twist on training-free cache reuse, but the abstract alone can't carry the 'competitive quality' claim; the noise-filter trick is the spot to probe. the 3 major comments →

arxiv 2508.16212 v2 pith:4PZH7T2R submitted 2025-08-22 cs.CV cs.AIcs.LG

OmniCache: A Trajectory-Oriented Global Perspective on Training-Free Cache Reuse for Diffusion Transformer Models

classification cs.CV cs.AIcs.LG
keywords diffusion transformerstraining-free accelerationcache reusedenoising trajectorynoise estimationimage generationsampling acceleration
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

This paper tries to show that diffusion transformer sampling can be made substantially faster without any training by exploiting redundancy that spans the whole denoising trajectory, not just between similar neighboring steps. The proposed method, OmniCache, analyzes the sampling trajectory of a diffusion transformer and decides which steps should recompute and which should reuse previously cached activations, spreading the reuse across the entire process instead of concentrating it in the late, more similar steps. At every reuse point it dynamically estimates the noise component that has accumulated in the cached activation and filters it out, so the reused value does not steer the sampling in the wrong direction. If correct, this means pretrained diffusion transformers for image and video generation can be deployed more cheaply, at higher throughput and lower latency, while keeping output quality competitive with full sampling.

Core claim

The authors claim that the denoising process in diffusion transformers has global redundancy: intermediate activations at one step remain informative far beyond the immediately adjacent steps, and existing caching methods miss most of this by only reusing cache where consecutive steps look similar, mostly late in the process. OmniCache is their training-free remedy. It treats the sampling run as a trajectory, distributes cache-reuse points across the whole trajectory, and when a step reuses a cached activation, it estimates the noise that has been added since the activation was computed and removes it before feeding the activation into the next layer. The paper reports that this accelerates

What carries the argument

The two load-bearing pieces are the trajectory-level cache-reuse schedule and the noise-filtering reuse rule. The schedule looks at the sampling trajectory of the diffusion transformer and selects which denoising steps are recomputed and which steps reuse cached activations, deliberately distributing reuse across early, middle, and late steps rather than clustering it where inter-step similarity is high. The reuse rule treats a cached activation as the true value plus an accumulated noise term; at reuse time it dynamically estimates that noise and subtracts it, so the model's next update receives a denoised approximation of the activation it would have computed. Together they make stale acti

Load-bearing premise

The load-bearing premise is that a cached activation, after the dynamically estimated noise is subtracted, is close enough to the activation the model would have computed at that step that reusing it does not steer the sampling off course.

What would settle it

On a fixed diffusion transformer and sampler, keep one reuse schedule and disable the noise filter while holding all seeds fixed: if the quality of generated images does not change, the filter is inert and the claimed mechanism is wrong; if the filtered cached activations have low cosine similarity to the activations from full recomputation at early and middle reuse steps, the global-reuse premise fails.

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

If this is right

  • Pretrained diffusion transformers can be accelerated without weight updates or extra training, so the method applies to existing models as-is.
  • Cache reuse no longer has to be confined to late-step segments; a global reuse schedule can capture more of the available computation savings.
  • The noise-estimation filter is what makes early and mid-trajectory reuse safe; without it, reusing stale activations would push the sampling direction off course.
  • Faster sampling at competitive quality removes a practical obstacle to real-time image and video generation with transformer-based diffusion models.

Where Pith is reading between the lines

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

  • The global schedule could be tuned per model, per sampler, or even per input; the paper proposes one analysis, but a learned or adaptive schedule might extract larger gains.
  • The noise-estimation-and-filter trick is not obviously limited to transformer diffusion: sequential generative models with known noise schedules could adopt the same correction when reusing hidden states.
  • OmniCache is orthogonal to training-based acceleration, so combining it with distillation or step-reduction methods may compound the speedup.
  • A direct empirical test of the core mechanism would be to compare the filtered cached activation against the true next-step activation: the gap should stay small across all scheduled reuse points.
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

3 major / 3 minor

Summary. OmniCache proposes a training-free cache reuse method for diffusion Transformer (DiT) sampling. The abstract claims that, unlike prior methods that concentrate cache reuse on later sampling steps, OmniCache distributes reuse across the entire denoising trajectory from a global perspective. During reuse it dynamically estimates the noise contained in cached activations and filters it out to preserve the sampling direction. The abstract further claims that extensive experiments show accelerated sampling with competitive generative quality. However, the supplied full text is an unreadable encoded/mojibake document, so the formal algorithm, equations, figures, and experimental tables could not be inspected in this review.

Significance. If the claims are substantiated, the method could be a practically useful contribution to efficient DiT inference, extending the line of training-free cache reuse methods with a trajectory-oriented schedule and a noise-filtering step. The idea is plausible and potentially novel. However, as submitted, the manuscript does not allow the claims to be verified: no quantitative results, baselines, ablations, definitions, or derivations are visible in the readable portion, and the full text is corrupted. The main empirical risk identified in the stress-test note—that the residual drift after noise filtering is not necessarily small and may compound over the trajectory—is directly relevant and is not addressed by the abstract. Thus, while the direction is promising, the significance cannot currently be assessed beyond the generic value of training-free acceleration.

major comments (3)
  1. [Full text (encoding)] The body of the manuscript is supplied as an unreadable character-encoded text (mojibake) from the introduction through the final sections. Equations, figures, tables, and the experimental section cannot be parsed. This blocks verification of the central empirical claim ('extensive experiments demonstrate...') and of the algorithm's definitions. The authors must provide a readable, correctly encoded manuscript.
  2. [Abstract] The load-bearing premise is that 'dynamically estimate the corresponding noise and filter it out' suffices to make stale cached activations approximate fresh ones. Hidden-state drift in DiTs comes not only from scheduler noise but from changing timestep embeddings, attention maps, and accumulated denoising directions; a noise estimator aimed at scheduler noise will not remove this residual. The paper needs a formal definition of the estimated noise and filter (with equations), per-layer/per-interval residual error statistics, and an ablation isolating the filter's effect on generative quality. Without these, the competitive-quality claim is unsupported.
  3. [Experiments (absent)] No quantitative experimental material is visible: no tables, no baselines, no model/dataset specifications, no speed/quality Pareto curves, and no comparisons with prior training-free cache methods. The abstract's 'competitive generative quality' is therefore an assertion. The revision should report quality metrics versus wall-clock time or FLOPs, compare with existing cache-reuse baselines, vary cache interval placement, and show the effect of the noise filter across a range of cache rates.
minor comments (3)
  1. [Abstract] 'DIT' should be written as 'DiT' (Diffusion Transformer) for consistency with the literature.
  2. [Abstract] The abstract should state the evaluated architectures, datasets, and step counts explicitly; currently it only promises 'extensive experiments' without any concrete context.
  3. [Full text (rendering)] The submitted text is corrupted beyond readability. If this is a PDF/TeX encoding issue, it should be fixed before any further review; the current version is not usable.

Circularity Check

0 steps flagged

No circularity found: the abstract presents a training-free heuristic with no derivation that reduces to its inputs.

full rationale

The paper's abstract makes no derivation claim that is equivalent to its inputs by construction. OmniCache is described as a training-free acceleration heuristic: it reuses cached activations across the diffusion trajectory and 'dynamically estimate[s] the corresponding noise and filter[s] it out.' This is an implemented heuristic, not a prediction derived from fitted parameters, and no target quantity is defined in terms of the paper's own output. No equation, no fitted constant, and no cited uniqueness theorem appears in the abstract, so there is no self-definitional step, no fitted input renamed as prediction, and no load-bearing self-citation to exhibit. The supplied full text is corrupted/encoded and unreadable, so no further equation-level or citation-level circularity can be quoted or verified. Under the rule that circularity must be demonstrated by a specific quote and reduction, none is found here. The empirical risks highlighted by the reader's take, such as the accuracy of the noise filter and transfer across schedulers or model scales, are correctness concerns rather than circularity. The absence of a readable body prevents a higher-confidence audit, but it also prevents any claim of circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

Limited to abstract-only review. The two listed free parameters are inferred, not confirmed: trajectory cache methods typically require placement thresholds, and the noise filter implies a tunable strength. Neither is documented in the abstract, and if tuned on evaluation benchmarks they would constitute circularity.

free parameters (2)
  • Cache placement and reuse thresholds
    Any trajectory-wide reuse rule needs per-step or per-stage thresholds (e.g., distance cutoffs for deciding when to reuse); the abstract reports none. Whether these are tuned on the test set is not visible.
  • Noise estimation strength or filtering coefficient
    The dynamic noise filtering step implies a coefficient or estimator; its derivation and any tuning are not described in the abstract.
axioms (3)
  • domain assumption The denoising trajectory exhibits exploitable global redundancy across the full sampling process, not just between neighboring steps.
    This is the premise of the method, stated in the abstract ('global redundancy inherent in the denoising process'). If false, reusing distant timestep computations would corrupt outputs.
  • domain assumption The mismatch between cached and true states can be modeled as noise that is dynamically estimable and filterable without biasing the sampling direction.
    The abstract's second mechanism ('dynamically estimate the corresponding noise and filter it out') relies on this. It is the load-bearing assumption behind the quality-preservation claim.
  • domain assumption Empirical gains measured on the tested DiT models, schedulers, and datasets transfer to other scales and step counts.
    The abstract claims acceleration 'while maintaining competitive generative quality' without stating the tested configurations; generalization across settings is assumed.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of OmniCache: A Trajectory-Oriented Global Perspective on Training-Free Cache Reuse for Diffusion Transformer Models." pith.science (2026). https://pith.science/paper/4PZH7T2R

@misc{pith2026250816212,
  author       = {Pith},
  title        = {Pith review of: OmniCache: A Trajectory-Oriented Global Perspective on Training-Free Cache Reuse for Diffusion Transformer Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4PZH7T2R}},
  note         = {Machine review of arXiv:2508.16212}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Diffusion models have emerged as a powerful paradigm for generative tasks such as image synthesis and video generation, with Transformer architectures further enhancing performance. However, the high computational cost of diffusion Transformers-stemming from a large number of sampling steps and complex per-step computations-presents significant challenges for real-time deployment. In this paper, we introduce OmniCache, a training-free acceleration method that exploits the global redundancy inherent in the denoising process. Unlike existing methods that determine caching strategies based on inter-step similarities and tend to prioritize reusing later sampling steps, our approach originates from the sampling perspective of DIT models. We systematically analyze the model's sampling trajectories and strategically distribute cache reuse across the entire sampling process. This global perspective enables more effective utilization of cached computations throughout the diffusion trajectory, rather than concentrating reuse within limited segments of the sampling procedure. In addition, during cache reuse, we dynamically estimate the corresponding noise and filter it out to reduce its impact on the sampling direction. Extensive experiments demonstrate that our approach accelerates the sampling process while maintaining competitive generative quality, offering a promising and practical solution for efficient deployment of diffusion-based generative models.

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Trajectory-Consistent Calibration for Cache-Accelerated Diffusion Models

    cs.CV 2026-05 unverdicted novelty 6.0

    TCC calibrates cached representations in diffusion sampling via an offline iterative procedure that accounts for trajectory shifts, improving FID from 29.83 to 27.35 on PixArt-alpha while preserving reuse policies.

Reference graph

Works this paper leans on

62 extracted references · 27 canonical work pages · cited by 1 Pith paper · 1 internal anchor

  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]

    Estimating the optimal covariance with imperfect mean in diffusion probabilistic models

    Fan Bao, Chongxuan Li, Jiacheng Sun, Jun Zhu, and Bo Zhang. Estimating the optimal covariance with imperfect mean in diffusion probabilistic models. arXiv preprint arXiv:2206.07309, 2022

  3. [3]

    All are worth words: A vit backbone for diffusion models

    Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22669--22679, 2023

  4. [4]

    Video generation models as world simulators

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. Unpublished, 2024

  5. [5]

    Large-vocabulary 3d diffusion model with transformer

    Ziang Cao, Fangzhou Hong, Tong Wu, Liang Pan, and Ziwei Liu. Large-vocabulary 3d diffusion model with transformer. arXiv preprint arXiv:2309.07920, 2023

  6. [6]

    On the trajectory regularity of ode-based diffusion sampling

    Defang Chen, Zhenyu Zhou, Can Wang, Chunhua Shen, and Siwei Lyu. On the trajectory regularity of ode-based diffusion sampling. In International Conference on Machine Learning, pages 7905--7934. PMLR, 2024 a

  7. [7]

    Tgate: Accelerating cross-attention in diffusion models via token-level cache reuse

    Michael Chen, Ning Zhao, and Wei Li. Tgate: Accelerating cross-attention in diffusion models via token-level cache reuse. Advances in Neural Information Processing Systems, 37, 2024 b

  8. [8]

    \( \)-dit: A training-free acceleration method tailored for diffusion transformers

    Pengtao Chen, Mingzhu Shen, Peng Ye, Jianjian Cao, Chongjun Tu, Christos-Savvas Bouganis, Yiren Zhao, and Tao Chen. \( \)-dit: A training-free acceleration method tailored for diffusion transformers. arXiv preprint arXiv:2406.01125, 2024 c

  9. [9]

    Gentron: Delving deep into diffusion transformers for image and video generation

    Shoufa Chen, Mengmeng Xu, Jiawei Ren, Yuren Cong, Sen He, Yanping Xie, Animesh Sinha, Ping Luo, Tao Xiang, and Juan-Manuel Perez-Rua. Gentron: Delving deep into diffusion transformers for image and video generation. arXiv preprint arXiv:2312.04557, 2023

  10. [10]

    Qncd: Quantization noise correction for diffusion models

    Huanpeng Chu, Wei Wu, Chengjie Zang, and Kun Yuan. Qncd: Quantization noise correction for diffusion models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10995--11003, 2024

  11. [11]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34, 2021

  12. [12]

    Structural pruning for diffusion models

    Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. Advances in Neural Information Processing Systems, 36, 2024

  13. [13]

    Generative adversarial networks

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63 0 (11): 0 139--144, 2020

  14. [14]

    Ptqd: Accurate post-training quantization for diffusion models

    Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024

  15. [15]

    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, pages 6626--6637, 2017

  16. [16]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33 0 (6): 0 6840--6851, 2020

  17. [17]

    Autoregressive diffusion models

    Emiel Hoogeboom, Alexey A Gritsenko, Jasmijn Bastings, Ben Poole, Rianne van den Berg, and Tim Salimans. Autoregressive diffusion models. arXiv preprint arXiv:2110.02037, 2021

  18. [18]

    Vbench: A comprehensive benchmark for video generation

    Ziqi Huang, Yinan He, Jiashuo Yu, Chenyang Zhang, et al. Vbench: A comprehensive benchmark for video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

  19. [19]

    Schuller

    Xin Jing, Yi Chang, Zijiang Yang, Jiangjian Xie, Andreas Triantafyllopoulos, and Bjoern W. Schuller. U-dit tts: U-diffusion vision transformer for text-to-speech. Proceedings of the ITG Conference on Speech Communication, pages 56--60, 2023

  20. [20]

    Adaptive caching for faster video generation with diffusion transformers

    Kumara Kahatapitiya, Haozhe Liu, Sen He, Ding Liu, Menglin Jia, Chenyang Zhang, Michael S Ryoo, and Tian Xie. Adaptive caching for faster video generation with diffusion transformers. arXiv preprint arXiv:2411.02397, 2024

  21. [21]

    Bk-sdm: Architecturally compressed stable diffusion for efficient text-to-image generation

    Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, and Shinkook Choi. Bk-sdm: Architecturally compressed stable diffusion for efficient text-to-image generation. Proceedings of the Workshop on Efficient Systems for Foundation Models at ICML2023, 2023

  22. [22]

    Kv caching in transformers for efficient inference

    Eric Kim and Frank Lee. Kv caching in transformers for efficient inference. Proceedings of the International Conference on Learning Representations, 2020

  23. [23]

    Improved precision and recall metric for assessing generative models

    Tuomas Kynk \"a nniemi, Tero Karras, Samuli Laine, and Timo Aila. Improved precision and recall metric for assessing generative models. In Advances in Neural Information Processing Systems, 2019

  24. [24]

    Q-align: A quality alignment metric for image-to-video generation

    Xi Li, Yifan Zhang, and Jie Chen. Q-align: A quality alignment metric for image-to-video generation. arXiv preprint arXiv:2303.XXXX, 2023

  25. [25]

    Snapfusion: Text-to-image diffusion model on mobile devices within two seconds

    Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. Advances in Neural Information Processing Systems, 36, 2024

  26. [26]

    Timestep embedding tells: It's time to cache for video diffusion model

    Feng Liu, Shiwei Zhang, Xiaofeng Wang, Yujie Wei, Haonan Qiu, Yuzhong Zhao, Yingya Zhang, Qixiang Ye, and Fang Wan. Timestep embedding tells: It's time to cache for video diffusion model. arXiv preprint arXiv:2411.19108, 2024

  27. [27]

    Vit-tts: Visual text-to-speech with scalable diffusion transformer

    Huadai Liu, Rongjie Huang, Xuan Lin, Wenqiang Xu, Maozong Zheng, Hong Chen, Jinzheng He, and Zhou Zhao. Vit-tts: Visual text-to-speech with scalable diffusion transformer. arXiv preprint arXiv:2305.12708, 2023

  28. [28]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35: 0 5775--5787, 2022

  29. [29]

    Vdt: General-purpose video diffusion transformers via mask modeling

    Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. Vdt: General-purpose video diffusion transformers via mask modeling. International Conference on Learning Representations, 2023

  30. [30]

    Knowledge distillation in iterative generative models for improved sampling speed

    Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021

  31. [31]

    Latent consistency models: Synthesizing high-resolution images with few-step inference

    Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023

  32. [32]

    Accelerating diffusion models via early stop of the diffusion process

    Zhaoyang Lyu, Xudong Xu, Ceyuan Yang, Dahua Lin, and Bo Dai. Accelerating diffusion models via early stop of the diffusion process. arXiv preprint arXiv:2205.12524, 2022

  33. [33]

    Deepcache: Accelerating diffusion models for free

    Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. arXiv preprint arXiv:2312.00858, 2023

  34. [34]

    Learning-to-cache: Accelerating diffusion transformer via layer caching

    Xinyin Ma, Gongfan Fang, Michael Bi Mi, and Xinchao Wang. Learning-to-cache: Accelerating diffusion transformer via layer caching. In Proceedings of NeurIPS 2024, 2024 a

  35. [35]

    Latte: Latent diffusion transformer for video generation

    Xin Ma, Yaohui Wang, Gengyun Jia, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024 b

  36. [36]

    Dit-3d: Exploring plain diffusion transformers for 3d shape generation

    Shentong Mo, Enze Xie, Ruihang Chu, Lanqing Hong, Matthias Niessner, and Zhenguo Li. Dit-3d: Exploring plain diffusion transformers for 3d shape generation. Advances in Neural Information Processing Systems, 36, 2024

  37. [37]

    Advances in kv caching techniques for large-scale language models

    Helen Nguyen and Isaac Patel. Advances in kv caching techniques for large-scale language models. Advances in Neural Information Processing Systems, 34, 2021

  38. [38]

    Scalable diffusion models with transformers

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

  39. [39]

    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. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684--10695, 2022

  40. [40]

    Progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022

  41. [41]

    Fora: Fast-forward caching in diffusion transformer acceleration

    Pratheba Selvaraju, Tianyu Ding, Tianyi Chen, Ilya Zharkov, and Luming Liang. Fora: Fast-forward caching in diffusion transformer acceleration. arXiv preprint arXiv:2407.01425, 2024

  42. [42]

    Post-training quantization on diffusion models

    Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1972--1981, 2023

  43. [43]

    Denoising diffusion implicit models

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

  44. [44]

    Generative modeling by estimating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019

  45. [45]

    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. arXiv preprint arXiv:2011.13456, 2020 b

  46. [46]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023

  47. [47]

    LiteFocus: Accelerated Diffusion Inference for Long Audio Synthesis

    Zhenxiong Tan, Xinyin Ma, Gongfan Fang, and Xinchao Wang. Litefocus: Accelerated diffusion inference for long audio synthesis. arXiv preprint arXiv:2407.10468, 2024

  48. [48]

    Efficient diffusion models for vision: A survey, 2022

    Anwaar Ulhaq, Naveed Akhtar, and Ganna Pogrebna. Efficient diffusion models for vision: A survey, 2022

  49. [49]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017

  50. [50]

    Ms-diffusion: Multi-subject zero-shot image personalization with layout guidance

    Xierui Wang, Siming Fu, Qihan Huang, Wanggui He, and Hao Jiang. Ms-diffusion: Multi-subject zero-shot image personalization with layout guidance. arXiv preprint arXiv:2406.07209, 2024 a

  51. [51]

    Y. Wang, X. Chen, J. Li, et al. Toca: Accelerating diffusion transformers with token-wise feature caching. arXiv preprint arXiv:2408.XXXX, 2024 b

  52. [52]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13 0 (4): 0 600--612, 2004

  53. [53]

    Diffusion-gan: Training gans with diffusion

    Zhendong Wang, Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Diffusion-gan: Training gans with diffusion. arXiv preprint arXiv:2206.02262, 2022

  54. [54]

    Cache me if you can: Accelerating diffusion models through block caching

    Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. arXiv preprint arXiv:2312.03209, 2023

  55. [55]

    Sparse videogen: Accelerating video diffusion transformers with spatial-temporal sparsity

    Haocheng Xi, Shuo Yang, Yilong Zhao, Chenfeng Xu, Muyang Li, Xiuyu Li, Yujun Lin, Han Cai, Jintao Zhang, Dacheng Li, et al. Sparse videogen: Accelerating video diffusion transformers with spatial-temporal sparsity. arXiv preprint arXiv:2502.01776, 2025

  56. [56]

    Diffusion probabilistic model made slim

    Xingyi Yang, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Diffusion probabilistic model made slim. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22552--22562, 2023

  57. [57]

    Cogvideox: Text-to-video diffusion models with an expert transformer

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024

  58. [58]

    Gddim: Generalized denoising diffusion implicit models

    Qinsheng Zhang, Molei Tao, and Yongxin Chen. Gddim: Generalized denoising diffusion implicit models. arXiv preprint arXiv:2206.05564, 2022

  59. [59]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 586--595, 2018

  60. [60]

    T-gate: Accelerating cross-attention in diffusion models via cache reuse

    Wentian Zhang, Haozhe Liu, Jinheng Xie, Mike Zheng, and Jürgen Schmidhuber. T-gate: Accelerating cross-attention in diffusion models via cache reuse. arXiv preprint arXiv:2404.02747, 2024

  61. [61]

    Pab: Pyramid attention broadcast for real-time video generation

    Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. Pab: Pyramid attention broadcast for real-time video generation. arXiv preprint arXiv:2408.12588, 2024

  62. [62]

    Opensora: Democratizing efficient video production for all, 2024

    Zangwei Zheng et al. Opensora: Democratizing efficient video production for all, 2024. GitHub repository, 2024

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