Pith. sign in

REVIEW 3 major objections 4 minor 43 references

Compute Only 16 Tokens in One Timestep: Accelerating Diffusion Transformers with Cluster-Driven Feature Caching

T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read By computing only one token per spatial cluster during skipped denoising steps, ClusCa claims up to 4.96x faster FLUX image generation and 6.21x faster HunyuanVideo generation while preserving or slightly improving quality.

desk verdict Solid, incremental training-free acceleration for diffusion transformers: the spatial-clustering plus temporal-caching combination is new and works in experiments, but the headline FLOPs number, the O=1/O=2 inconsistency, and missing long-gap cluster-stability analysis need fixing before the results are taken at face value. read the letter →

arxiv 2509.10312 v1 pith:ZPSQQJBC submitted 2025-09-12 cs.CV

Zhixin Zheng , Xinyu Wang , Chang Zou , Shaobo Wang , Linfeng Zhang This is my paper · ORCID
classification cs.CV
keywords diffusiontransformerfeaturecachingspatialclusteringK-Meanstext-to-imagegenerationtext-to-videoinferenceaccelerationtokenreuse
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

ClusCa aims to prove that diffusion transformers hide a second, largely untapped redundancy: tokens within a single denoising step are similar enough that a few representatives can carry a whole step. The paper's strategy is to cluster tokens at every full-calculation timestep, compute one token per cluster during the skipped steps, and blend the fresh representative values into the cached values of all other tokens. This adds a spatial reuse channel on top of existing temporal caching, and the empirical claim is that it both speeds up inference and slows the error accumulation that limits long caching cycles. Concretely, the paper reports 4.96x speedup on FLUX with an ImageReward of 99.49%, slightly above the uncached model, and up to 6.21x on HunyuanVideo with a VBench score near the baseline. If this holds, training-free inference acceleration for image and video diffusion transformers can come from exploiting spatial structure, not only from skipping steps or pruning tokens.

What carries the argument

The machinery is a cluster-conditional cache update. K-Means groups image or video tokens at the last layer of a full-calculation step; at partial steps one token per cluster is recomputed; and the cluster's freshly computed mean $\mu(i)$ is interpolated with each non-computed token's cached value using propagation ratio $\gamma$. The load-bearing object is the cluster assignment vector, whose temporal stability (ARI above 0.8) lets a single representative stand for many tokens, and whose spatial tightness (intra-cluster distance roughly 100x smaller than the global token distance) makes the interpolation accurate.

What would settle it

Measure intra-cluster feature distance at each skipped timestep inside a cache cycle on prompts with small fast-moving objects or rapid lighting changes; if the distance grows past the point where the weighted cache update diverges from a full recomputation, the affected regions should show ghosting or detail loss and the speedup-quality trade-off should collapse.

Watch

Extended reading notes

Core claim

The central claim is that, during the denoising trace, tokens fall into clusters that keep their shape across adjacent timesteps (adjusted Rand index above 0.8), and tokens within a cluster not only lie close in feature space but move along similar trajectories. ClusCa exploits this by dividing timesteps into cache cycles: at the first step of a cycle every token is computed and clustered with K-Means; over the next N-1 steps only K tokens, one per cluster, are recomputed. The refreshed representative features are spread to co-clustered tokens through a weighted update, $C(x_i)=\gamma \mu(i)+(1-\gamma)C(x_i)$, where $\mu(i)$ is the mean of freshly computed representatives in the cluster and $\gamma$ is a propagation ratio. The paper argues that this spatial reuse, combined with temporal reuse from the cache, keeps all tokens updated without computing all of them, and that it is this propagation rather than token importance that prevents quality collapse at high speedup.

Load-bearing premise

The paper assumes that clusters formed at a full-calculation step stay valid for the next several skipped steps, so one randomly chosen token from each cluster can stand for all tokens in that cluster, without any stated bound on when motion or fine detail breaks that coherence.

Editorial extensions

If this is right

  • Feature caching methods that previously treated spatial tokens as independent can now be layered with cluster-driven propagation, so the gains should stack with temporal-cache, step-reduction, and attention-compression approaches.
  • At high speedup ratios, ClusCa reports stable FID on ImageNet with DiT where FORA, ToCa, and DuCa degrade, suggesting that spatial propagation specifically suppresses the error accumulation that limits long cache intervals.
  • On video, the method reports better VBench fidelity sub-metrics (PSNR, SSIM, LPIPS) than the compared cache baselines at 5.54x, indicating that cluster reuse can preserve temporal consistency rather than destroy it.
  • Because the method is training-free and inference-only, it applies to already-deployed diffusion transformers without retraining or weight modification.

Reading between the lines

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

  • If cluster coherence is what matters, then the random choice of one token per cluster implies that token importance scores are unnecessary; cheaper alternatives to K-Means, such as spatial tiling or hashing, might deliver similar gains.
  • The reported ImageReward gain over the uncached FLUX model may be metric-specific; a direct human-preference study would separate genuine aesthetic improvement from reward-model alignment.
  • The method's premise is most vulnerable in video scenes with fast motion, occlusion, or rapid lighting change, so a natural extension is to make reclustering frequency or propagation ratio depend on estimated motion or intra-cluster variance rather than a fixed cycle length.
  • Because ClusCa refreshes representatives from the network while forecasting methods predict the drift of cached values, the two ideas are complementary and could be combined to extend cache-cycle length further.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes ClusCa, a training-free, plug-and-play acceleration method for diffusion transformers. At the first timestep of each cache cycle, all tokens are computed and clustered with K-Means; at the following N-1 timesteps only one token per cluster (K tokens total) is computed, and its updated feature is propagated to the other members of the cluster by a weighted combination with the previously cached feature. Experiments on DiT-XL/2, FLUX.1-dev, and HunyuanVideo report large FLOPs reductions (up to 4.96x on FLUX and 6.21x on HunyuanVideo) with roughly preserved or even improved quality metrics, plus ablations over the propagation ratio, cluster count, and cache interval.

Significance. If the reported results hold, ClusCa is a meaningful contribution to the feature-caching literature: it is training-free, architecture-agnostic in principle, and evaluated on external benchmarks (ImageNet, DrawBench, VBench) against recent baselines (FORA, ToCa, DuCa, TaylorSeer). The use of ARI to quantify cluster stability over timesteps is a step in the right direction, and the code release supports reproducibility. The main caveats are that the headline acceleration figure is FLOPs-based rather than wall-clock-based, and that the central cluster-stability assumption is validated only at adjacent timesteps, not at the cache-cycle lengths actually used in the experiments.

major comments (3)
  1. [Abstract and Table 1] The abstract's claim of "4.96x acceleration on FLUX" is a FLOPs-based speedup (748.48 TFLOPs vs 3719.50 TFLOPs), while the same configuration (ClusCa, N=6, O=1) shows a wall-clock latency speedup of only 3.63x (7.10s vs 25.82s) in Table 1. Since the paper is aimed at inference acceleration, the user-facing metric is latency, and the abstract should state both numbers and clearly label which one is being reported; otherwise the headline overstates the practical speedup by roughly 37%.
  2. [Section 3.2 and 3.3] Observation 2 reports ARI above 0.8 only for adjacent timesteps, but ClusCa reuses cluster assignments for N-1 skipped timesteps with N=5, 6, and 7 in Tables 1-3. The load-bearing assumption is that clusters remain valid across the full cache cycle, so the paper should report ARI at Δt = N-1 (at least for N=5,6,7) and, ideally, the correlation of per-token feature updates within a cluster versus across clusters. The current PCA visualization in Figure 2(b) covers only nine tokens and is qualitative, so the co-variation claim is not yet backed by a measurement of the exact property the propagation rule relies on.
  3. [Section 3.4 and Section 4.5] The cache-update equation C(x_i) = γ·μ(i) + (1-γ)·C(x_i) uses C(x_i) on both sides without distinguishing the old and new cache values, and the behavior of the update when O=1 is essentially a moving average between a randomly selected representative token's feature and the previous cached feature. Because a different token is selected randomly at each partial-calculation step, the propagation injects stochastic variation into non-computed tokens; the paper should either provide a variance or error analysis of this mechanism or include an ablation comparing random selection against a deterministic choice such as the token nearest to the cluster centroid.
minor comments (4)
  1. [Section 4.2 vs Table 1] The text states that at N=5 the ImageReward values of ToCa and DuCa decrease to 0.9731 and 0.9896, respectively, but Table 1 lists no such values; the corresponding rows show ToCa(N=6) at 0.9802 and DuCa(N=5) at 0.9955. Please correct the text to match the table or clarify which configurations are being cited.
  2. [Figure 5 caption] The caption says Δt is the timestep span between compared clusters and also describes the plot as "ARI between adjacent timesteps"; please make clear whether the reported ARI is for adjacent steps only or for a range of Δt values, and if so, which Δt values are used.
  3. [References] References [19] and [20] appear to be the same paper (TaylorSeer, arXiv:2503.06923) and are cited in Sections 2.2 and 3.1 with different numbers; please merge them or renumber consistently.
  4. [Table 1] The "Efficient Attention" column is not defined in the table or the text; please state what the checkmark means and whether all methods, including ClusCa, use the same attention implementation when latency is measured.

Circularity Check

1 steps flagged · score 2.0 of 10

No central circularity: headline quality/speed numbers are external benchmark measurements; only the spatial-similarity motivation leans on a definitional property of K-Means.

  1. self definitional [Section 3.2, Observation 1 and Figure 2(a)]
    "A fundamental property of clustering is that intra-cluster distances are significantly smaller than global distances(Figure 2 a). ... These two observations demonstrate the similarity between tokens within the same cluster in both their features and the variations on their features, indicating that it is possible to compute only one token in each cluster and then reuse its features for the remaining tokens in the same cluster."

    The static half of Observation 1 is true by construction: K-Means is defined as the partition that minimizes intra-cluster pairwise distances (Section 3.3 gives the arg min over S of sum |S_i|^{-1} sum ||x-y||^2). Using this definitional property as evidence that one token can represent its cluster is therefore self-referential rather than independent support. The load-bearing dynamic half, co-variation of feature updates within a cluster, is only qualitatively illustrated by the PCA sketch in Figure 2(b), with no quantitative co-variation metric or long-horizon ARI reported. This makes the motivation partly circular, but the paper's headline results (FID, ImageReward, VBench) are external benchmark measurements, so the circularity is minor and does not force the reported numbers.

full rationale

ClusCa's central claims—4.96x acceleration on FLUX, 6.21x on HunyuanVideo, and the FID/ImageReward/VBench scores—are empirical measurements against external benchmarks, not quantities derived from fitted equations. The update rule C(x_i)=F(x_i) for computed tokens and C(x_i)=gamma*mu(i)+(1-gamma)*C(x_i) for others is a definition of the algorithm; its quality is evaluated, not predicted. The propagation ratio gamma is selected via a validation sweep in Table 4/Figure 10, which is standard hyperparameter tuning rather than a fitted input renamed as a prediction. The TaylorSeer, ToCa, and DuCa citations overlap with the present authors' group, but they are used as baselines and related work, not as the load-bearing justification for ClusCa's effectiveness. The only mildly circular element is the spatial-similarity motivation, which leans on K-Means' definitional property that intra-cluster distances are small; this supports the narrative but does not determine the measured benchmark outcomes. Overall, the paper is self-contained against external benchmarks and exhibits no significant circularity.

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

The method rests on empirical regularities (spatial similarity and cluster stability) plus tuned hyperparameters (gamma, K, N, O). No new physical or mathematical entities are introduced.

free parameters (4)
  • propagation ratio gamma = 0.005 (DiT-XL/2, N=6/7/8)
    Tuned via ablation on ImageNet FID (Table 4). Controls the balance between the new cluster mean and the old cached feature. Chosen by hand/ablation, not derived.
  • number of clusters K = 16 for DiT/FLUX, 32 or 64 for HunyuanVideo
    Selected per model to balance FID/ImageReward against speed (Table 5). Not derived from first principles.
  • cache interval N = 3 to 7 depending on target speed
    Cache cycle length, chosen per experiment to reach a target speedup. Larger N accumulates more error.
  • tokens per cluster O = 1 and 2
    Section 3.3 says O=1 is fixed throughout, but Table 1 reports O=2 configurations and the best ImageReward (0.9961) uses O=2. This multiplies the number of computed tokens by O, so it is a free parameter despite the text claiming otherwise.
assumptions (4)
  • domain assumption Tokens in the same K-Means cluster have similar features and similar temporal dynamics.
    Section 3.2, Observations 1 and 2. Supported by Figure 2 PCA trajectories, but no theoretical guarantee or failure-case analysis is given.
  • domain assumption Cluster assignments from the full-calculation timestep remain valid for the following N-1 skip timesteps.
    Section 3.3. Relies on ARI > 0.8 in Figure 5. If clusters drift due to fast motion or fine details, the representative token becomes unrepresentative.
  • domain assumption The weighted sum gamma*mu(i) + (1-gamma)*C(x_i) approximates the true uncomputed feature for non-computed tokens.
    Section 3.4. The mixing rule is chosen heuristically and gamma is tuned, not derived.
  • domain assumption Evaluation metrics (FID, ImageReward, CLIP, VBench) are reliable indicators of generation quality for this comparison.
    Standard in the field, but ImageReward is a learned preference model and may be noisy, and no error bars are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Compute Only 16 Tokens in One Timestep: Accelerating Diffusion Transformers with Cluster-Driven Feature Caching." pith.science (2026). https://pith.science/paper/ZPSQQJBC

@misc{pith2026250910312,
  author       = {Pith},
  title        = {Pith review of: Compute Only 16 Tokens in One Timestep: Accelerating Diffusion Transformers with Cluster-Driven Feature Caching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZPSQQJBC}},
  note         = {Machine review of arXiv:2509.10312}
}
read the original abstract

Diffusion transformers have gained significant attention in recent years for their ability to generate high-quality images and videos, yet still suffer from a huge computational cost due to their iterative denoising process. Recently, feature caching has been introduced to accelerate diffusion transformers by caching the feature computation in previous timesteps and reusing it in the following timesteps, which leverage the temporal similarity of diffusion models while ignoring the similarity in the spatial dimension. In this paper, we introduce Cluster-Driven Feature Caching (ClusCa) as an orthogonal and complementary perspective for previous feature caching. Specifically, ClusCa performs spatial clustering on tokens in each timestep, computes only one token in each cluster and propagates their information to all the other tokens, which is able to reduce the number of tokens by over 90%. Extensive experiments on DiT, FLUX and HunyuanVideo demonstrate its effectiveness in both text-to-image and text-to-video generation. Besides, it can be directly applied to any diffusion transformer without requirements for training. For instance, ClusCa achieves 4.96x acceleration on FLUX with an ImageReward of 99.49%, surpassing the original model by 0.51%. The code is available at https://github.com/Shenyi-Z/Cache4Diffusion.

Figures

Figures reproduced from arXiv: 2509.10312 by the authors.

Figure 1
Figure 1. Visualization of similarity in two dimensions [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 4
Figure 4. The Overview of ClusCa: (a) In full-calculation step, ClusCa caches tokens and reuses them across timesteps by [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 5
Figure 5. (a) Visualization of clustering results in adjacent steps. (b) Adjusted rand index (ARI) of clusters in adjacent steps, [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figures from the paper (4 more)
Figure 6
Figure 6. Figure 6: Qualitative comparison on FLUX. ClusCa excels in [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Visualization of different generation methods on various prompts. ClusCa produces high-quality, accurate images, [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 9
Figure 9. Figure 9: PCA visualization on the trajectories of fea￾tures in different timesteps. “DiT” here denotes the orig￾inal model without accelera￾tion, which can be considered as the ground-truth trajectory. Propagation Ratio (×10!") FID sFID [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 11
Figure 11. Figure 11: Time consumption visualization of each part [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 7 canonical work pages

  1. [1]

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. 2023. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127(2023)

    arXiv 2023

  2. [2]

    Daniel Bolya and Judy Hoffman. 2023. Token merging for fast stable diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 4599–4603

    2023

  3. [3]

    Pengtao Chen, Mingzhu Shen, Peng Ye, Jianjian Cao, Chongjun Tu, Christos- Savvas Bouganis, Yiren Zhao, and Tao Chen. 2024. Δ-DiT: A Training-Free Acceleration Method Tailored for Diffusion Transformers.arXiv preprint arXiv:2406.01125(2024)

    arXiv 2024

  4. [4]

    Xinle Cheng, Zhuoming Chen, and Zhihao Jia. 2025. CAT Pruning: Cluster-Aware Token Pruning For Text-to-Image Diffusion Models. arXiv:2502.00433 [cs.CV] https://arxiv.org/abs/2502.00433

    arXiv 2025

  5. [5]

    Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. 2022. Flashat- tention: Fast and memory-efficient exact attention with io-awareness.Advances in neural information processing systems35 (2022), 16344–16359

    2022

  6. [6]

    Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion models beat gans on image synthesis.Advances in neural information processing systems34 (2021), 8780–8794

    2021

  7. [7]

    Gongfan Fang, Xinyin Ma, and Xinchao Wang. 2023. Structural Pruning for Diffusion Models. arXiv:2305.10924 [cs.LG] https://arxiv.org/abs/2305.10924

    arXiv 2023

  8. [8]

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. 2021. Clipscore: A reference-free evaluation metric for image captioning.arXiv preprint arXiv:2104.08718(2021)

    arXiv 2021

Show all 43 references
  1. [9]

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2017. GANs Trained by a Two Time-Scale Update Rule Con- verge to a Local Nash Equilibrium. InAdvances in Neural Information Processing Systems 30: Annual Conference on Neural Information...

    2017

  2. [10]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models.Advances in neural information processing systems33 (2020), 6840–6851

    2020

  3. [11]

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuan- han Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. 2024. Vbench: Comprehensive benchmark suite for video generative models. InProceedings of the IEEE/CVF Conference on Computer Vision ...

    2024

  4. [12]

    2021.OpenCLIP

    Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. 2021.OpenCLIP. doi:10.5281/zenodo.5143773 If you use this software...

    2021 doi

  5. [13]

    Minchul Kim, Shangqian Gao, Yen-Chang Hsu, Yilin Shen, and Hongxia Jin. 2024. Token fusion: Bridging the gap between token pruning and token merging. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 1383–1392

    2024

  6. [14]

    Sungbin Kim, Hyunwuk Lee, Wonho Cho, Mincheol Park, and Won Woo Ro. 2025. Ditto: Accelerating Diffusion Model via Temporal Value Similarity. In2025 IEEE International Symposium on High Performance Computer Architecture (HPCA). IEEE, 338–352

    2025

  7. [15]

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

    2024

  8. [16]

    Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shang- hang Zhang, and Kurt Keutzer. 2023. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 17535– 17545

    2023

  9. [17]

    Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. 2024. Hunyuan- dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding.arXiv preprint arXiv:2405...

    2024 arXiv

  10. [18]

    Feng Liu, Shiwei Zhang, Xiaofeng Wang, Yujie Wei, Haonan Qiu, Yuzhong Zhao, Yingya Zhang, Qixiang Ye, and Fang Wan. 2025. Timestep Embedding Tells: It’s Time to Cache for Video Diffusion Model. InProceedings of the Computer Vision and Pattern Recognition Conference. 7353–7363

    2025

  11. [20]

    Jiacheng Liu, Chang Zou, Yuanhuiyi Lyu, Junjie Chen, and Linfeng Zhang. 2025. From reusing to forecasting: Accelerating diffusion models with taylorseers. arXiv preprint arXiv:2503.06923(2025)

    2025 arXiv

  12. [21]

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

    2022 arXiv

  13. [22]

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu

  14. [23]

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. 2025. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. Machine Intelligence Research(2025), 1–22

    2025

  15. [24]

    Xinyin Ma, Gongfan Fang, and Xinchao Wang. 2024. Deepcache: Accelerating diffusion models for free. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 15762–15772

    2024

  16. [25]

    William Peebles and Saining Xie. 2023. Scalable diffusion models with transform- ers. InProceedings of the IEEE/CVF international conference on computer vision. 4195–4205

    2023

  17. [26]

    Bernstein, Alexander C

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Li Fei-Fei. 2015. ImageNet Large Scale Visual Recognition Challenge.Int. J. Comput. Vis.115, 3 (2015), 21...

    2015 doi

  18. [27]

    Omid Saghatchian, Atiyeh Gh Moghadam, and Ahmad Nickabadi. 2025. Cached Adaptive Token Merging: Dynamic Token Reduction and Redundant Computa- tion Elimination in Diffusion Model.arXiv preprint arXiv:2501.00946(2025)

    2025 arXiv

  19. [28]

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. 2022. Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural inform...

    2022

  20. [29]

    Pratheba Selvaraju, Tianyu Ding, Tianyi Chen, Ilya Zharkov, and Luming Liang

  21. [30]

    Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. 2023. Post- training quantization on diffusion models. InProceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition. 1972–1981

    2023

  22. [31]

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

    2020 arXiv

  23. [32]

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. 2023. Consistency models. (2023)

    2023

  24. [33]

    Xingwu Sun, Yanfeng Chen, Yiqing Huang, Ruobing Xie, Jiaqi Zhu, Kai Zhang, Shuaipeng Li, Zhen Yang, Jonny Han, Xiaobo Shu, et al. 2024. Hunyuan-large: An open-source moe model with 52 billion activated parameters by tencent.arXiv preprint arXiv:2411.02265(2024)

    2024 arXiv

  25. [34]

    Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. 2024. Cache me if you can: Accelerating diffusion models through block caching. InProceedings of the IEEE/CVF Conference on Computer ...

    2024

  26. [35]

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. 2023. Imagereward: Learning and evaluating human prefer- ences for text-to-image generation.Advances in Neural Information Processing Systems36 (2023), 15903–15935

    2023

  27. [36]

    Zhihang Yuan, Hanling Zhang, Lu Pu, Xuefei Ning, Linfeng Zhang, Tianchen Zhao, Shengen Yan, Guohao Dai, and Yu Wang. 2024. Ditfastattn: Attention compression for diffusion transformer models.Advances in Neural Information Processing Systems37 (2024), 1196–1219

    2024

  28. [37]

    Evelyn Zhang, Jiayi Tang, Xuefei Ning, and Linfeng Zhang. 2025. Training-Free and Hardware-Friendly Acceleration for Diffusion Models via Similarity-based Token Pruning. InProceedings of the AAAI Conference on Artificial Intelligence

    2025

  29. [38]

    Evelyn Zhang, Bang Xiao, Jiayi Tang, Qianli Ma, Chang Zou, Xuefei Ning, Xum- ing Hu, and Linfeng Zhang. 2024. Token pruning for caching better: 9 times acceleration on stable diffusion for free.arXiv preprint arXiv:2501.00375(2024)

    2024 arXiv

  30. [39]

    Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. 2024. Real-time video generation with pyramid attention broadcast.arXiv preprint arXiv:2408.12588 (2024)

    2024 arXiv

  31. [40]

    Haowei Zhu, Dehua Tang, Ji Liu, Mingjie Lu, Jintu Zheng, Jinzhang Peng, Dong Li, Yu Wang, Fan Jiang, Lu Tian, et al. 2024. Dip-go: A diffusion pruner via few-step gradient optimization.Advances in Neural Information Processing Systems37 (2024), 92581–92604

    2024

  32. [41]

    Chang Zou, Xuyang Liu, Ting Liu, Siteng Huang, and Linfeng Zhang. 2024. Accelerating diffusion transformers with token-wise feature caching.arXiv preprint arXiv:2410.05317(2024)

    2024 arXiv

  33. [42]

    Chang Zou, Evelyn Zhang, Runlin Guo, Haohang Xu, Conghui He, Xuming Hu, and Linfeng Zhang. 2024. Accelerating diffusion transformers with dual feature caching.arXiv preprint arXiv:2412.18911(2024). Compute Only 16 Tokens in One Timestep: Accelerating Diffusion Transformers wit...

    2024

  34. [2022]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.Advances in neural information processing systems35 (2022), 5775–5787

    2022

  35. [2024]

    Fora: Fast-forward caching in diffusion transformer acceleration.arXiv preprint arXiv:2407.01425(2024)

    2024 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.