Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

LESA claims diffusion acceleration can be learned rather than guessed, hitting 6.25x on Qwen-Image

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 ↗

T0 review · deepseek-v4-flash

2026-08-02 21:18 UTC pith:OYI5CF6Q

load-bearing objection LESA is a solid caching extension worth a referee, but the abstract's HunyuanVideo PSNR number disagrees with its own Table 3 and must be fixed. the 4 major comments →

arxiv 2602.20497 v3 pith:OYI5CF6Q submitted 2026-02-24 cs.CV cs.AI

LESA: Learnable Stage-Aware Predictors for Diffusion Model Acceleration

Peiliang Cai , Jiacheng Liu , Haowen Xu , Xinyu Wang , Chang Zou , Linfeng Zhang This is my paper
classification cs.CV cs.AI
keywords diffusion modelsfeature cachingdiffusion transformer accelerationKolmogorov-Arnold Networkstage-aware predictiontext-to-image generationtext-to-video generationtemporal feature modeling
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 paper claims that feature evolution during diffusion denoising is not uniformly smooth: it is drastic under high noise, stable in the middle, and detail-refining at the end. Because existing caching methods use a single fixed reuse or polynomial-forecast rule across all timesteps, they lose quality under aggressive acceleration. To fix this, the authors introduce LESA, a learnable predictor that splits the denoising trajectory into stages and assigns each stage a dedicated expert built on a Kolmogorov–Arnold Network. The predictor is trained first on ground-truth features and then in a closed-loop autoregressive manner, making it robust to its own drift during fast inference. The paper reports 5.00x speedup on FLUX.1-dev with a 1.0% ImageReward drop, 6.25x on Qwen-Image with a 20.2% ImageReward improvement over TaylorSeer, and 5.00x on HunyuanVideo with a 24.7% PSNR improvement over TaylorSeer.

Core claim

The central claim is that stage-aware learnable prediction, rather than training-free reuse or polynomial forecasting, is what allows diffusion transformers to be accelerated without sacrificing fidelity. The paper shows experimentally that a compact KAN-based predictor, modulated by a scalar learned from relative timestep offsets and partitioned into three noise-level experts, consistently outperforms fixed-rule caching baselines across five models and both image and video generation. The authors interpret this as evidence that diffusion feature trajectories are non-smooth and stage-dependent, and that learning stage-specific temporal mappings captures dynamics that Taylor-style extrapolati

What carries the argument

The load-bearing mechanism is a residual predictor combining a linear projection of the last K cached features with a scalar-output Kolmogorov–Arnold Network (KAN), a neural network that learns smooth univariate spline functions instead of using fixed activations. The KAN maps relative timestep offsets to a scalar alpha that modulates the residual update, separating spatial feature transformation from temporal modulation. Three stage experts, using window sizes K=4 for high noise and K=8 for later stages, are assigned to distinct noise-level phases, and the whole system is trained in two stages: ground-truth guided and closed-loop autoregressive.

Load-bearing premise

The load-bearing premise is that the hand-set stage boundaries and window sizes, chosen after inspecting feature-dynamics plots of the tested models, generalize, and that the 100 training prompts do not overlap the 200 evaluation prompts enough to inflate the reported quality metrics.

What would settle it

Train LESA on a prompt set proven disjoint from the evaluation set, and fix the stage boundaries a priori on one model before applying them to others; if the reported gains over TaylorSeer shrink or disappear, the claim of general stage-aware learning is undermined.

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

If this is right

  • If the central claim holds, feature caching for diffusion models no longer has to be training-free; a small learned predictor can beat Taylor-based forecasts at the same speedup.
  • Acceleration would generalize across model families and even to distilled models, where the paper reports quality improvements rather than degradation.
  • Text-to-video generation would also benefit, indicating that temporal redundancy in diffusion transformers is exploitable with stage-aware prediction.
  • The method's low cache memory (0.81 GB on FLUX) and parameter efficiency could make high-speed generation practical on memory-constrained hardware.

Where Pith is reading between the lines

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

  • The stage boundaries and window sizes are hand-set after inspecting feature-dynamics plots of the very models being evaluated; if these are per-model tuning choices, the method may require reconfiguration when moving to a new model rather than being fully plug-and-play.
  • The paper does not state whether the 100 custom training prompts are disjoint from the 200 DrawBench evaluation prompts; if there is overlap, the reported quality gains could be partly inflated by memorization rather than genuine temporal modeling.
  • The ablation shows the stage-aware KAN advantage grows with the inference interval N, suggesting the benefit is concentrated in aggressive acceleration regimes—a testable prediction for future work.
  • A natural next step, not explored in the paper, is to learn the stage boundaries and window sizes end-to-end instead of selecting them offline, which could reveal whether the stage segmentation itself or just the KAN modulation is the key ingredient.

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 / 5 minor

Summary. The paper proposes LESA, a learned feature-caching predictor for accelerating diffusion transformer inference. The method combines a KAN-based scalar temporal modulation with a multi-expert, stage-aware architecture: the denoising trajectory is split into stages, each assigned a dedicated predictor with its own window size, and training proceeds in two phases (ground-truth guided, then closed-loop autoregressive). Experiments are reported on FLUX.1-dev/schnell, Qwen-Image/Lightning, HunyuanVideo, and Stable Diffusion XL, with 5.00–6.25x FLOPs-based speedups and claims of improved or comparable quality relative to TaylorSeer and other caching baselines.

Significance. If the results are accurate, LESA is a meaningful empirical contribution: it demonstrates that a small learned predictor can outperform hand-crafted Taylor forecasting for feature caching across multiple large-scale DiT models, and it includes useful ablations (prompt-count sensitivity, memory comparison) plus a code-release commitment. The multi-model validation is a strength. However, the paper's headline empirical claim is undermined by an internal inconsistency in the HunyuanVideo abstract number, and the stage-segmentation mechanism is insufficiently specified for reproduction. The contribution is potentially significant, but the current version needs verification and clarification before the SOTA claim can be accepted.

major comments (4)
  1. [Abstract / §4.2, Table 3] The abstract claims "5.00x acceleration on HunyuanVideo with a 24.7% PSNR improvement over TaylorSeer." In Table 3, at 5.00x FLOPs, TaylorSeer(N=5,O=1) has PSNR 17.29 and LESA(N=7) has PSNR 21.43, giving (21.43−17.29)/17.29 = 23.9%. The LESA(N=8) row gives (21.05−17.29)/17.29 = 21.7%. No row pairing in the table yields 24.7%. Since this is a headline SOTA claim, please correct the number or explicitly specify the baseline/setting used to obtain 24.7%.
  2. [§3.2, §7.1, Fig. 2] The paper never states the timestep ranges defining the three stages, nor the criterion for placing the boundaries. The window sizes K=4 and K=8 and the expert allocation depend on these boundaries, yet they are free parameters selected after inspecting feature-dynamics plots from the very models evaluated in Tables 1–3. This makes the central "stage-aware" mechanism non-reproducible and the cross-model generalization claim unsupported as written. Please report the exact stage boundaries, explain how they were chosen, and include an experiment that fixes the boundaries on one model and applies them unchanged to the other models.
  3. [§7.2 vs §7.1] Training uses 100 custom prompts (Section 7.2) while evaluation uses 200 DrawBench prompts (Section 7.1). The paper does not state whether the two sets are disjoint. Although Table 9 shows performance is almost invariant to the number of training prompts, overlap with the evaluation set should be explicitly ruled out. Please disclose the relationship between the 100 training prompts and the DrawBench-200 test prompts.
  4. [Abstract, Tables 1–3] The headline speedups are quoted from the FLOPs-reduction column, but the measured wall-clock latency speedups are lower. For example, FLUX.1-dev LESA(N=7) shows 4.48x latency speedup vs 5.00x FLOPs speedup (Table 1); Qwen-Image LESA(N=10) shows 5.69x latency vs 6.25x FLOPs (Table 2). The abstract's "acceleration" claims should either be labeled as FLOPs reduction or based on the measured latency, to avoid overstating practical runtime gains.
minor comments (5)
  1. [Table 1] The table title says "for FLUX.1-schnell" but the table contains FLUX.1-dev rows and the section heading covers both models. Retitle as "FLUX.1-dev and FLUX.1-schnell."
  2. [§3.2] The paper advertises a "two-stage training strategy" but the text enumerates three components: Data Preparation, Ground-Truth Guided Training, and Closed-Loop Autoregressive Training. Please renumber or clarify the terminology.
  3. [Tables 1–4] No error bars or repeated-run statistics are reported. For small metric differences (e.g., Table 1 ImageReward 0.98 for LESA vs 1.02 for TaylorSeer at comparable speedup), reporting variance or multiple seeds would substantially strengthen the comparisons.
  4. [Fig. 2] The subplot labels, axis labels, and annotations are very small and hard to read. Please enlarge fonts and clarify which trajectory corresponds to which model.
  5. [Eqs. (5)–(7)] The KAN produces a single scalar modulation that is broadcast across all feature channels. The paper should explicitly acknowledge this expressivity limitation and justify why per-channel or per-token modulation is not needed.

Circularity Check

0 steps flagged

No significant circularity: LESA is a supervised temporal regressor trained on ground-truth features; the abstract/table discrepancy is a correctness issue, not a circular reduction.

full rationale

LESA's derivation chain is not circular. The predictor is a supervised temporal regressor: Eq. (7) computes ĥ_{t−1} = h_t + αz from past cached features {h_{t+K−1},...,h_t}, with α produced by a KAN over relative timestep offsets (Eqs. 4–6), and is trained by an L1 loss against ground-truth DiT features (Sec. 3.2, 'Training Procedure'). The target feature h_{t−1} is not an input to the predictor, so the prediction is not equal to the fit by construction. Evaluation is performed against the original model's outputs on DrawBench-200 (Sec. 7.1), not against the training targets. The training/evaluation prompt overlap is a potential weakness, but Table 9 shows quality is nearly invariant to training prompt count (PSNR 30.18 for 1 prompt vs 30.18 for 100 prompts on Qwen-Image at N=7), so any overlap is not load-bearing. The abstract's HunyuanVideo '24.7% PSNR improvement' is not reproduced by Table 3, which yields 23.9% for LESA(N=7) vs TaylorSeer(N=5) (21.43 vs 17.29) and 21.7% for LESA(N=8); this is an internal-consistency/correctness concern, not a circular reduction. Self-citations, including TaylorSeer [17], HiCache [4], and FreqCa [15], serve as baselines and related work, but no load-bearing claim reduces to a self-citation, and no uniqueness theorem or ansatz is imported from the authors' prior work. The stage boundaries and window sizes (K=4, K=8) are design choices informed by feature-dynamics plots of the evaluated models, which is a possible overfitting/test-selection concern, but it is not an Eq.-X=Eq.-Y or fitted-parameter-called-prediction circularity. Overall, the central derivation is self-contained and externally benchmarked.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim rests on several hand-chosen hyperparameters (stage count, window sizes, training prompts) and on unstated assumptions about the Markov property of features and the stage-separability of the dynamics. No new physical entities are introduced.

free parameters (6)
  • Number of stages = 3
    Stage count chosen by hand from trajectory plots in Fig. 2; not derived.
  • Window sizes K = 4 (high-noise), 8 (mid/low-noise)
    Section 3.2: K=4 for rapid high-noise transitions, K=8 for stable phases; chosen without quantitative selection.
  • Stage boundaries = Not numerically specified
    The noise-level cutoffs dividing the three stages are never given; they are free modeling choices fit to the observed models.
  • KAN hidden dimension = 256
    Section 7.2; architecture hyperparameter.
  • Training prompt count = 100
    Section 7.2: 'custom set of 100 prompts'; not justified and possible overlap with eval.
  • Inference interval N = N=5,7,10 depending on experiment
    The skip interval is chosen per experiment; different N values change the acceleration/quality tradeoff.
axioms (5)
  • standard math Kolmogorov-Arnold representation theorem
    Basis for KAN architecture (Eq. 3, 6).
  • domain assumption Feature trajectory is stage-dependent and can be segmented into three discrete regimes
    Main motivation from Fig. 2; if false, stage experts add nothing beyond a single predictor.
  • domain assumption Past K features are sufficient to predict the next feature
    Eq. (4)-(7) maps a finite window of cached features to the target; assumes Markov property of order K.
  • ad hoc to paper Scalar modulation alpha captures all temporal variation
    The separation-of-variables form (linear projection in space, scalar KAN in time) is introduced without proof and is a restrictive ansatz.
  • domain assumption 100 training prompts generalize to the 200-prompt evaluation set
    Needed for the reported generalization; not demonstrated by a separate validation split.

pith-pipeline@v1.3.0-alltime-deepseek · 18930 in / 11130 out tokens · 96390 ms · 2026-08-02T21:18:02.287496+00:00 · methodology

0 comments
read the original abstract

Diffusion models have achieved remarkable success in image and video generation tasks. However, the high computational demands of Diffusion Transformers (DiTs) pose a significant challenge to their practical deployment. While feature caching is a promising acceleration strategy, existing methods based on simple reusing or training-free forecasting struggle to adapt to the complex, stage-dependent dynamics of the diffusion process, often resulting in quality degradation and failing to maintain consistency with the standard denoising process. To address this, we propose a LEarnable Stage-Aware (LESA) predictor framework based on two-stage training. Our approach leverages a Kolmogorov-Arnold Network (KAN) to accurately learn temporal feature mappings from data. We further introduce a multi-stage, multi-expert architecture that assigns specialized predictors to different noise-level stages, enabling more precise and robust feature forecasting. Extensive experiments show our method achieves significant acceleration while maintaining high-fidelity generation. Experiments demonstrate 5.00x acceleration on FLUX.1-dev with minimal quality degradation (1.0% drop), 6.25x speedup on Qwen-Image with a 20.2% quality improvement over the previous SOTA (TaylorSeer), and 5.00x acceleration on HunyuanVideo with a 24.7% PSNR improvement over TaylorSeer. State-of-the-art performance on both text-to-image and text-to-video synthesis validates the effectiveness and generalization capability of our training-based framework across different models. Our code is available at https://github.com/caipeiliang2004/LESA.

Figures

Figures reproduced from arXiv: 2602.20497 by Chang Zou, Haowen Xu, Jiacheng Liu, Linfeng Zhang, Peiliang Cai, Xinyu Wang.

Figure 1
Figure 1. Figure 1: Images generated by Qwen-image using LESA with a 6.25× acceleration. Abstract Diffusion models have achieved remarkable success in im￾age and video generation tasks. However, the high compu￾tational demands of Diffusion Transformers (DiTs) pose a significant challenge to their practical deployment. While feature caching is a promising acceleration strategy, exist￾ing methods based on simple reusing or trai… view at source ↗
Figure 2
Figure 2. Figure 2: Cosine Similarity curves and PCA-projected trajectories show that feature evolution differs markedly across diffusion models, with non-smooth, stage-dependent dynamics that challenge the common assumption of simple or continuous temporal change. quality and diversity of generated content. However, DiT models face significant computational challenges during in￾ference due to their deep transformer architect… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the training-based stage-aware learnable predictor framework. (a) Training progress of the LESA uses outputs from the previous K steps as input and updates the predictor with the DiT activation output. (b) Inference process with the learned predictor skips part of the DiT computations by predicting features. (c) Predictor architecture processes features from previous steps with a linear project… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of LESA and TaylorSeer under different speedup ratios. LESA preserves better quality metrics and percep￾tual metrics On QwenImage and FLUX. learn how features evolve under imperfect inputs, resulting in more robust behavior during accelerated inference. For a comprehensive description of the training details, please refer to the supplementary material. 4. Experiments Evaluation and Metrics For t… view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of Sampling Process Image and Final Im￾age. On FLUX, LESA demonstrates significant speedup ratio with excellent spatial organization and accurate color representation. Qualitative Study As shown in [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative Comparison of Generated Images. On Qwen-Image, LESA delivers higher speedup with high image quality. substantial degradation of perceptual quality, and existing baselines exhibit a similar trade-off. In contrast, LESA si￾multaneously achieves faster inference and markedly better visual fidelity. With N =7, it attains the best overall perfor￾mance, improving PSNR by 23.9% over TaylorSeer; even u… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative Comparison of Early, Middle, and Late Frames From Generated Videos. On HunyuanVideo, LESA constructs continuous dynamic process with excellent subject accuracy under high speedup ratio [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗

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. Focused Forcing: Content-Aware Per-Frame KV Selection for Efficient Autoregressive Video Diffusion

    cs.CV 2026-05 unverdicted novelty 5.0

    Focused Forcing is a training-free per-frame KV selection method that combines attention scores with diversity metrics and head-importance estimation to accelerate autoregressive video diffusion up to 1.48x while impr...

Reference graph

Works this paper leans on

51 extracted references · 15 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023. 1

    Pith/arXiv arXiv 2023

  2. [2]

    δ-dit: A training-free acceleration method tailored for diffusion transformers.arXiv preprint arXiv:2406.01125,

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

    Pith/arXiv arXiv

  3. [3]

    Struc- tural pruning for diffusion models.arXiv preprint arXiv:2305.10924, 2023

    Gongfan Fang, Xinyin Ma, and Xinchao Wang. Struc- tural pruning for diffusion models.arXiv preprint arXiv:2305.10924, 2023. 3

    Pith/arXiv arXiv 2023

  4. [4]

    Hicache: Training-free accelera- tion of diffusion models via hermite polynomial-based feature caching, 2025

    Liang Feng, Shikang Zheng, Jiacheng Liu, Yuqi Lin, Qinming Zhou, Peiliang Cai, Xinyu Wang, Junjie Chen, Chang Zou, Yue Ma, and Linfeng Zhang. Hicache: Training-free accelera- tion of diffusion models via hermite polynomial-based feature caching, 2025. 3

    2025

  5. [5]

    A temporal kolmogorov- arnold transformer for time series forecasting.arXiv preprint arXiv:2406.02486, 2024

    Remi Genet and Hugo Inzirillo. A temporal kolmogorov- arnold transformer for time series forecasting.arXiv preprint arXiv:2406.02486, 2024. 3

    Pith/arXiv arXiv 2024

  6. [6]

    Clipscore: A reference-free evaluation metric for image captioning.arXiv preprint arXiv:2104.08718, 2021

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning.arXiv preprint arXiv:2104.08718, 2021. 5

    Pith/arXiv arXiv 2021

  7. [7]

    Denoising Dif- fusion Probabilistic Models, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Dif- fusion Probabilistic Models, 2020. arXiv:2006.11239 [cs]. 1

    Pith/arXiv arXiv 2020

  8. [8]

    Dif- fusion dataset condensation: Training your diffusion model faster with less data.arXiv preprint arXiv:2507.05914, 2025

    Rui Huang, Shitong Shao, Zikai Zhou, Pukun Zhao, Hangyu Guo, Tian Ye, Lichen Bai, Shuo Yang, and Zeke Xie. Dif- fusion dataset condensation: Training your diffusion model faster with less data.arXiv preprint arXiv:2507.05914, 2025. 3

    arXiv 2025

  9. [9]

    Ryoo, and Tian Xie

    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, 2024. 3

    2024

  10. [10]

    Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024. 3, 2

    Pith/arXiv arXiv 2024

  11. [11]

    Black Forest Labs. Flux. https://github.com/ black-forest-labs/flux, 2024. 2

    2024

  12. [12]

    Faster diffusion: Rethinking the role of unet encoder in diffu- sion models.arXiv preprint arXiv:2312.09608, 2023

    Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster diffusion: Rethinking the role of unet encoder in diffu- sion models.arXiv preprint arXiv:2312.09608, 2023. 2

    Pith/arXiv arXiv 2023

  13. [13]

    Hunyuan- DiT: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding

    Zhimin Li, Jianwei Zhang, and and others Lin. Hunyuan- DiT: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding. 3

  14. [14]

    Timestep embedding tells: It’s time to cache for video diffusion model, 2024

    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, 2024. 2, 3

    2024

  15. [15]

    Freqca: Accelerating diffusion models via frequency-aware caching,

    Jiacheng Liu, Peiliang Cai, Qinming Zhou, Yuqi Lin, Deyang Kong, Benhao Huang, Yupei Pan, Haowen Xu, Chang Zou, Junshu Tang, Shikang Zheng, and Linfeng Zhang. Freqca: Accelerating diffusion models via frequency-aware caching,

  16. [16]

    A survey on cache methods in diffu- sion models: Toward efficient multi-modal generation.arXiv preprint arXiv:2510.19755, 2025

    Jiacheng Liu, Xinyu Wang, Yuqi Lin, Zhikai Wang, Peiru Wang, Peiliang Cai, Qinming Zhou, Zhengan Yan, Zexuan Yan, Zhengyi Shi, et al. A survey on cache methods in diffu- sion models: Toward efficient multi-modal generation.arXiv preprint arXiv:2510.19755, 2025. 2

    arXiv 2025

  17. [17]

    From reusing to forecasting: Accelerating diffusion models with taylorseers, 2025

    Jiacheng Liu, Chang Zou, Yuanhuiyi Lyu, Junjie Chen, and Linfeng Zhang. From reusing to forecasting: Accelerating diffusion models with taylorseers, 2025. 2, 3

    2025

  18. [18]

    Speca: Accelerating diffusion transformers with speculative feature caching

    Jiacheng Liu, Chang Zou, Yuanhuiyi Lyu, Kaixin Li, Shaobo Wang, and Linfeng Zhang. Speca: Accelerating diffusion transformers with speculative feature caching. InProceedings of the 33rd ACM International Conference on Multimedia (MM ’25), page to appear, Dublin, Ireland, 2025. ACM. 2, 3

    2025

  19. [19]

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

    Xingchao Liu, Chengyue Gong, et al. Flow straight and fast: Learning to generate and transfer data with rectified flow. InThe Eleventh International Conference on Learning Representations, 2023. 3

    2023

  20. [20]

    Hou, and Max Tegmark

    Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljacic, Thomas Y . Hou, and Max Tegmark. Kan: Kolmogorov-arnold networks.ArXiv, abs/2404.19756, 2024. 2, 3

    Pith/arXiv arXiv 2024

  21. [21]

    C-kan: A new approach for integrating convolutional layers with kolmogorov–arnold networks for time-series forecasting.Mathematics, 12(19):3022, 2024

    Ioannis E Livieris. C-kan: A new approach for integrating convolutional layers with kolmogorov–arnold networks for time-series forecasting.Mathematics, 12(19):3022, 2024. 3

    2024

  22. [22]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.Advances in Neural Information Processing Systems, 35:5775–5787, 2022

    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, 35:5775–5787, 2022. 3

    2022

  23. [23]

    Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models.arXiv preprint arXiv:2211.01095, 2022

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models.arXiv preprint arXiv:2211.01095, 2022. 3

    Pith/arXiv arXiv 2022

  24. [24]

    Learning-to-cache: Accelerating diffusion transformer via layer caching.arXiv preprint arXiv:2406.01733, 2024

    Xinyin Ma, Gongfan Fang, Michael Bi Mi, and Xinchao Wang. Learning-to-cache: Accelerating diffusion transformer via layer caching.arXiv preprint arXiv:2406.01733, 2024. 3

    Pith/arXiv arXiv 2024

  25. [25]

    Deepcache: Accelerating diffusion models for free

    Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15762–15772, 2024. 2, 3

    2024

  26. [26]

    Scalable Diffusion Models with Transformers, 2023

    William Peebles and Saining Xie. Scalable Diffusion Models with Transformers, 2023. arXiv:2212.09748 [cs]. 1, 3

    Pith/arXiv arXiv 2023

  27. [27]

    High-Resolution Image Syn- thesis with Latent Diffusion Models, 2022

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-Resolution Image Syn- thesis with Latent Diffusion Models, 2022. arXiv:2112.10752 [cs]. 1

    Pith/arXiv arXiv 2022

  28. [28]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. InMedical image computing and computer-assisted 9 intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234–241. Springer, 2015. 3

    2015

  29. [29]

    Moghadam, and Ahmad Nick- abadi

    Omid Saghatchian, Atiyeh Gh. Moghadam, and Ahmad Nick- abadi. Cached adaptive token merging: Dynamic token re- duction and redundant computation elimination in diffusion model, 2025. 3

    2025

  30. [30]

    Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Moham- mad Norouzi. Photorealistic text-to-image diffusion mod- els with deep language understanding. InProceedings of the IEEE/CVF Confere...

    2022

  31. [31]

    Fora: Fast-forward caching in diffusion transformer acceleration.arXiv preprint arXiv:2407.01425,

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

    Pith/arXiv arXiv

  32. [32]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. InInternational Conference on Learning Representations, 2021. 3

    2021

  33. [33]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. InInternational Conference on Machine Learning, pages 32211–32252. PMLR, 2023. 3

    2023

  34. [34]

    Kolmogorov-arnold networks (kans) for time series analysis.arXiv preprint arXiv:2405.08790, 2024

    Cristian J Vaca-Rubio, Luis Blanco, Roberto Pereira, and M`arius Caus. Kolmogorov-arnold networks (kans) for time series analysis.arXiv preprint arXiv:2405.08790, 2024. 3

    arXiv 2024

  35. [35]

    Wan: Open and advanced large- scale video generative models

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pan- deng Li, Pingyu Wu, Ruihang Chu, Ruili Feng, Shiwei Zhang, Siyang Sun, Tao Fang,...

  36. [36]

    Bovik, H.R

    Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: From error visibility to structural similarity.IEEE Transactions on Image Processing, 13(4): 600–612, 2004. 5

    2004

  37. [37]

    Qwen-image technical report.ArXiv, abs/2508.02324, 2025

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, Yuxiang Chen, Zecheng Tang, Zekai Zhang, Zhengyi Wang, An Yang, Bowen Yu, Chen Cheng, Da-Wei Liu, De mei Li, Hang Zhang, Hao Meng, Hu Wei, Ji-Li Ni, Kai Chen, Kuan Cao, Liang Peng, Lin Qu, Min Wu, Peng Wang, Shuting Yu, Tingkun Wen, Wens...

    Pith/arXiv arXiv 2025

  38. [38]

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

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation, 2023. 5

    2023

  39. [39]

    Perflow: Piecewise rectified flow as universal plug-and-play accelerator.Advances in Neu- ral Information Processing Systems, 37:78630–78652, 2024

    Hanshu Yan, Xingchao Liu, Jiachun Pan, Jun Hao Liew, Qiang Liu, and Jiashi Feng. Perflow: Piecewise rectified flow as universal plug-and-play accelerator.Advances in Neu- ral Information Processing Systems, 37:78630–78652, 2024. 3

    2024

  40. [40]

    Cogvideox: Text-to-video diffusion models with an ex- pert 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, Bin Xu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an ex- pert transformer. InThe Thirteenth International Conference on Learning Repres...

    2025

  41. [41]

    DiTFastattn: Attention compression for diffusion transformer models

    Zhihang Yuan, Hanling Zhang, Lu Pu, Xuefei Ning, Linfeng Zhang, Tianchen Zhao, Shengen Yan, Guohao Dai, and Yu Wang. DiTFastattn: Attention compression for diffusion transformer models. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3

    2024

  42. [42]

    Token pruning for caching better: 9 times acceleration on stable diffusion for free, 2024

    Evelyn Zhang, Bang Xiao, Jiayi Tang, Qianli Ma, Chang Zou, Xuefei Ning, Xuming Hu, and Linfeng Zhang. Token pruning for caching better: 9 times acceleration on stable diffusion for free, 2024. 3

    2024

  43. [43]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric, 2018. 5

    2018

  44. [44]

    Forecast then calibrate: Feature caching as ode for efficient diffusion transformers, 2025

    Shikang Zheng, Liang Feng, Xinyu Wang, Qinming Zhou, Peiliang Cai, Chang Zou, Jiacheng Liu, Yuqi Lin, Junjie Chen, Yue Ma, and Linfeng Zhang. Forecast then calibrate: Feature caching as ode for efficient diffusion transformers, 2025. 3

    2025

  45. [45]

    Compute only 16 tokens in one timestep: Accelerating Diffusion Transformers with Cluster-Driven Fea- ture Caching

    Zhixin Zheng, Xinyu Wang, Chang Zou, Shaobo Wang, and Linfeng Zhang. Compute only 16 tokens in one timestep: Accelerating Diffusion Transformers with Cluster-Driven Fea- ture Caching. InProceedings of the 33rd ACM International Conference on Multimedia (MM ’25), page to appear, Dublin, Ireland, 2025. ACM. 3

    2025

  46. [46]

    Dip-go: A diffusion pruner via few-step gradient optimization, 2024

    Haowei Zhu, Dehua Tang, Ji Liu, Mingjie Lu, Jintu Zheng, Jinzhang Peng, Dong Li, Yu Wang, Fan Jiang, Lu Tian, Span- dan Tiwari, Ashish Sirasao, Jun-Hai Yong, Bin Wang, and Emad Barsoum. Dip-go: A diffusion pruner via few-step gradient optimization, 2024. 3

    2024

  47. [47]

    Accelerating diffusion transformers with dual feature caching, 2024

    Chang Zou, Evelyn Zhang, Runlin Guo, Haohang Xu, Con- ghui He, Xuming Hu, and Linfeng Zhang. Accelerating diffusion transformers with dual feature caching, 2024. 3

    2024

  48. [48]

    Accelerating diffusion transformers with token- wise feature caching

    Chang Zou, Xuyang Liu, Ting Liu, Siteng Huang, and Lin- feng Zhang. Accelerating diffusion transformers with token- wise feature caching. InProceedings of the 13th International Conference on Learning Representations (ICLR 2025). ICLR,

    2025

  49. [50]

    Main Terms

    Method Details 6.1. Preliminary Diffusion Transformer Architecture.The Diffusion Trans- former (DiT) processes an input xt ={x i}H×W i=1 through a hierarchical structure G=g 1 ◦g 2 ◦ · · · ◦gL. Each block gl, where the superscript l denotes the layer index ranging from 1 to L, is composed as gl =F l SA ◦ Fl CA ◦ Fl MLP and sequentially applies self-attent...

  50. [51]

    Experiments Details 7.1. Model Configuration & Evaluation and Metrics We evaluate our proposed method on five representative large-scale diffusion models: FLUX.1-dev[ 11], FLUX.1- schnell[11], Qwen-Image[37], Qwen-Image-Lightning[37], and HunyuanVideo [10]. FLUX.1-dev and FLUX.1-schnellFor both FLUX.1-dev and FLUX.1-schnell, image generation is performed ...

  51. [2025]

    2, 3 10 LESA: Learnable Stage-Aware Predictors for Diffusion Model Acceleration Supplementary Material

    accepted to ICLR 2025. 2, 3 10 LESA: Learnable Stage-Aware Predictors for Diffusion Model Acceleration Supplementary Material

    2025