Pith. sign in

REVIEW 3 major objections 5 minor 54 references

This paper claims that the attention patterns in video diffusion transformers are artifacts of channel-level spatial and temporal correlations, making 85% of self-attention computation reusable with near-zero quality loss.

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

2026-08-03 22:06 UTC pith:PKYNM4LJ

load-bearing objection Timeripple is a well-motivated attention-reuse scheme for vDiTs with a genuinely new channel-level analysis, but the similarity metric in §3.3 is underspecified and the headline speedup is an estimate, so the paper needs careful revision before the claims can be trusted. the 3 major comments →

arxiv 2511.12035 v2 pith:PKYNM4LJ submitted 2025-11-15 cs.AR cs.CV

Timeripple: Accelerating vDiTs by Understanding the Spatio-Temporal Correlations in Latent Space

Wenxuan Miao , Yulin Sun , Aiyue Chen , Jing Lin , Yiwu Yao , Yiming Gan , Jieru Zhao , Jingwen Leng
show 2 more authors
Minyi Guo Yu Feng
This is my paper
classification cs.AR cs.CV
keywords video diffusion transformerself-attention accelerationspatio-temporal correlationpartial attention score reuselatent space redundancytraining-free accelerationadaptive thresholdVBench
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.

This paper tries to establish that the self-attention computation dominating video diffusion transformers (vDiTs) is largely redundant. The authors show that the varied patterns in vDiT attention maps are not intrinsic to attention, but are produced by spatial and temporal correlations among tokens along the channel dimension. They then propose a training-free reuse strategy: when a token's channel-window values are close to a neighbor's along the time, x, or y axis, the neighbor's partial attention score is reused instead of recomputed. On four open vDiT models they report up to 85% fewer self-attention operations with under 0.06% VBench quality loss, and an estimated up to 2.7x end-to-end speedup. Because self-attention accounts for about 78% of vDiT inference time, this targets the dominant bottleneck without retraining.

Core claim

On the paper's own terms, the central claim is that the spatial and temporal structure of video latents is so strong that per-channel dot products forming the attention map are locally constant along the time, x, and y axes, in both query and key. Because rotary position embedding splits channels into those three semantic groups, tokens adjacent along one axis have near-identical values in the corresponding channels, so their partial attention scores are nearly identical. The paper argues that a cheap standard-error test (window size 2) can mark such tokens as reusable, allowing up to 85% of dot-product work to be skipped while the final video stays essentially unchanged (VBench drop under 0

What carries the argument

The mechanism is channel-windowed partial-score reuse. For each query and key token, the method computes the standard error across a small channel window with its immediate neighbor along the temporal, x, and y axes (Eq. 3); passing any axis's threshold marks the token reusable. Reusable tokens borrow the neighbor's already-computed partial dot product channel-by-channel, while only non-reusable tokens are sparsely computed, using an OR-aggregated mask to decide what to skip. A second component, the adaptive threshold schedule (Eq. 4), scales the reuse threshold linearly between i_min and i_max denoising steps, matching the observed decay of reconstruction error so the induced error stays ro

Load-bearing premise

The load-bearing premise is that tokens whose channel-window values are within a small standard error of a neighbor produce partial attention scores close enough that reusing them — under a threshold schedule calibrated on about ten prompts — keeps output error negligible across the full 950-prompt benchmark suite.

What would settle it

A concrete check: run a vDiT with the reuse method on a fixed prompt set, and for every reused token compute the true partial attention score it replaced; if more than a few percent of reused scores differ from the true values by an amount larger than the delta test implies, the reuse criterion is failing. Alternatively, a real kernel executing the exact sparse pattern should reproduce the estimated speedup on an H100; an end-to-end wall-clock speedup well below the claimed 2.7x would show the performance gain is not realized.

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

If this is right

  • The reuse is training-free and only assumes a RoPE layout with time/x/y channel groups, so it transfers across vDiT architectures; the paper demonstrates this on four open models.
  • At equal compute savings, reuse keeps an order of magnitude lower output MSE than token masking, implying that redundant attention work should be reused, not zeroed.
  • Because reuse happens inside a single self-attention pass with no cross-step caching, its memory overhead is small compared with methods that store intermediate features.
  • The threshold schedule depends on the denoising step, not the prompt, so one table of thresholds applies across inputs, simplifying deployment.
  • Reuse is compatible with existing sparse masking: combining them yields a further speedup with only minor quality loss, showing the two strategies are complementary.

Where Pith is reading between the lines

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

  • We infer that the reported speedup, estimated by proportionally reducing measured self-attention latency, will only become a wall-clock win once a kernel or IO-aware algorithm implements the unstructured sparse pattern; until then the 2.7x figure is prospective.
  • We infer that the same channel-correlation reuse principle could transfer to other RoPE-based generative models with axis-wise channel groups, such as image diffusion transformers, a direction the paper does not pursue.
  • We infer that the threshold schedule's prompt-independence, observed on about ten prompts, deserves a wider calibration test before relying on it across the full 950-prompt benchmark suite.
  • We infer that the single-axis, window-size-2 similarity test is conservative; a joint spatial-temporal tolerance could yield higher savings, though the paper does not evaluate this.

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 / 5 minor

Summary. The paper proposes Timeripple, a training-free technique to accelerate self-attention in video diffusion transformers (vDiTs) by reusing partial attention scores for tokens that are spatially or temporally correlated along the channel dimension. The authors first analyze attention patterns and attribute them to channel-level spatio-temporal correlations, then design a reuse scheme based on a thresholded standard-error metric, an adaptive per-step threshold schedule (Eq. 4), and evaluate on four vDiTs against several baselines. They report up to 85% self-attention savings, near-identical VBench scores, and up to 2.7x end-to-end speedup.

Significance. The core idea — exploiting channel-level spatio-temporal correlations in latent space rather than imposing attention-map sparsity patterns from LLMs — is timely and potentially influential. The paper includes a broad evaluation on four public vDiTs with 950 VBench prompts, multiple quality metrics, and per-model hyperparameter tables, which is a strength. If the reuse predicate is made precise and the speedup is validated on real hardware, this could be a meaningful step toward efficient video diffusion inference. However, as written, the load-bearing claims rest on an under-specified reuse metric and an estimated, not measured, end-to-end speedup.

major comments (3)
  1. [§3.3, Eq. (3)] The reuse condition is not defined as a pairwise similarity. Eq. (3) computes the standard deviation of a single K-channel window a, but the text says "compute the standard error ∆ between every two adjacent frames" and "Token pairs with ∆ below a predefined threshold." If a is a single token's channel window, two tokens can both have small Δ yet very different means, producing very different partial attention scores. If a is instead the elementwise difference of two tokens, then any constant-offset pair (e.g., [1,1,1] vs. [101,101,101]) has Δ=0 even though the partial-score error is large. The manuscript provides no bound on the approximation error in the attention map as a function of Δ. This is load-bearing: the 85% savings claim presupposes that the threshold condition identifies pairs whose partial scores are actually close. The authors must specify exactly which vector enters Eq. (
  2. [§4.2, Table 2 (Performance)] The reported speedups are not measured end-to-end. The text states: "we estimate speedup by proportionally reducing the self-attention latency based on the amount of computation reduced," because no existing kernel supports the resulting unstructured sparsity. Thus the Latency and Speedup columns for Timeripple rows (e.g., 260.85 s and 2.66x for HunyuanVideo) are synthetic, not wall-clock results from an implementation. Since self-attention is only ~78% of runtime, the headline "up to 2.7× end-to-end speedup" (Abstract, Introduction) is contingent on a kernel that does not exist. Please relabel these as estimated/theoretical speedups, give the exact formula used, report the overhead of the reuse checks, and either temper the abstract/conclusion claims or implement a prototype to measure actual latency.
  3. [§3.3, Fig. 9, Eq. (4)] The adaptive framework is described as an "analytical model," but Eq. (4) is a linear interpolation with four per-model hyperparameters (Table 1) selected so that the measured per-step MSE is flat: the text says "we select threshold values such that they induce the same MSE at each step." This is a fitted calibration, not a predictive model derived from attention error statistics. The manuscript should explicitly call Eq. (4) a calibrated schedule, report sensitivity of results to the Table 1 hyperparameters, and demonstrate that the schedule generalizes beyond the ~10 prompts used in Figs. 8–9 to the full 950-prompt benchmark.
minor comments (5)
  1. [Fig. 7] The baselines Mask-1 and Mask-2 are not defined in the text or caption, and the MSE metric is not specified (over tokens? frames? prompts?). Please clarify.
  2. [References] Citations are inconsistent: HunyuanVideo is cited as [17] in Table 1 and as [37] in §3.2; CogVideo appears as [14] while CogVideoX is [45]. Please unify the citation style.
  3. [Eq. (3)] The formula computes a standard deviation, but the text calls it "standard error." Also, the notation does not define K or explain how a window is constructed along the temporal, x, or y axes. The threshold values θ_x and θ_y mentioned in the text are not listed in Table 1.
  4. [Abstract and Table 2] The abstract states "<0.06% loss on VBench" without specifying the aggregation; Table 2 shows a 0.14-point loss on Wan2.1 (81.17 vs. 81.03). Report the average or the per-model values explicitly.
  5. [Table 2] For Timeripple rows, the "Theoretical Speedup" and "Speedup" columns are identical, which is confusing because the text says the speedup is estimated. Please add a footnote or legend making clear that these are estimates, not wall-clock measurements.

Circularity Check

0 steps flagged

No significant circularity: threshold calibration is disclosed and quality metrics are measured independently; Eq. (3) ambiguity is a correctness issue, not circularity.

full rationale

No significant circularity in the claimed derivation chain. The paper proposes an empirical acceleration method: §3.2 attributes attention-map patterns to channel-level spatio-temporal correlations; §3.3 defines a reuse rule based on a threshold Δ<θ (Eq. 3); Eq. (4) sets per-step thresholds with hyperparameters in Table 1, chosen from the MSE-vs-timestep trend measured on 10 prompts (Fig. 9); and §4 evaluates the method on the 950-prompt VBench suite plus frame-wise image metrics. None of these steps reduces by construction to the claimed result. The per-step threshold schedule is calibrated to equalize MSE, but the headline claims (85% compute savings, <0.06% VBench loss, up to 2.7× speedup) are measured outputs, not consequences of the calibration formula: VBench scores are not fed back into threshold selection, and the per-step MSE is not the same object as the final quality metric. This is standard hyperparameter selection on a small prompt subset followed by held-out evaluation on a large benchmark, not a fitted parameter renamed as a prediction. The only self-citation ([9], real-time gaze tracking) appears in related work and is not load-bearing. The paper's limitation section explicitly discloses that the speedup is estimated by proportionally reducing measured self-attention latency because no kernel supports the unstructured sparsity; that is an unmeasured constant, not a circular step. The ambiguity in Eq. (3)—whether Δ is an intra-window standard deviation or a pairwise difference—is a legitimate reproducibility/correctness concern about whether Δ bounds the approximation error, but it is not a case where the result is defined in terms of its inputs. Overall, the derivation is self-contained against external benchmarks and no circular step was found.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central claim rests on four per-model threshold hyperparameters (Table 1), a hand-chosen window size, a linear scheduling ansatz (Eq. 4), and three empirical transfer assumptions (token-value similarity → attention-score reuse validity; 10-prompt sensitivity trend → 950-prompt benchmark; the four studied vDiTs sharing RoPE channel-group structure). No new physical or architectural entities are posited. These are engineering calibration choices rather than derivations — typical for systems papers, but it means the '<0.06% loss' and '85% savings' numbers describe a tuned system, not a parameter-free prediction.

free parameters (3)
  • θ_t,max, θ_t,min, i_min, i_max per model = HunyuanVideo 0.2/0.5/10/20; Wan2.1 0.4/0.6/10/48; CogVideoX 0.2/0.5/10/28; Open-Sora-Plan 0.4/0.8/20/48
    Table 1 hyperparameters calibrate the linear threshold schedule (Eq. 4) so per-step MSE is roughly flat (Fig. 9). They are chosen per model with no described held-out tuning protocol; the quality claims describe a tuned system.
  • reuse window size K = 2
    Window size K=2 is chosen from the sensitivity sweep (Fig. 11); larger windows degrade either quality or savings. Hand-selected, not derived.
  • channel-wise threshold coefficient α (Sec. 4.4 variant) = not given (deferred to supplementary)
    Used only for the sensitivity variant in Table 4 with individually set channel thresholds; not part of the main claim but another hand-set scale.
axioms (5)
  • standard math Attention is computed as Softmax(QK^T/√d_k)V with Rotary Position Embedding — the standard vDiT formulation (Eqs. 1–2).
    Background math the method operates on; taken as given from prior work (HunyuanVideo, Wan2.1, CogVideoX, Open-Sora-Plan).
  • domain assumption RoPE channel groups carry distinct semantic roles — early channels encode temporal position, later channels encode x/y spatial position (§3.1: 'the d channels are often partitioned by semantic roles').
    The whole reuse design (three check axes, channel groups) presupposes this structural fact about the four models. Stated, not proven; the manipulative-reuse visuals (Fig. 5) are the supporting evidence.
  • domain assumption Δ-similarity of token channel windows (Eq. 3) transfers to validity of partial attention-score reuse with bounded final-video error.
    This transfer from input-token similarity to attention-output error is the load-bearing premise of the reuse idea; supported only empirically by Fig. 7's MSE comparison on one model and the quality tables.
  • domain assumption The MSE-vs-denosing-step sensitivity trend measured on ~10 prompts (Figs. 8–9) generalizes to the full 950-prompt VBench suite.
    The adaptive schedule (Eq. 4) is derived from this trend. If the trend is prompt-dependent in deployment, the quality-equivalence claim fails. No held-out validation is reported.
  • ad hoc to paper The linear-in-i threshold interpolation (Eq. 4), with θ_t,max applied for steps after i_max, is an adequate error-control model.
    The 'analytical model' is a hand-picked linear schedule fitted to the observed sensitivity curve, not derived. The i_min/i_max values in the text (§3.3: 11/21) disagree with Table 1 (10/20).

pith-pipeline@v1.3.0-alltime-deepseek · 13480 in / 21822 out tokens · 191091 ms · 2026-08-03T22:06:41.444754+00:00 · methodology

0 comments
read the original abstract

The recent surge in video generation has shown the growing demand for high-quality video synthesis using large vision models. Existing video generation models are predominantly based on the video diffusion transformer (vDiT), however, they suffer from substantial inference delay due to self-attention. While prior studies have focused on reducing redundant computations in self-attention, they often overlook the inherent spatio-temporal correlations in video streams and directly leverage sparsity patterns from large language models to reduce attention computations. In this work, we take a principled approach to accelerate self-attention in vDiTs by leveraging the spatio-temporal correlations in the latent space. We show that the attention patterns within vDiT are primarily due to the dominant spatial and temporal correlations at the token channel level. Based on this insight, we propose a lightweight and adaptive reuse strategy that approximates attention computations by reusing partial attention scores of spatially or temporally correlated tokens along individual channels. We demonstrate that our method achieves significantly higher computational savings (85\%) compared to state-of-the-art techniques over 4 vDiTs, while preserving almost identical video quality ($<$0.06\% loss on VBench).

Figures

Figures reproduced from arXiv: 2511.12035 by Aiyue Chen, Jieru Zhao, Jing Lin, Jingwen Leng, Minyi Guo, Wenxuan Miao, Yiming Gan, Yiwu Yao, Yu Feng, Yulin Sun.

Figure 1
Figure 1. Figure 1: An illustration of spatial and temporal patterns in one head of multi-head attention maps. Due to the space limit, only 4 frames [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Examples of attention maps with different patterns. For visualization, we present only a fraction ( [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The overview of vDiT architectures. A vDiT consists of [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The execution breakdown of four popular vDiT models [ [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The effects of how different channel groups govern the final generation quality. Here, we maliciously reuse [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The overview of our reuse method. 1) Both [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: MSE comparison be￾tween our reusing method and two masking baselines. Our loss is much lower. 0 10 20 30 40 50 Timestep 10 2 10 3 10 4 10 5 MSE Loss [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: The average reuse technique accuracy across the denoising [PITH_FULL_IMAGE:figures/full_fig_p005_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The qualitative evaluations of TIMERIPPLE against other methods. “→” shows the major artifacts in prior studies. onward, and TIMERIPPLE 75%+SVG70%, which combines our reuse strategy with sparse masking. Video Quality. TIMERIPPLE achieves almost the best VBench scores across all models and baselines. On Hun￾yuanVideo, TIMERIPPLE 85% surpasses the original model, possibly due to regularization effects. Howe… view at source ↗
Figure 11
Figure 11. Figure 11: Sensitivity of our method to the reuse window size. [PITH_FULL_IMAGE:figures/full_fig_p008_11.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

54 extracted references · 24 linked inside Pith

  1. [1]

    Image and video compression standards: algorithms and architec- tures

    Vasudev Bhaskaran and Konstantinos Konstantinides. Image and video compression standards: algorithms and architec- tures. 1997. 1

    1997

  2. [2]

    Token merging for fast sta- ble diffusion

    Daniel Bolya and Judy Hoffman. Token merging for fast sta- ble diffusion. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4599–4603,

  3. [3]

    Delta-dit: A training-free acceleration method tailored for diffusion transformers.arXiv preprint arXiv:2406.01125, 2024

    Pengtao Chen, Mingzhu Shen, Peng Ye, Jianjian Cao, Chongjun Tu, Christos-Savvas Bouganis, Yiren Zhao, and Tao Chen. Delta-dit: A training-free acceleration method tailored for diffusion transformers.arXiv preprint arXiv:2406.01125, 2024. 6, 7

    Pith/arXiv arXiv 2024

  4. [4]

    Generating long sequences with sparse transformers.arXiv preprint arXiv:1904.10509, 2019

    Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers.arXiv preprint arXiv:1904.10509, 2019. 2

    Pith/arXiv arXiv 1904

  5. [5]

    Diffusion models in vision: A survey

    Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 45(9):10850–10869, 2023. 2

    2023

  6. [6]

    Flashattention-2: Faster attention with bet- ter parallelism and work partitioning.arXiv preprint arXiv:2307.08691, 2023

    Tri Dao. Flashattention-2: Faster attention with bet- ter parallelism and work partitioning.arXiv preprint arXiv:2307.08691, 2023. 8, 9

    Pith/arXiv arXiv 2023

  7. [7]

    Flashattention: Fast and memory-efficient exact at- tention with io-awareness.Advances in Neural Information Processing Systems, 35:16344–16359, 2022

    Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christo- pher R´e. Flashattention: Fast and memory-efficient exact at- tention with io-awareness.Advances in Neural Information Processing Systems, 35:16344–16359, 2022. 5, 8, 9

    2022

  8. [8]

    Efficient-vdit: Ef- ficient video diffusion transformers with attention tile.arXiv preprint arXiv:2502.06155, 2025

    Hangliang Ding, Dacheng Li, Runlong Su, Peiyuan Zhang, Zhijie Deng, Ion Stoica, and Hao Zhang. Efficient-vdit: Ef- ficient video diffusion transformers with attention tile.arXiv preprint arXiv:2502.06155, 2025. 1, 2, 3

    Pith/arXiv arXiv 2025

  9. [9]

    Real-time gaze tracking with event- driven eye segmentation

    Yu Feng, Nathan Goulding-Hotta, Asif Khan, Hans Reyser- hove, and Yuhao Zhu. Real-time gaze tracking with event- driven eye segmentation. In2022 IEEE Conference on Vir- tual Reality and 3D User Interfaces (VR), pages 399–408. IEEE, 2022. 1

    2022

  10. [10]

    Veo 2: Our state-of-the-art video generation model,

    Google. Veo 2: Our state-of-the-art video generation model,

  11. [11]

    Boot: Data-free distillation of denois- ing diffusion models with bootstrapping

    Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Lingjie Liu, and Joshua M Susskind. Boot: Data-free distillation of denois- ing diffusion models with bootstrapping. InICML 2023 Workshop on Structured Probabilistic Inference&Genera- tive Modeling, 2023. 1

    2023

  12. [12]

    Mod- ern lossless compression techniques: Review, comparison and analysis

    Apoorv Gupta, Aman Bansal, and Vidhi Khanduja. Mod- ern lossless compression techniques: Review, comparison and analysis. In2017 Second International Conference on Electrical, Computer and Communication Technologies (ICECCT), pages 1–8. IEEE, 2017. 1

    2017

  13. [13]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 2

    2020

  14. [14]

    Cogvideo: Large-scale pretraining for text-to-video generation via transformers.arXiv preprint arXiv:2205.15868, 2022

    Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers.arXiv preprint arXiv:2205.15868, 2022. 1, 3, 4

    Pith/arXiv arXiv 2022

  15. [15]

    Vbench: Comprehensive bench- mark suite for video generative models

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive bench- mark suite for video generative models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807–21818, 2024. 6

    2024

  16. [16]

    Minference 1.0: Accel- erating pre-filling for long-context llms via dynamic sparse attention.Advances in Neural Information Processing Sys- tems, 37:52481–52515, 2024

    Huiqiang Jiang, Yucheng Li, Chengruidong Zhang, Qian- hui Wu, Xufang Luo, Surin Ahn, Zhenhua Han, Amir Abdi, Dongsheng Li, Chin-Yew Lin, et al. Minference 1.0: Accel- erating pre-filling for long-context llms via dynamic sparse attention.Advances in Neural Information Processing Sys- tems, 37:52481–52515, 2024. 2, 6, 7

    2024

  17. [17]

    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. 2, 6, 7, 8

    Pith/arXiv arXiv 2024

  18. [18]

    Kuaishou Unveils Proprietary Video Generation Model ‘Kling’; Testing Now Available, 2024

    Kuaishou. Kuaishou Unveils Proprietary Video Generation Model ‘Kling’; Testing Now Available, 2024. 1

    2024

  19. [19]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. InPro- ceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023. 9

    2023

  20. [20]

    Autodiffusion: Training-free optimization of time steps and architectures for automated diffusion model acceleration

    Lijiang Li, Huixia Li, Xiawu Zheng, Jie Wu, Xuefeng Xiao, Rui Wang, Min Zheng, Xin Pan, Fei Chao, and Rongrong Ji. Autodiffusion: Training-free optimization of time steps and architectures for automated diffusion model acceleration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7105–7114, 2023. 1, 4

    2023

  21. [21]

    Open-sora plan: Open-source large video generation model.arXiv preprint arXiv:2412.00131, 2024

    Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Li- uhan Chen, et al. Open-sora plan: Open-source large video generation model.arXiv preprint arXiv:2412.00131, 2024. 1, 2, 3, 4, 6, 7

    Pith/arXiv arXiv 2024

  22. [22]

    Flow matching for generative mod- eling.arXiv preprint arXiv:2210.02747, 2022

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matt Le. Flow matching for generative mod- eling.arXiv preprint arXiv:2210.02747, 2022. 2

    Pith/arXiv arXiv 2022

  23. [23]

    Rec-gpt4v: Multimodal recommendation with large vision- language models.arXiv preprint arXiv:2402.08670, 2024

    Yuqing Liu, Yu Wang, Lichao Sun, and Philip S Yu. Rec-gpt4v: Multimodal recommendation with large vision- language models.arXiv preprint arXiv:2402.08670, 2024. 1

    Pith/arXiv arXiv 2024

  24. [24]

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

    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,

  25. [25]

    Dpm-solver++: Fast solver for guided 9 sampling of diffusion probabilistic models.arXiv preprint arXiv:2211.01095, 2022

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongx- uan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided 9 sampling of diffusion probabilistic models.arXiv preprint arXiv:2211.01095, 2022. 2

    Pith/arXiv arXiv 2022

  26. [26]

    Image and video compres- sion with neural networks: A review.IEEE Transactions on Circuits and Systems for Video Technology, 30(6):1683– 1698, 2019

    Siwei Ma, Xinfeng Zhang, Chuanmin Jia, Zhenghui Zhao, Shiqi Wang, and Shanshe Wang. Image and video compres- sion with neural networks: A review.IEEE Transactions on Circuits and Systems for Video Technology, 30(6):1683– 1698, 2019. 1

    2019

  27. [27]

    Sora: Bring your imagination to life with text, im- age, or video, 2024

    OpenAI. Sora: Bring your imagination to life with text, im- age, or video, 2024. 1

    2024

  28. [28]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 4195–4205,

  29. [29]

    Runway Gen 2: The Next Step Forward for Gener- ative AI, 2024

    Runway. Runway Gen 2: The Next Step Forward for Gener- ative AI, 2024. 1

    2024

  30. [30]

    Progressive distillation for fast sampling of diffusion models.arXiv preprint arXiv:2202.00512, 2022

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models.arXiv preprint arXiv:2202.00512, 2022. 1

    Pith/arXiv arXiv 2022

  31. [31]

    Seaweed-7b: Cost-effective train- ing of video generation foundation model.arXiv preprint arXiv:2504.08685, 2025

    Team Seawead, Ceyuan Yang, Zhijie Lin, Yang Zhao, Shanchuan Lin, Zhibei Ma, Haoyuan Guo, Hao Chen, Lu Qi, Sen Wang, et al. Seaweed-7b: Cost-effective train- ing of video generation foundation model.arXiv preprint arXiv:2504.08685, 2025. 1

    Pith/arXiv arXiv 2025

  32. [32]

    Make-a-video: Text-to-video generation without text-video data.arXiv preprint arXiv:2209.14792,

    Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data.arXiv preprint arXiv:2209.14792,

    Pith/arXiv arXiv

  33. [33]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 2

    Pith/arXiv arXiv 2010

  34. [34]

    Score-based generative modeling through stochastic differential equa- tions.arXiv preprint arXiv:2011.13456, 2020

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equa- tions.arXiv preprint arXiv:2011.13456, 2020. 2

    Pith/arXiv arXiv 2011

  35. [35]

    Mochi 1.https :/ /github .com/ genmoai/models, 2024

    Genmo Team. Mochi 1.https :/ /github .com/ genmoai/models, 2024. 1

    2024

  36. [36]

    Vidi: Large multimodal models for video un- derstanding and editing.arXiv preprint arXiv:2504.15681,

    Vidi Team, Celong Liu, Chia-Wen Kuo, Dawei Du, Fan Chen, Guang Chen, Jiamin Yuan, Lingxi Zhang, Lu Guo, Lusha Li, et al. Vidi: Large multimodal models for video un- derstanding and editing.arXiv preprint arXiv:2504.15681,

    Pith/arXiv arXiv

  37. [37]

    Tencent launches and open-sources Hunyuan video-generation model, 2024

    Tencent. Tencent launches and open-sources Hunyuan video-generation model, 2024. 1, 3, 4

    2024

  38. [38]

    Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    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, Pandeng Li, Pingyu Wu, Ruihang Chu, Ruili Feng, Shiwei Zhang, Siyang Sun, Tao Fang, Tianxing Wa...

    Pith/arXiv arXiv 2025

  39. [39]

    Towards real-time multi-object tracking

    Zhongdao Wang, Liang Zheng, Yixuan Liu, Yali Li, and Shengjin Wang. Towards real-time multi-object tracking. In European conference on computer vision, pages 107–122. Springer, 2020. 1

    2020

  40. [40]

    Sparse videogen: Accelerating video diffusion transformers with spatial-temporal sparsity.arXiv preprint arXiv:2502.01776, 2025

    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. 1, 2, 3, 5, 6, 7

    Pith/arXiv arXiv 2025

  41. [41]

    Training-free and adaptive sparse attention for efficient long video generation.arXiv preprint arXiv:2502.21079, 2025

    Yifei Xia, Suhan Ling, Fangcheng Fu, Yujie Wang, Huixia Li, Xuefeng Xiao, and Bin Cui. Training-free and adaptive sparse attention for efficient long video generation.arXiv preprint arXiv:2502.21079, 2025. 1, 2, 3

    Pith/arXiv arXiv 2025

  42. [42]

    Efficient streaming language models with attention sinks.arXiv preprint arXiv:2309.17453, 2023

    Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks.arXiv preprint arXiv:2309.17453, 2023. 2

    Pith/arXiv arXiv 2023

  43. [43]

    Easyanimate: A high-performance long video generation method based on transformer architecture.arXiv preprint arXiv:2405.18991,

    Jiaqi Xu, Xinyi Zou, Kunzhe Huang, Yunkuo Chen, Bo Liu, MengLi Cheng, Xing Shi, and Jun Huang. Easyanimate: A high-performance long video generation method based on transformer architecture.arXiv preprint arXiv:2405.18991,

    arXiv

  44. [44]

    Diffusion models: A comprehensive survey of methods and applications.ACM Computing Surveys, 56(4): 1–39, 2023

    Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Run- sheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming- Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications.ACM Computing Surveys, 56(4): 1–39, 2023. 2

    2023

  45. [45]

    Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiao- han Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024. 2, 6, 7

    Pith/arXiv arXiv 2024

  46. [46]

    Flashinfer: Efficient and customizable attention engine for llm inference serving

    Zihao Ye, Lequn Chen, Ruihang Lai, Wuwei Lin, Yineng Zhang, Stephanie Wang, Tianqi Chen, Baris Kasikci, Vinod Grover, Arvind Krishnamurthy, et al. Flashinfer: Efficient and customizable attention engine for llm inference serving. arXiv preprint arXiv:2501.01005, 2025. 9

    Pith/arXiv arXiv 2025

  47. [47]

    One-step diffusion with distribution matching distillation

    Tianwei Yin, Micha ¨el Gharbi, Richard Zhang, Eli Shecht- man, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 6613–6623, 2024. 1

    2024

  48. [48]

    Ditfastattn: Attention compression for diffusion transformer models.Advances in Neural Information Pro- cessing Systems, 37:1196–1219, 2024

    Zhihang Yuan, Hanling Zhang, Lu Pu, Xuefei Ning, Lin- feng Zhang, Tianchen Zhao, Shengen Yan, Guohao Dai, and Yu Wang. Ditfastattn: Attention compression for diffusion transformer models.Advances in Neural Information Pro- cessing Systems, 37:1196–1219, 2024. 1, 2, 3

    2024

  49. [49]

    Make pixels dance: High- dynamic video generation

    Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High- dynamic video generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8850–8860, 2024. 1

    2024

  50. [50]

    H2o: Heavy-hitter ora- 10 cle for efficient generative inference of large language mod- els.Advances in Neural Information Processing Systems, 36: 34661–34710, 2023

    Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher R´e, Clark Barrett, et al. H2o: Heavy-hitter ora- 10 cle for efficient generative inference of large language mod- els.Advances in Neural Information Processing Systems, 36: 34661–34710, 2023. 2

    2023

  51. [51]

    Real-time video generation with pyramid attention broad- cast.arXiv preprint arXiv:2408.12588, 2024

    Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. Real-time video generation with pyramid attention broad- cast.arXiv preprint arXiv:2408.12588, 2024. 2, 5, 6, 7

    Pith/arXiv arXiv 2024

  52. [52]

    Efficiently program- ming large language models using sglang

    Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Jeff Huang, Chuyue Sun, Cody Hao Yu, Shiyi Cao, Christos Kozyrakis, Ion Stoica, Joseph E Gonzalez, et al. Efficiently program- ming large language models using sglang. 2023. 9

    2023

  53. [53]

    Open-sora: Democratizing efficient video production for all, 2024

    Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1

    2024

  54. [54]

    Accelerating diffusion transformers with token- wise feature caching.arXiv preprint arXiv:2410.05317,

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

    Pith/arXiv arXiv