Pith. sign in

REVIEW 5 major objections 5 minor 105 references

Video diffusion self-attention can be cut by up to 85% by reusing partial results across similar latent channels, with a co-designed accelerator reaching 5.9x speedup.

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 03:45 UTC pith:AYYBB4T4

load-bearing objection A well-executed co-design with plausible speedup claims, but the headline numbers are conditional on a reuse mapping and thresholds tuned on the same models used for validation. the 5 major comments →

arxiv 2607.13770 v1 pith:AYYBB4T4 submitted 2026-07-15 cs.AR cs.AI

Kaleido: Algorithm-Hardware Co-Design for Video Diffusion Transformers by Exploiting Latent Space Correlations

Wenxuan Miao , Haosong Liu , Weiming Hu , Zihan Liu , Aiyue Chen , Jianlin Yu , Yiwu Yao , Yiming Gan
show 4 more authors
Jieru Zhao Jingwen Leng Minyi Guo Yu Feng
This is my paper
classification cs.AR cs.AI
keywords video diffusion transformersalgorithm-hardware co-designchannel-wise reuseself-attention accelerationrotary position embeddingspatio-temporal correlationslatent space reusediffusion accelerator
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 sets out to show that self-attention, not the number of denoising timesteps, is the real bottleneck in modern video diffusion transformers, and that this attention is not randomly sparse but structured by how token channels encode space and time. It claims each channel group (temporal, horizontal, vertical) carries a distinct positional signal, so a token that is similar to its neighbor in that channel can inherit the neighbor's partial attention score instead of recomputing it. On that basis it proposes a lightweight channel-wise reuse algorithm and a matching accelerator, and reports that the combination preserves near-reference video quality while cutting up to 85% of operations and achieving up to 5.9x speedup and 16.0x energy savings over prior accelerator designs. A sympathetic reader would care because this is a path to fast video generation that does not sacrifice the dense, pixel-level consistency that token-dropping or sparse-attention methods tend to lose.

Core claim

The central claim is that the dominant attention patterns in vDiTs—repetitive spatial tiles and strong temporal diagonals—are caused by rotary position embedding applied channel-wise. Rather than skipping entries of the attention map, the paper reuses previously computed partial results: for each channel group, adjacent tokens along a predefined axis are compared; tokens closer than a first threshold inherit a full partial score, while tokens closer than a second threshold reuse the high bits and only compute the remaining bits. With a co-designed systolic-array-like accelerator and data dispatcher, the method accelerates four mainstream vDiTs with up to 5.9x speedup and 16.0x energy savings

What carries the argument

The load-bearing object is the channel-wise reuse algorithm plus the hardware that executes it. The algorithm splits token channels into t-, x-, and y-groups according to rotary position embedding (RoPE), the standard positional encoding that gives each channel group its own space-time frequency; it measures the absolute difference between adjacent tokens along a fixed direction per group and, using two power-of-two thresholds, decides whether the second token fully reuses, partially reuses (high bits only), or recomputes the partial attention score of the first. The accelerator side is a reconfigurable processing element built on an 8-bit fixed-point multiply-accumulate unit with extra regi

Load-bearing premise

The load-bearing premise is that the t-, x-, and y-channel groups created by the position encoding keep their roles across all layers, timesteps, prompts, and models, so that reusing along the 'other two' directions is always the least damaging choice; if a layer or denoising stage violates that ordering, the fixed mapping injects errors that per-prompt averages can hide.

What would settle it

Run the same channel-wise reuse on a single vDiT using fast-motion prompts and report per-clip PSNR instead of the aggregate: a clip where adjacent frames genuinely differ but the reuse mask still labels many channels as reusable, with PSNR dropping below the reported range, would falsify the claimed universality of the reuse-direction mapping. A cheaper check is to swap a channel group's reuse axis to its own dimension on a held-out model and test whether the MSE ordering shown in the paper actually holds; if it flips, the root-cause story fails.

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

If this is right

  • If self-attention is the true bottleneck in modern vDiTs, then timestep-reduction techniques alone will hit a ceiling; attention-level reuse is where the remaining speedups are.
  • Channel-level reuse, because it reuses rather than drops computation, should generalize to any operation whose activations inherit the spatio-temporal correlations; the paper demonstrates this on linear projection layers as well as attention.
  • If the claimed 85% ideal operation reduction holds, the practical ~60% reduction is mostly a scheduling and mapping gap, so future hardware with reuse-mask-aware dataflow could approach the ideal without changing the algorithm.
  • At higher resolutions, the approach reports both higher speedup and higher quality, implying the benefit grows as token counts grow.
  • The design achieves 5.2-6.6x speedup and 14.8-18.4x energy savings over a current GPU baseline across the four evaluated models, including a distilled 4-step model.

Where Pith is reading between the lines

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

  • A testable corollary of the root-cause explanation is that a vDiT variant using a different positional encoding (or no RoPE) would not show the same tile-and-diagonal structure, so the fixed reuse-direction mapping would need to be re-tuned for that encoding.
  • Because reuse decisions are made per channel and per timestep, the thresholds could be made adaptive across the denoising schedule—tighter early where global coherence matters, looser later—without changing the hardware.
  • Since the technique targets a different redundancy than timestep caching, the two are complementary; a distilled 4-step model still benefits strongly, suggesting that stacking channel reuse with block caching could compound speedups.
  • The 'reuse, don't skip' principle may transfer to other dense generative workloads such as image or 3D generation, where token pruning degrades quality; the same channel-similarity argument could be tested on those models.

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

5 major / 5 minor

Summary. The paper proposes Kaleido, an algorithm-hardware co-design for accelerating video diffusion transformers (vDiTs). The central observation is that RoPE-based channel groups (t, x, y) encode different spatio-temporal information, and that latent tokens exhibit channel-wise correlations. On this basis, the paper introduces a channel-wise reuse algorithm: for each channel group, adjacent tokens along a fixed direction are compared, and if their values are sufficiently similar, the partial attention or activation product is reused, either fully or partially, using 8-bit fixed-point arithmetic with (5,3)-bit prefix thresholds. The hardware contribution is a systolic-array-like accelerator with reconfigurable processing elements that support the reuse dataflow, plus a data dispatcher with a scheduler and matcher to improve PE utilization. The design is evaluated with RTL synthesis in 16nm technology, cycle-level simulation, and comparison against GPUs and four prior accelerators on HunyuanVideo, Wan, CogVideoX, and TurboDiffusion. The paper reports up to 5.9x speedup and 16.0x energy savings over prior accelerators while retaining high PSNR, SSIM, LPIPS, and VBench scores.

Significance. If the claims hold, Kaleido is a meaningful contribution. It goes beyond LLM-style sparse attention by identifying a mechanism — RoPE-driven channel-wise spatio-temporal correlation — and turns that mechanism into a concrete algorithm and hardware implementation. The paper contains real engineering: RTL synthesis, post-layout area/power estimates, a cycle-level simulator, ablations separating the PE and dispatcher contributions, and sensitivity studies on PE array size, buffer size, thresholds, and fast-motion prompts. The reuse algorithm is simple and appears complementary to step-reduction and caching techniques. However, the evaluation as presented is not yet fully convincing. The quality comparison in Fig. 17 is not matched at equal compute reduction or speedup, a relevant baseline (Ditto) is excluded from the quality comparison, and the reuse mapping and thresholds are selected on the same models and metrics later used for the headline quality claims. These issues are correctable with additional experiments, but they currently prevent acceptance.

major comments (5)
  1. [§7.1, Fig. 17] The quality comparison is not at matched operating points. The PSNR/SSIM/LPIPS/VBench bars compare Kaleido against prior accelerators at their default configurations, but Fig. 26 shows that AdapTiV and Exion trade quality for speedup as their compression settings change. Without controlling compute reduction, operation count, or achieved speedup, the claim that Kaleido is '>17 dB' better partly reflects different operating points rather than algorithmic superiority. Please provide a matched-speedup or matched-operation-reduction quality comparison, or a quality-vs-speedup Pareto plot for all methods.
  2. [§7.1 and §6] Ditto is excluded from the quality comparison with the statement 'the algorithm proposed in Ditto is lossless.' If Ditto is truly lossless, it is the most relevant baseline for evaluating a quality-preserving acceleration scheme: a lossless method trivially dominates in quality, and the meaningful comparison is speed/energy at equal quality or quality degradation at equal speed. If Ditto is lossless only in a restricted sense, that should be stated and it should be included in Fig. 17. The current asymmetric treatment — Ditto appears in performance but not quality — makes the 'higher generative quality than prior methods' claim incomplete.
  3. [§4.1, Fig. 7, §7.4, Fig. 23] The reuse mapping (t→x, x→y, y→t) and thresholds ((θth1, θth2) = (5,3) bits) are selected using MSE measurements on the same four models and the same quality metrics that are later reported as evidence. Fig. 7 only establishes that, for a given channel group, reuse along the two non-own axes is better than reuse along the own axis; it does not uniquely justify the specific cyclic mapping, nor does the paper show that this mapping is stable across layers, denoising stages, prompts, or held-out models. Table 1 shows that the channel partitions differ across models, and Fig. 24 shows that fast-motion content changes the reuse ratio. The paper should provide a held-out-model evaluation or a per-layer/per-timestep breakdown of reuse ratio and error; without this, the headline speedup and quality numbers are conditional on in-sample tuning.
  4. [Abstract, §7.2] The headline 'up to 5.9x speedup and 16.0x energy savings over state-of-the-art accelerators' does not match the body. Section 7.2 reports 6.6x, 5.6x, 5.2x, and 6.1x speedup over A100, whose average is 5.9x, and 14.8x, 15.6x, 15.2x, and 18.4x energy savings over A100, whose average is 16.0x. Thus 'up to' is applied to averages, and the baseline is A100, not a state-of-the-art accelerator. The text also states that AdapTiV exhibits higher energy savings than Kaleido. Please correct the abstract and conclusion to state the actual baselines and whether the reported numbers are maximum or average.
  5. [§4.1, §5.2] The paper does not analyze how the reuse approximation error propagates. A 'full reuse' decision for two 8-bit values with the same top 5 bits permits a difference of up to 7 LSBs; after multiplication by another 8-bit value and accumulation over many tokens, the error in the partial attention score is not bounded by the 5-bit prefix. The paper states that the thresholds preserve the most significant bits, but it does not provide an analytical error bound or a per-layer/per-timestep error accumulation study. Since the end-to-end metrics in Fig. 17 are the only quality evidence, and since the thresholds are tuned on those same metrics, an explicit error analysis or layerwise validation would substantially strengthen the claim that the approximation is safe beyond the specific evaluated prompts.
minor comments (5)
  1. [Abstract] The abstract says 'three mainstream vDiT models,' but Section 6 and Section 7 evaluate four models (HunyuanVideo, Wan, CogVideoX, TurboDiffusion). Please align the wording.
  2. [§6] The hardware configuration description is ambiguous: '128×128 PE arrays, each with 4×4 PEs' and later 'baselines scaled to the same number of PEs, 512×512' should be clarified so the total PE count and array organization are immediately understandable.
  3. [Fig. 26] The configuration annotations (75%, 50%, 25%, etc.) are difficult to read. Please define them in the caption or in the text so the operating points are explicit.
  4. [§8] Related work includes TimeRipples [52], which appears to be a closely related study of spatio-temporal correlations in vDiT latent space. The paper should explicitly state the differences between TimeRipples and Kaleido, especially regarding the reuse mapping and the hardware support.
  5. [Fig. 17e] The qualitative comparison would benefit from zoomed crops to show the artifacts and detail differences; the current figure is difficult to inspect at page scale.

Circularity Check

0 steps flagged

No significant circularity; design choices are empirical but not definitionally tied to the reported metrics.

full rationale

The core derivation chain is empirical rather than logical. The reuse-direction mapping and thresholds are chosen from measurements (Fig. 7 and Fig. 23) on the same models and metrics later reported (Fig. 17), but this is standard in-sample hyperparameter selection rather than a constructional equivalence: the reported PSNR/VBench numbers are measured outputs of the resulting system, not quantities that are equal to the thresholds or mapping by definition. The hardware speedup and energy savings come from independent RTL synthesis and cycle-level simulation, with sensitivity studies that do not reduce to the algorithm's parameters. Self-citations (Astraea [42], TimeRipples [52]) appear only as related work and are not load-bearing for the main claim. No equation in the paper is shown to be equivalent to its own inputs, and no externally falsifiable result is replaced by a self-citation chain. The generalization concern (the fixed t->x, x->y, y->t mapping is only tested on four models) is a robustness limitation, not circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 2 invented entities

The central claim rests on two fitted design choices (threshold widths and the channel-to-axis reuse mapping), an unstated window size, and the domain assumption that RoPE channel groups encode separable axes. The hardware entities are internally simulated but have no external, falsifiable evidence beyond this paper.

free parameters (3)
  • theta_th1/theta_th2 bit-preservation widths = 5 bits / 3 bits
    Chosen via the sensitivity sweep in Fig. 23 over the same speedup/PSNR tradeoff later reported as evidence; not derived from theory.
  • Channel-group reuse direction mapping = t->x, x->y, y->t
    Selected from MSE measurements in Fig. 7 on the same quality metric used to claim PSNR preservation; a discrete design choice fitted to observed data.
  • Window size w for non-attention reuse (Eq. 5) = unspecified
    The window-based similarity metric needs a window size, but no value or sensitivity analysis is reported.
axioms (5)
  • domain assumption RoPE channel partition into t/x/y axes implies each channel group encodes only its own spatiotemporal direction, so reuse along other axes is benign.
    Central to the algorithm. Supported only by the MSE experiments in Fig. 7, with no formal proof or cross-model validation shown.
  • domain assumption Adjacent latent video tokens are similar enough at channel level that reusing high-bit partial results is perceptually lossless.
    Assumed for all layers, timesteps, prompts, and models in Sec. 4; partially tested on a subset of prompts.
  • ad hoc to paper 8-bit fixed-point with 4-bit partial computation has bounded error that does not accumulate across attention and MLP layers.
    Stated in Sec. 4.1 step 3; no numerical error analysis or bit-level simulation is provided.
  • domain assumption Custom accelerator baselines scaled to 512x512 PEs at 1 GHz in 16nm are comparable to real A100/H100 GPU measurements.
    Sec. 6: energy comparisons mix simulated custom chips with measured GPUs; process and memory-technology differences are not fully controlled.
  • ad hoc to paper One fixed reuse direction per channel group is sufficient for all operations.
    A hardware-simplification choice, acknowledged to reduce ideal 85% operation reduction to about 60% (Sec. 7.2).
invented entities (2)
  • Reconfigurable PE with R0/R1/R2 accumulation registers and mode-control logic no independent evidence
    purpose: Enables full/half reuse across 2x2 attention-tile groups and different input dataflows.
    Described in Sec. 5.2; evaluated only in the paper's own RTL/cycle-level simulation, no fabricated chip.
  • Data dispatcher composed of scheduler and matcher no independent evidence
    purpose: Clusters tokens by reuse-pattern Hamming distance and merges compatible channels to improve PE utilization.
    Described in Sec. 5.3; characterized only in the paper's own simulator, no independent measurement.

pith-pipeline@v1.3.0-alltime-deepseek · 29017 in / 14037 out tokens · 136844 ms · 2026-08-02T03:45:12.591451+00:00 · methodology

0 comments
read the original abstract

Video diffusion transformers (vDiTs) generate high quality video but introduce extremely high compute cost due to the long diffusion timesteps and self attention computation. As diffusion timesteps are reduced, the computation cost of self attention becomes the dominant bottleneck. Existing acceleration approaches largely inherit sparse attention techniques from large language models, which fail to consider the unique spatiotemporal correlation of video data. This paper presents Kaleido, an algorithm hardware codesign that accelerates all operations in vDiTs by exploiting channel-wise spatiotemporal correlations in latent space. Based on this insight, we propose a lightweight channelwise reuse algorithm that skips redundant computations by reusing partial results while preserving higher generative quality than prior methods (>17 dB). To efficiently support this algorithm, we design a systolic array like accelerator with reconfigurable processing elements and a lightweight data dispatcher to mitigate irregular sparsity and data access patterns introduced by our reuse algorithm. Evaluations across three mainstream vDiT models show that Kaleido achieves up to 5.9x speedup and 16.0x energy savings over state of the art accelerators.

Figures

Figures reproduced from arXiv: 2607.13770 by Aiyue Chen, Haosong Liu, Jianlin Yu, Jieru Zhao, Jingwen Leng, Minyi Guo, Weiming Hu, Wenxuan Miao, Yiming Gan, Yiwu Yao, Yu Feng, Zihan Liu.

Figure 1
Figure 1. Figure 1: An example of a diffusion process. The random-noised [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The execution breakdown of eight mainstream [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: The ratio between token sequence length and the num [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Example of various attention patterns, which can be [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: The MSE loss introduced by reusing different direc [PITH_FULL_IMAGE:figures/full_fig_p004_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: An overview of our channel-wise reuse algorithm, which consists of four steps. We first compute the similarity between [PITH_FULL_IMAGE:figures/full_fig_p005_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: MSE comparison be￾tween our reusing method and two skipping techniques. We pick 20 random prompts. 0 10 20 30 40 50 Diffusion Timestep 10 3 10 4 MSE Loss IQR [PITH_FULL_IMAGE:figures/full_fig_p006_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: The overall architecture design, which comprises a [PITH_FULL_IMAGE:figures/full_fig_p006_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: The computation order of self-attention computation when reusing along the t-axis. Each PE array is responsible for a group of tiles, as highlighted in colors. The numbers denote the computation order. 8-bit W1 4-bit A1 4-bit A2 <<4 ➕ ➕ ➕ ✖ M M M ➕ \ 4 \ 4 \ \ 8 8-bit 8 W2 ✖ M 0 0 [PITH_FULL_IMAGE:figures/full_fig_p007_12.png] view at source ↗
Figure 14
Figure 14. Figure 14: An overview of our reuse-aware dataflow for a single PE. In this example, 3 blocks in [PITH_FULL_IMAGE:figures/full_fig_p008_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Examples of two scheduling cases with identical data [PITH_FULL_IMAGE:figures/full_fig_p008_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: An overview of our data dispatcher design, which consists of a scheduler and a matcher. The scheduler groups tokens [PITH_FULL_IMAGE:figures/full_fig_p009_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: compares the video generation quality of Kaleido with the three approximation methods shown above. Across all four met￾rics, Kaleido consistently delivers the best quality. For instance, Kaleido improves reconstruction fidelity by a large margin in terms of PSNR. Kaleido achieves 29.9 dB, 24.5 dB, 24.4 dB, and 26.5 dB on HunyuanVideo, Wan, TurboDiffusion, and CogVideoX, respec￾tively. In particular, Kalei… view at source ↗
Figure 19
Figure 19. Figure 19: Ablation study of the different hardware components [PITH_FULL_IMAGE:figures/full_fig_p011_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Ablation study of per￾formance and quality on Hun￾yuanVideo, with and without applying Kaleido to linear projection layers. 2x2 4x4 8x8 16x16 32x32 64x64 # of PE 0.85 0.90 0.95 1.00 1.05 Speedup HunyuanVideo Wan TurboDiffusion [PITH_FULL_IMAGE:figures/full_fig_p011_20.png] view at source ↗
Figure 25
Figure 25. Figure 25: Scalability analysis on generating 720p high￾resolution videos. 4 5 6 7 8 Speedup 15 20 25 30 PSNR (dB) Qth = 15 EXION Kaleido AdapTiV 75% 50% 25% 50% 65% 70% Better [PITH_FULL_IMAGE:figures/full_fig_p012_25.png] view at source ↗
Figure 23
Figure 23. Figure 23: Sensitivity of speedup and generative quality to reuse thresholds, 𝜃th1 and𝜃th2. General Fast-Motion 0.0 2.0 4.0 6.0 8.0 Speedup Speedup PSNR 15 25 35 PSNR (dB) [PITH_FULL_IMAGE:figures/full_fig_p012_23.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

105 extracted references · 27 linked inside Pith

  1. [1]

    Akshay Kulkarni. 2026. Predis.ai Announces Launch of New State-of-the-Art (SOTA) AI Models for Ad Videos. https://www.manilatimes.net/2026/01/09/tmt- newswire/globenewswire/predisai-announces-launch-of-new-state-of-the- art-sota-ai-models-for-ad-videos/2256039

    arXiv 2026

  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]

    Mark Buckler, Philip Bedoukian, Suren Jayasuriya, and Adrian Sampson

  4. [4]

    Junsong Chen, Yuyang Zhao, Jincheng Yu, Ruihang Chu, Junyu Chen, Shuai Yang, Xianbang Wang, Yicheng Pan, Daquan Zhou, Huan Ling, et al. 2025. SANA- Video: Efficient Video Generation with Block Linear Diffusion Transformer. (2025). arXiv:2509.24695 [cs.CV] https://arxiv.org/abs/2509.24695

    arXiv 2025

  5. [5]

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

    Pith/arXiv arXiv 2024

  6. [6]

    Yu-Hsin Chen, Tien-Ju Yang, Joel Emer, and Vivienne Sze. 2019. Eyeriss v2: A flexible accelerator for emerging deep neural networks on mobile devices. IEEE Journal on Emerging and Selected Topics in Circuits and Systems9, 2 (2019), 292–308

    2019

  7. [7]

    Zheng Chen, Zichen Zou, Kewei Zhang, Xiongfei Su, Xin Yuan, Yong Guo, and Yulun Zhang. 2025. DOVE: Efficient One-Step Diffusion Model for Real-World Video Super-Resolution. InNeurIPS

    2025

  8. [8]

    Chris McKay. [n. d.]. World Labs’ AI System Can Generate a 3D World from an Image. https://www.maginative.com/article/world-labs-ai-system-can- generate-a-3d-world-from-an-image/

  9. [9]

    Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah

  10. [10]

    Tri Dao. 2023. Flashattention-2: Faster attention with better parallelism and work partitioning.arXiv preprint arXiv:2307.08691(2023)

    Pith/arXiv arXiv 2023

  11. [11]

    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

  12. [12]

    Hangliang Ding, Dacheng Li, Runlong Su, Peiyuan Zhang, Zhijie Deng, Ion Sto- ica, and Hao Zhang. 2025. Efficient-vDiT: Efficient Video Diffusion Transformers With Attention Tile.arXiv preprint arXiv:2502.06155(2025)

    Pith/arXiv arXiv 2025

  13. [13]

    Yu Feng, Zihan Liu, Jingwen Leng, Minyi Guo, and Yuhao Zhu. 2025. Cicero: Addressing Algorithmic and Architectural Bottlenecks in Neural Rendering by Radiance Warping and Memory Optimizations. InProceedings of the 51st Annual International Symposium on Computer Architecture(Buenos Aires, Argentina) (ISCA ’24). IEEE Press, 1293–1308. https://doi.org/10.110...

    arXiv 2025

  14. [14]

    Yu Feng, Paul Whatmough, and Yuhao Zhu. 2019. Asv: Accelerated stereo vision system. InProceedings of the 52nd Annual IEEE/ACM International Symposium on Microarchitecture. 643–656

    2019

  15. [15]

    Ashish Gondimalla, Noah Chesnut, Mithuna Thottethodi, and T. N. Vijaykumar

  16. [16]

    Google. 2024. Veo 2: Our state-of-the-art video generation model. https: //deepmind.google/technologies/veo/veo-2/

    2024

  17. [17]

    Google DeepMind. [n. d.]. Veo3: Our state-of-the-art video generation model. https://deepmind.google/models/veo/ 13 Conference’17, July 2017, Washington, DC, USA Wenxuan Miao et al

    2017

  18. [18]

    Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. 2024. LTX-Video: Realtime Video Latent Diffusion.arXiv preprint arXiv:2501.00103(2024)

    Pith/arXiv arXiv 2024

  19. [19]

    Tae Jun Ham, Sung Jun Jung, Seonghak Kim, Young H Oh, Yeonhong Park, Yoonho Song, Jung-Hun Park, Sanghee Lee, Kyoung Park, Jae W Lee, et al. 2020. Aˆ 3: Accelerating attention mechanisms in neural networks with approximation. In2020 IEEE International Symposium on High Performance Computer Architecture (HPCA). IEEE, 328–341

    2020

  20. [21]

    Jaehoon Heo, Adiwena Putra, Jieon Yoon, Sungwoong Yune, Hangyeol Lee, Ji-Hoon Kim, and Joo-Young Kim. 2025. EXION: Exploiting Inter-and Intra- Iteration Output Sparsity for Diffusion Models. In2025 IEEE International Sym- posium on High Performance Computer Architecture (HPCA). 324–337. https: //doi.org/10.1109/HPCA61900.2025.00034

    arXiv 2025

  21. [22]

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

    2020

  22. [23]

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

    Pith/arXiv arXiv 2022

  23. [24]

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

    2024

  24. [25]

    Team HunyuanWorld. 2025. HunyuanWorld 1.0: Generating Immersive, Ex- plorable, and Interactive 3D Worlds from Words or Pixels.arXiv preprint(2025)

    2025

  25. [26]

    Micron Technology Inc. 2025. Micron System Power Calculators. https://www. micron.com/support/tools-and-utilities/power-calc

    2025

  26. [27]

    Jack Parker-Holder, Shlomi Fruchter. [n. d.]. Genie 3: A new frontier for world models. https://deepmind.google/blog/genie-3-a-new-frontier-for-world- models/

  27. [28]

    Hanhwi Jang, Joonsung Kim, Jae-Eon Jo, Jaewon Lee, and Jangwoo Kim. 2019. MnnFast: a fast and scalable system architecture for memory-augmented neural networks. InProceedings of the 46th International Symposium on Computer Ar- chitecture(Phoenix, Arizona)(ISCA ’19). Association for Computing Machinery, New York, NY, USA, 250–263. https://doi.org/10.1145/3...

    arXiv 2019

  28. [29]

    Jun-Woo Jang, Sehwan Lee, Dongyoung Kim, Hyunsun Park, Ali Shafiee Ardestani, Yeongjae Choi, Channoh Kim, Yoojin Kim, Hyeongseok Yu, Hamzah Abdel-Aziz, Jun-Seok Park, Heonsoo Lee, Dongwoo Lee, Myeong Woo Kim, Hanwoong Jung, Heewoo Nam, Dongguen Lim, Seungwon Lee, Joon-Ho Song, Suknam Kwon, Joseph Hassoun, SukHwan Lim, and Changkyu Choi. 2021. Sparsity-Awa...

    arXiv 2021

  29. [30]

    Huiqiang Jiang, Yucheng Li, Chengruidong Zhang, Qianhui Wu, Xufang Luo, Surin Ahn, Zhenhua Han, Amir Abdi, Dongsheng Li, Chin-Yew Lin, et al. 2024. Minference 1.0: Accelerating pre-filling for long-context llms via dynamic sparse attention.Advances in Neural Information Processing Systems37 (2024), 52481– 52515

    2024

  30. [31]

    Norman P Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder Bajwa, Sarah Bates, Suresh Bhatia, Nan Boden, Al Borchers, et al. 2017. In-datacenter performance analysis of a tensor processing unit. InProceedings of the 44th Annual ACM/IEEE International Symposium on Computer Architecture

    2017

  31. [32]

    Katie Deighton. [n. d.]. Coca-Cola Injects ‘Holidays Are Coming’ Ads With an Upgraded Dose of AI. https://www.wsj.com/articles/coca-cola-injects- holidays-are-coming-ads-with-an-upgraded-dose-of-ai-bc8921e2

  32. [33]

    Sungbin Kim, Hyunwuk Lee, Wonho Cho, Mincheol Park, and Won Woo Ro

  33. [34]

    Weihao Kong, Yifan Hao, Qi Guo, Yongwei Zhao, Xinkai Song, Xiaqing Li, Mo Zou, Zidong Du, Rui Zhang, Chang Liu, et al. 2024. Cambricon-d: Full-network differential acceleration for diffusion models. In2024 ACM/IEEE 51st Annual International Symposium on Computer Architecture (ISCA). IEEE, 903–914

    2024

  34. [35]

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

    Pith/arXiv arXiv 2024

  35. [36]

    Kuaishou. 2024. Kuaishou Unveils Proprietary Video Generation Model ‘Kling’; Testing Now Available. https://ir.kuaishou.com/news-releases/news-release- details/kuaishou-unveils-proprietary-video-generation-model-kling

    2024

  36. [37]

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. InProceedings of the 29th symposium on operating systems principles. 611–626

    2023

  37. [38]

    Jianze Li, Jiezhang Cao, Zichen Zou, Xiongfei Su, Xin Yuan, Yulun Zhang, Yong Guo, and Xiaokang Yang. 2025. Unleashing the Power of One-Step Diffusion based Image Super-Resolution via a Large-Scale Diffusion Discriminator. (2025). arXiv:2410.04224 [cs.CV] https://arxiv.org/abs/2410.04224

    Pith/arXiv arXiv 2025

  38. [39]

    Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, et al. 2024. Open-Sora Plan: Open-Source Large Video Generation Model.arXiv preprint arXiv:2412.00131 (2024)

    Pith/arXiv arXiv 2024

  39. [40]

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le

  40. [41]

    Feng Liu, Shiwei Zhang, Xiaofeng Wang, Yujie Wei, Haonan Qiu, Yuzhong Zhao, Yingya Zhang, Qixiang Ye, and Fang Wan. 2024. Timestep Embedding Tells: It’s Time to Cache for Video Diffusion Model. (2024). arXiv:2411.19108 [cs.CV] https://arxiv.org/abs/2411.19108

    Pith/arXiv arXiv 2024

  41. [42]

    Haosong Liu, Yuge Cheng, Wenxuan Miao, Zihan Liu, Aiyue Chen, Jing Lin, Yiwu Yao, Chen Chen, Jingwen Leng, Yu Feng, and Minyi Guo. 2025. Astraea: A Token-wise Acceleration Framework for Video Diffusion Transformers. (2025). arXiv:2506.05096 [cs.CV] https://arxiv.org/abs/2506.05096

    arXiv 2025

  42. [43]

    Joseph Liu, Joshua Geddes, Ziyu Guo, Haomiao Jiang, and Mahesh Kumar Nandwana. 2024. SmoothCache: A Universal Inference Acceleration Technique for Diffusion Transformers.arXiv preprint arXiv:2411.10510(2024)

    Pith/arXiv arXiv 2024

  43. [44]

    Jinlai Liu, Jian Han, Bin Yan, Hui Wu, Fengda Zhu, Xing Wang, Yi Jiang, Bingyue Peng, and Zehuan Yuan. 2025. InfinityStar: Unified Spacetime AutoRegressive Modeling for Visual Generation. (2025). arXiv:2511.04675 [cs.CV] https://arxiv. org/abs/2511.04675

    arXiv 2025

  44. [45]

    Zhi-Gang Liu, Paul N Whatmough, Yuhao Zhu, and Matthew Mattina. 2022. S2ta: Exploiting structured sparsity for energy-efficient mobile cnn acceleration. In 2022 IEEE International Symposium on High-Performance Computer Architecture (HPCA). IEEE, 573–586

    2022

  45. [47]

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

  46. [48]

    Liqiang Lu, Yicheng Jin, Hangrui Bi, Zizhang Luo, Peng Li, Tao Wang, and Yun Liang. 2021. Sanger: A Co-Design Framework for Enabling Sparse Attention using Reconfigurable Architecture. InMICRO-54: 54th Annual IEEE/ACM In- ternational Symposium on Microarchitecture(Virtual Event, Greece)(MICRO ’21). Association for Computing Machinery, New York, NY, USA, 9...

    arXiv 2021

  47. [49]

    Guoqing Ma, Haoyang Huang, Kun Yan, Liangyu Chen, Nan Duan, Shengming Yin, Changyi Wan, Ranchen Ming, Xiaoniu Song, Xing Chen, Yu Zhou, Deshan Sun, Deyu Zhou, Jian Zhou, Kaijun Tan, Kang An, Mei Chen, Wei Ji, Qiling Wu, Wen Sun, Xin Han, Yanan Wei, Zheng Ge, Aojie Li, Bin Wang, Bizhu Huang, Bo Wang, Brian Li, Changxing Miao, Chen Xu, Chenfei Wu, Chenguang...

    Pith/arXiv arXiv 2025

  48. [50]

    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

  49. [51]

    Mostafa Mahmoud, Kevin Siu, and Andreas Moshovos. 2018. Diffy: a déjà vu- free differential deep neural network accelerator. InProceedings of the 51st An- nual IEEE/ACM International Symposium on Microarchitecture(Fukuoka, Japan) (MICRO-51). IEEE Press, 134–147. https://doi.org/10.1109/MICRO.2018.00020

    arXiv 2018

  50. [52]

    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

  51. [53]

    Micron Technology, Inc. 2023. Micron DDR5 128GB RDIMM. https://assets.micron.com/adobe/assets/urn%3Aaaid%3Aaem%3A6ffd17ac- e709-469d-9473-a0a904681dd9/renditions/original/as/128gb-ddr5-rdimm- product-brief.pdf Product Brief, Rev. A

    2023

  52. [54]

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

    Pith/arXiv arXiv 2022

  53. [55]

    Nvidia. 2021. NVIDIA A100 PCIe 80 GB. https://www.nvidia.com/en-us/data- center/a100/

    2021

  54. [56]

    Nvidia. 2022. NVIDIA H100 PCIe 80 GB. https://www.nvidia.com/en-us/data- center/h100/

    2022

  55. [57]

    OpenAI. 2024. Sora: Bring your imagination to life with text, image, or video. https://openai.com/sora/

    2024

  56. [58]

    Subhankar Pal, Jonathan Beaumont, Dong-Hyeon Park, Aporva Amarnath, Siying Feng, Chaitali Chakrabarti, Hun-Seok Kim, David Blaauw, Trevor Mudge, and Ronald Dreslinski. 2018. OuterSPACE: An Outer Product Based Sparse Matrix Multiplication Accelerator. In2018 IEEE International Symposium on High Performance Computer Architecture (HPCA). 724–736. https://doi...

    arXiv 2018

  57. [59]

    Wenxuan Miao, Yulin Sun, Aiyue Chen, Jing Lin, Yiwu Yao, Yiming Gan, Jieru Zhao, Jingwen Leng, Minyi Guo, and Yu Feng. 2026. TimeRipples: Accelerating vDiTs by Understanding the Spatio-Temporal Correlations in Latent Space. In 14 Kaleido: Algorithm-Hardware Co-Design for Video Diffusion Transformers Conference’17, July 2017, Washington, DC, USA Proceeding...

    2026

  58. [60]

    Lenny Rachitsky. 2025. The Godmother of AI on jobs, robots and why world models are next. https://www.lennysnewsletter.com/p/the-godmother-of-ai

    2025

  59. [61]

    Nvidia. [n. d.]. NVIDIA RTX PRO 6000 Blackwell. https://www.nvidia.com/en- us/products/workstations/professional-desktop-gpus/rtx-pro-6000/

  60. [62]

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

    Pith/arXiv arXiv 2022

  61. [63]

    Team Seawead, Ceyuan Yang, Zhijie Lin, Yang Zhao, Shanchuan Lin, Zhibei Ma, Haoyuan Guo, Hao Chen, Lu Qi, Sen Wang, et al . 2025. Seaweed-7B: Cost-Effective Training of Video Generation Foundation Model.arXiv preprint arXiv:2504.08685(2025)

    Pith/arXiv arXiv 2025

  62. [64]

    Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al . 2022. Make- a-video: Text-to-video generation without text-video data.arXiv preprint arXiv:2209.14792(2022)

    Pith/arXiv arXiv 2022

  63. [65]

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

    Pith/arXiv arXiv 2020

  64. [67]

    Zhuoran Song, Feiyang Wu, Xueyuan Liu, Jing Ke, Naifeng Jing, and Xiaoyao Liang. 2020. Vr-dann: Real-time video recognition via decoder-assisted neural network acceleration. In2020 53rd Annual IEEE/ACM International Symposium on Microarchitecture (MICRO). IEEE, 698–710

    2020

  65. [68]

    Runway. 2024. Runway Gen 2: The Next Step Forward for Generative AI. https://medium.com/@dbhatt245/runway-gen-2-the-next-step-forward- for-generative-ai-an-introduction-b85bc90d3e45

    2024

  66. [69]

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

    2024

  67. [70]

    Meituan LongCat Team, Xunliang Cai, Qilong Huang, Zhuoliang Kang, Hongyu Li, Shijun Liang, Liya Ma, Siyu Ren, Xiaoming Wei, Rixu Xie, and Tong Zhang

  68. [71]

    Robbyant Team, Zelin Gao, Qiuyu Wang, Yanhong Zeng, Jiapeng Zhu, Ka Leong Cheng, Yixuan Li, Hanlin Wang, Yinghao Xu, Shuailei Ma, Yihang Chen, Jie Liu, Yansong Cheng, Yao Yao, Jiayi Zhu, Yihao Meng, Kecheng Zheng, Qingyan Bai, Jingye Chen, Zehong Shen, Yue Yu, Xing Zhu, Yujun Shen, and Hao Ouyang

  69. [72]

    Tencent. 2024. Tencent launches and open-sources Hunyuan video-generation model. https://technode.com/2024/12/04/tencent-launches-and-open-sources- hunyuan-video-generation-model/

    2024

  70. [73]

    Zhuoran Song, Chunyu Qi, Fangxin Liu, Naifeng Jing, and Xiaoyao Liang. 2024. CMC: Video Transformer Acceleration via CODEC Assisted Matrix Condensing. InProceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2(La Jolla, CA, USA)(ASPLOS ’24). Association for Computing Machinery,...

    arXiv 2024

  71. [74]

    Hanrui Wang, Zhekai Zhang, and Song Han. 2021. SpAtten: Efficient Sparse Attention Architecture with Cascade Token and Head Pruning. In2021 IEEE International Symposium on High-Performance Computer Architecture (HPCA). 97–110. https://doi.org/10.1109/HPCA51647.2021.00018

    arXiv 2021

  72. [75]

    Nitish Srivastava, Hanchen Jin, Jie Liu, David Albonesi, and Zhiru Zhang. 2020. MatRaptor: A Sparse-Sparse Matrix Multiplication Accelerator Based on Row- Wise Product. In2020 53rd Annual IEEE/ACM International Symposium on Mi- croarchitecture (MICRO). 766–780. https://doi.org/10.1109/MICRO50266.2020. 00068

    arXiv 2020

  73. [76]

    Yannan Nellie Wu, Po-An Tsai, Saurav Muralidharan, Angshuman Parashar, Vivienne Sze, and Joel Emer. 2023. HighLight: Efficient and Flexible DNN Acceleration with Hierarchical Structured Sparsity. InProceedings of the 56th Annual IEEE/ACM International Symposium on Microarchitecture (MICRO ’23). Association for Computing Machinery, New York, NY, USA, 1106–...

    arXiv 2023

  74. [77]

    Haocheng Xi, Shuo Yang, Yilong Zhao, Chenfeng Xu, Muyang Li, Xiuyu Li, Yujun Lin, Han Cai, Jintao Zhang, Dacheng Li, et al. 2025. Sparse VideoGen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity. arXiv preprint arXiv:2502.01776(2025)

    Pith/arXiv arXiv 2025

  75. [78]

    LongCat-Video Technical Report. (2025). arXiv:2510.22200 [cs.CV] https://arxiv.org/abs/2510.22200

    arXiv 2025

  76. [79]

    Jiaqi Xu, Xinyi Zou, Kunzhe Huang, Yunkuo Chen, Bo Liu, MengLi Cheng, Xing Shi, and Jun Huang. 2024. Easyanimate: A high-performance long video genera- tion method based on transformer architecture.arXiv preprint arXiv:2405.18991 (2024)

    arXiv 2024

  77. [80]

    Lu Xu, Sijie Zhu, Chunyuan Li, Chia-Wen Kuo, Fan Chen, Xinyao Wang, Guang Chen, Dawei Du, Ye Yuan, and Longyin Wen. 2025. Beyond raw videos: Un- derstanding edited videos with large multimodal model. InProceedings of the Computer Vision and Pattern Recognition Conference. 503–512

    2025

  78. [81]

    Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. 2023. Diffusion models: A comprehensive survey of methods and applications.Comput. Surveys56, 4 (2023), 1–39

    2023

  79. [82]

    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, Jin- gren 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 ...

  80. [83]

    Wan: Open and Advanced Large-Scale Video Generative Models.arXiv preprint arXiv:2503.20314(2025)

    Pith/arXiv arXiv 2025

Showing first 80 references.