Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

UniCP: A Unified Caching and Pruning Framework for Efficient Video Generation

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read UniCP claims a unified, error-aware framework that combines attention caching and PCA-based pruning can speed up diffusion-transformer video generation by up to 1.6× on a single GPU with little quality loss.

desk verdict Plausible combination of known caching and pruning tricks, but the paper's own speed-matched comparisons show FasterCache beating UniCP, and the PCA 'optimality' claim does not survive contact with the algebra. read the letter →

arxiv 2502.04393 v1 pith:BKAH3ZTL submitted 2025-02-06 cs.CV

Wenzhang Sun , Qirui Hou , Donglin Di , Jiahui Yang , Yongjia Ma , Jianxun Cui This is my paper · ORCID
classification cs.CV
keywords diffusiontransformersvideogenerationattentioncachingmodelpruningPCA-basedcompressionerror-awareschedulinginferenceaccelerationcachewindow
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

UniCP claims that the standard tricks for speeding up diffusion-transformer video generation—caching attention computations across denoising steps and pruning redundant attention dimensions—work best when combined, with the choice made per block and per timestep from the measured error. The paper argues that attention differences between adjacent steps form a U-shaped curve with sudden spikes, so fixed cache intervals either waste computation or let errors accumulate. Its three mechanisms, an error-aware dynamic cache window, PCA-based slicing, and a dynamic weight shift, let each attention block reuse an old attention output, reuse an old attention map, or compute on pruned query and key dimensions depending on an error threshold δ. If correct, this yields up to a 1.6× single-GPU speedup on current video-generation DiT models with only small quality loss, and it extends acceleration into the high-error regions where caching alone fails.

What carries the argument

The load-bearing machinery is a per-block decision rule that uses local attention error as a proxy for whether computation can be skipped. For a block $i$ at timestep $t$, the error is $\|o_i - o_{i+k}\|_2$ between attention outputs or $\|a_i - a_{i+k}\|_2$ between attention maps for candidate cache steps $k$; if either falls below a threshold $\delta$, the block reuses that previous value. When neither passes, the block falls back to a PCA-based Slicing that compresses the query and key projections via $Q = Z_q D^\top R^\top$ and $K = Z_k D^\top R^\top$, where $R$ holds eigenvectors of $X^\top X$ and $D$ selects a thin set of columns, so the attention map is computed in lower dimension. A Dynamic Weight Shift then coordinates the two modes by storing both pre-pruning and post-pruning weights and switching on a spatial-temporal cache map, preventing the two strategies from interfering.

What would settle it

A concrete test: run the method on a video-generation model with a fixed δ and compare generated videos against the original while corrupting the error signal—for instance, by randomizing the cache decisions while keeping the same compute budget. If random caching with the same average skip rate matches UniCP's quality, then the error-threshold signal is not doing the work; if quality collapses, the error signal is load-bearing. Alternatively, compute the correlation between the block-level L2 attention differences and downstream frame-level FID/CLIP score over many prompts; a near-zero correlation would falsify the premise.

Watch

Extended reading notes

Core claim

The paper's central claim is that the U-shaped attention-error profile across denoising steps is not smooth enough for fixed caching schedules: sudden error spikes appear even at the bottom of the curve, and the two ends show large discrepancies that make caching harmful. The proposed solution treats error as the control signal: each attention block measures the L2 distance between its current and previous attention output, or between attention maps, compares it with a per-step threshold δ, and then chooses among caching the attention output, caching the attention map, or pruning the query and key projections through PCA. The reported experiments on three video-generation DiT models show that this error-aware routing keeps VBench scores and similarity metrics nearly flat while cutting latency and MACs, and it outperforms caching-only baselines at similar or better efficiency.

Load-bearing premise

The load-bearing premise is that a local, instantaneous L2 difference between attention outputs or attention maps at adjacent steps can be thresholded with a single scalar δ to decide when to skip computation, without tracking accumulated quality loss; if local attention similarity does not track global video fidelity, the whole scheme loses its error control.

Editorial extensions

If this is right

  • On the three video-generation models tested, UniCP reaches up to a 1.6× single-GPU speedup with quality metrics close to the original model.
  • The error threshold δ becomes a continuous speed–quality dial: raising it from 0.025 to 0.175 lowers latency substantially while quality degrades only slightly.
  • Blocks in the high-error regions at the ends of the denoising trajectory, which caching-only methods cannot accelerate, are handled instead by pruning, so the framework covers the whole trajectory.
  • Ablation results indicate that dynamically choosing between caching the attention output and caching the attention map is better than committing to either one alone.

Reading between the lines

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

  • A per-block or per-timestep learned threshold would likely beat the single global δ used here, since the paper's own figures show error profiles differ across blocks; this is a natural next step the authors do not take.
  • Because the local L2 criterion ignores accumulated error, adding a running estimate of total introduced error could tighten the quality guarantee and allow larger speedups at the same visual fidelity.
  • The same cache-or-prune routing should transfer to image-generation DiTs and to cross-attention layers, but the experiments only cover spatial and temporal attention in video models.
  • If the measured 1.6× speedup persists at longer videos and higher resolutions, the practical cost of high-quality video generation drops substantially; that scaling is not demonstrated in the paper.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. UniCP proposes a training-free acceleration framework for DiT-based video generation that combines an error-aware dynamic cache window (EDCW), PCA-based slicing of attention queries and keys (PCAS), and a dynamic weight shift (DWS) to switch between cached and pruned computation. The method is evaluated on OpenSora 1.2, Latte, and CogVideoX against PAB and FasterCache using VBench, LPIPS, SSIM, PSNR, latency, and MACs. The central claim is that UniCP delivers up to a 1.6x speedup without compromising video quality and outperforms existing methods in both performance and efficiency.

Significance. The idea of unifying caching and pruning with adaptive, error-driven control is a plausible extension of the existing training-free acceleration literature for video diffusion. The paper includes a threshold sweep (E1–E5) and ablations of the caching strategies, which are useful elements. However, the manuscript provides no code, no error bars, and its own Table I contradicts the headline comparative claim. If the empirical results held, the contribution would be moderate; as presented, the evidence does not establish that UniCP improves on the FasterCache baseline at matched speedups. The paper's significance is therefore substantially below the bar for a top computer vision venue.

major comments (4)
  1. [Section IV-C, Table I (and Abstract)] At matched speedups, FasterCache outperforms UniCP on all four reported quality metrics on Latte (1.59x vs 1.61x: VBench 76.92 vs 76.82, LPIPS 0.0952 vs 0.0978, SSIM 0.8548 vs 0.8471, PSNR 23.75 vs 23.65) and CogVideoX (1.60x vs 1.64x: VBench 78.34 vs 77.72, LPIPS 0.0850 vs 0.0866, SSIM 0.9572 vs 0.9237, PSNR 28.66 vs 28.37). On OpenSora, the comparison is mixed: UniCP E5 has slightly better LPIPS and SSIM but lower VBench and PSNR. No configuration shows UniCP simultaneously faster and better than FasterCache. Moreover, for each model, VBench declines monotonically from E1 to E5 relative to the original model (e.g., OpenSora from 78.17 to 77.34 vs original 78.69). This directly undercuts the abstract's 'without compromising video quality' and 'outperforms existing methods' claims.
  2. [Algorithm 1 and Section III-C] The DWS algorithm as written is ambiguous: after the first 'for' loop sets c=T and breaks, control proceeds to the second 'for' loop and then to 'Apply PCA-based Slicing,' so the supposed dynamic switching between caching and pruning is not actually exclusive. If the intent is that output caching, map caching, and slicing are mutually exclusive alternatives, the pseudocode must include explicit 'else if' structure and a check that terminates the routine once c=T. As written, the method's operation is underspecified.
  3. [Section III-B, Eqs. (2)-(3)] The PCA formulation is not self-consistent as stated. Substituting Zq and Zk gives Q = XWq R D D^T R^T and K = XWk R D D^T R^T; because D D^T is a projection rather than the identity, the resulting Q and K are approximations, not the original matrices, and the paper never states this or quantifies the approximation error. More importantly, the paper does not explain how the eigenvector matrix R is obtained during inference. If R is computed online per block and timestep, the eigendecomposition cost O(m^3) would likely dominate the attention savings, contradicting the MACs numbers in Table I. A per-component complexity analysis is needed.
  4. [Section IV-C, Table II] The ablation in Table II does not support the claimed advantage of the dynamic selection strategy: the 'dynamic select' row has lower VBench (78.17) than both 'attention output' (78.23) and 'attention map' (78.26), with only marginal latency differences (49.07 vs 49.19/49.27 s). Without repeated trials or confidence intervals, these differences are not meaningful, so the DWS contribution is not validated either by the main table or by the ablation.
minor comments (6)
  1. [Section V] The sentence 'We presents UniCP' should be 'We present UniCP.'
  2. [Table I] The CogVideoX PAB PSNR entry reads '28,51' (comma) instead of '28.51.'
  3. [Section III-B] The notation for dimensions is inconsistent: X is introduced as R^{m×m}, but attention inputs are typically m×d; the dimensions of X, Wq, Wk, and the head dimension should be clarified.
  4. [Section IV-A] The threshold labels E1–E5 are introduced as δ = 0.025, 0.05, 0.75, 0.125, 0.175; E3's 0.75 appears to be a typo for 0.075 given the monotonic progression.
  5. [Algorithm 1] The indentation and capitalization of 'Break' should be fixed for readability, and the pseudocode should be formatted consistently with standard algorithm environments.
  6. [Section III-A] Figures 3 and 4 are not directly referenced in the text at their first occurrence; add explicit references to guide the reader.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: UniCP's thresholds are explicit user parameters and its quality/efficiency claims are tested against external baselines and metrics.

full rationale

The paper's derivation chain is algorithmic rather than predictive: EDCW (Eq. 1) defines the cache step as a function of the user-supplied error threshold delta and measured attention output/map differences, so nothing is being 'predicted' from a fitted constant. The thresholds E1-E5 are enumerated in Section IV-C as explicit settings, not hidden parameters tuned to match the reported VBench, LPIPS, SSIM, or PSNR values. The U-shaped error observation is attributed to the external prior work PAB [9], not to a self-citation, and FasterCache [24] is also an external baseline. PCAS invokes the standard L2-optimality property of PCA, an external mathematical result, and DWS is a switching mechanism between cached and pruned outputs. The paper's central claims are therefore not equivalent to their inputs by construction. A skeptic might dispute whether the evidence in Table I supports the 'outperforms existing methods' headline once speed is matched, but that is a soundness/correctness concern, not circularity, and no equation in the paper reduces to its own output. No self-citation chain, uniqueness theorem, or renamed fitted parameter is load-bearing. Accordingly, the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The framework relies on an error threshold as a free knob, on the U-shaped attention-difference assumption from PAB, and on the validity of applying PCA to attention inputs. No new physical or mathematical entities are introduced.

free parameters (3)
  • error threshold delta = 0.025, 0.05, 0.075 (printed as 0.75), 0.125, 0.175 (E1-E5)
    Controls when caching or pruning is triggered; no principled selection rule is given, and the reported speed/quality trade-off depends directly on it.
  • search window size K = not specified
    Used in Algorithm 1 to bound how far back caching looks; no value or sensitivity analysis is provided.
  • PCA slice ratio = 0.1 to 0.4 (dynamic)
    Portion of query/key dimensions removed; chosen from Fig. 8 rather than derived. DWS selects per-block k below a threshold curve.
assumptions (4)
  • standard math Eckart-Young theorem: top-k eigenvector projection is L2-optimal for reconstructing X.
    Section III-B claims PCAS reconstruction is L2-optimal, invoking this classical result, but Eq. 2-3 replace the query/key product with a projected version whose optimality is not established.
  • domain assumption Adjacent-step attention differences follow a U-shaped pattern with occasional spikes.
    Motivated by PAB [9] and Fig. 2; EDCW's design and the distinction between cacheable middle steps and uncacheable ends depend on it.
  • domain assumption Per-step thresholded differences between attention outputs and maps predict final video quality.
    Algorithm 1 makes caching and pruning decisions from ||oi - oi+k|| and ||ai - ai+k|| alone; no accumulation or downstream fidelity model is used.
  • ad hoc to paper Switching between cached outputs, pruned weights, and original weights per block and timestep does not introduce artifacts.
    Section III-C asserts DWS unifies the two strategies without analyzing boundary effects or discontinuities across steps.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UniCP: A Unified Caching and Pruning Framework for Efficient Video Generation." pith.science (2026). https://pith.science/paper/BKAH3ZTL

@misc{pith2026250204393,
  author       = {Pith},
  title        = {Pith review of: UniCP: A Unified Caching and Pruning Framework for Efficient Video Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BKAH3ZTL}},
  note         = {Machine review of arXiv:2502.04393}
}
read the original abstract

Diffusion Transformers (DiT) excel in video generation but encounter significant computational challenges due to the quadratic complexity of attention. Notably, attention differences between adjacent diffusion steps follow a U-shaped pattern. Current methods leverage this property by caching attention blocks, however, they still struggle with sudden error spikes and large discrepancies. To address these issues, we propose UniCP a unified caching and pruning framework for efficient video generation. UniCP optimizes both temporal and spatial dimensions through. Error Aware Dynamic Cache Window (EDCW): Dynamically adjusts cache window sizes for different blocks at various timesteps, adapting to abrupt error changes. PCA based Slicing (PCAS) and Dynamic Weight Shift (DWS): PCAS prunes redundant attention components, and DWS integrates caching and pruning by enabling dynamic switching between pruned and cached outputs. By adjusting cache windows and pruning redundant components, UniCP enhances computational efficiency and maintains video detail fidelity. Experimental results show that UniCP outperforms existing methods in both performance and efficiency.

Figures

Figures reproduced from arXiv: 2502.04393 by the authors.

Figure 1
Figure 1. Accelerating video generation methods like [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Visualization of attention differences in Open-Sora. (a) Conventional U-shaped error distribution and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 5
Figure 5. After acquiring the spatial-temporal cache [PITH_FULL_IMAGE:figures/full_fig_p003_5.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Visualization of the PCAS. PCAS reduces the computational cost of the attention mechanism by pruning redundant dimensions in the query and key matrices. approach is then applied to the attention map (ai). The first strategy offers greater computational savings but allo…
Figure 6
Figure 6. Figure 6: Video generation samples under various error [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]
Figure 7
Figure 7. Figure 7: Visualization of generated video quality, la [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 8
Figure 8. Figure 8: Visual results across various slice ratios. [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. DiffSparse: Accelerating Diffusion Transformers with Learned Token Sparsity

    cs.CV 2026-04 conditional novelty 6.0 of 10

    A learnable cost predictor plus dynamic programming allocates layer-wise token sparsity for diffusion transformers, removing forced full steps and cutting ~54% compute on PixArt-α without quality loss.

  2. Predictive Feature Caching for Training-free Acceleration of Molecular Geometry Generation

    cs.LG 2025-10 conditional novelty 4.0 of 10

    Predictive feature caching, borrowed from image diffusion, speeds up molecular flow-matching generation by 2-3x at near-matched quality by forecasting hidden features instead of recomputing them.

Reference graph

Works this paper leans on

28 extracted references · 14 canonical work pages · cited by 2 Pith papers

  1. [1]

    Scalable diffusion models with transformers,

    William Peebles and Saining Xie, “Scalable diffusion models with transformers,” arXiv preprint arXiv:2212.09748 , 2022

    arXiv 2022

  2. [2]

    Ditfastattn: Attention compression for diffusion transformer models,

    Zhihang Yuan, Pu Lu, Hanling Zhang, Xuefei Ning, Lin- feng Zhang, Tianchen Zhao, Shengen Yan, Guohao Dai, and Yu Wang, “Ditfastattn: Attention compression for diffusion transformer models,” 2024

    work page 2024

  3. [3]

    Denoising dif- fusion probabilistic models,

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

    work page 2020

  4. [4]

    High-resolution image synthesis with latent diffusion models,

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer, “High-resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 10684–10695

    work page 2022

  5. [5]

    Stable video diffusion: Scaling latent video diffusion models to large datasets,

    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

    arXiv 2023

  6. [6]

    Denoising diffusion implicit models,

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

    arXiv 2010

  7. [7]

    Deepcache: Accelerating diffusion models for free,

    Xinyin Ma, Gongfan Fang, and Xinchao Wang, “Deepcache: Accelerating diffusion models for free,” in The IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024

    work page 2024

  8. [8]

    Faster diffusion: Rethinking the role of unet encoder in diffusion models,

    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 diffusion models,” arXiv e-prints, pp. arXiv–2312, 2023

    work page 2023

Show all 28 references
  1. [9]

    Real- time video generation with pyramid attention broadcast,

    Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You, “Real- time video generation with pyramid attention broadcast,” 2024

    2024

  2. [10]

    Delta-dit: A training-free acceleration method tailored for diffusion transformers,

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

    2024 arXiv

  3. [11]

    Temporal generative adversarial nets with singular value clipping,

    Masaki Saito, Eiichi Matsumoto, and Shunta Saito, “Temporal generative adversarial nets with singular value clipping,” in Proceedings of the IEEE international conference on computer vision, 2017, pp. 2830–2839

    2017

  4. [12]

    Video-to- video synthesis,

    Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro, “Video-to- video synthesis,” arXiv preprint arXiv:1808.06601 , 2018

    2018 arXiv

  5. [13]

    Scaling autoregressive video models,

    Dirk Weissenborn, Oscar Täckström, and Jakob Uszkor- eit, “Scaling autoregressive video models,” arXiv preprint arXiv:1906.02634, 2019

    1906 arXiv

  6. [14]

    Video diffusion models,

    Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet, “Video diffusion models,” Advances in Neural Information Processing Systems, vol. 35, pp. 8633–8646, 2022

    2022

  7. [15]

    Latte: Latent diffusion transformer for video generation,

    Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao, “Latte: Latent diffusion transformer for video generation,” arXiv preprint arXiv:2401.03048, 2024

    2024 arXiv

  8. [16]

    Cogvideo: Large-scale pretraining for text-to-video generation via transformers,

    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

    2022 arXiv

  9. [17]

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

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

    2024 arXiv

  10. [18]

    Vasa-1: Lifelike audio-driven talking faces generated in real time,

    Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo, “Vasa-1: Lifelike audio-driven talking faces generated in real time,” arXiv preprint arXiv:2404.10667 , 2024

    2024 arXiv

  11. [19]

    Scaling rectified flow transformers for high-resolution image synthesis,

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim En- tezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al., “Scaling rectified flow transformers for high-resolution image synthesis,” in Forty- first International Conference on Ma...

    2024

  12. [20]

    Hunyuanvideo: A systematic framework for large video generative models,

    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

    2024 arXiv

  13. [21]

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

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

    2022

  14. [22]

    One- step diffusion with distribution matching distillation,

    Tianwei Yin, Michaël Gharbi, Richard Zhang, Eli Shechtman, Frédo Durand, William T Freeman, and Taesung Park, “One- step diffusion with distribution matching distillation,” in CVPR, 2024

    2024

  15. [23]

    Slicegpt: Compress large language models by deleting rows and columns,

    Saleh Ashkboos, Maximilian L Croci, Marcelo Gennari do Nascimento, Torsten Hoefler, and James Hensman, “Slicegpt: Compress large language models by deleting rows and columns,” arXiv preprint arXiv:2401.15024 , 2024

    2024 arXiv

  16. [24]

    Fastercache: Training-free video diffusion model acceleration with high quality,

    Zhengyao Lv, Chenyang Si, Junhao Song, Zhenyu Yang, Yu Qiao, Ziwei Liu, and Kwan-Yee K Wong, “Fastercache: Training-free video diffusion model acceleration with high quality,” arXiv preprint arXiv:2410.19355 , 2024

    2024 arXiv

  17. [25]

    Vbench: Comprehensive benchmark 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 benchmark suite for video generative models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and ...

    2024

  18. [26]

    The unreasonable effectiveness of deep features as a perceptual metric,

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 586–595

    2018

  19. [27]

    Image quality assessment: from error visibility to structural similarity,

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE transactions on image processing , vol. 13, no. 4, pp. 600–612, 2004

    2004

  20. [28]

    Peak signal-to-noise ratio revisited: Is simple beautiful?,

    Jari Korhonen and Junyong You, “Peak signal-to-noise ratio revisited: Is simple beautiful?,” in 2012 Fourth international workshop on quality of multimedia experience . IEEE, 2012, pp. 37–38

    2012

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.