Pith. sign in

REVIEW 3 major objections 6 minor 80 references

HyperVAttention makes training-free sparse attention for video diffusion both faster and more faithful by cutting clustering cost and packing irregular clusters into GPU tiles.

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 · grok-4.5

2026-07-12 05:27 UTC pith:LMFIHSXG

load-bearing objection Real systems win on video DiT sparse attention: local/hybrid clustering plus CTA-aware merging moves the measured Pareto; locality assumptions are load-bearing and under-ablated, not fatal. the 3 major comments →

arxiv 2607.03012 v1 pith:LMFIHSXG submitted 2026-07-03 cs.CV cs.AIcs.LG

HyperVAttention: Efficient Sparse Attention with Spatio-Temporal Clustering for Video Diffusion

Dongyeun Lee , Amir Zandieh , Vahab Mirrokni , Junmo Kim , Insu Han This is my paper
classification cs.CV cs.AIcs.LG
keywords video diffusion transformerssparse attentionspatio-temporal clusteringtraining-free accelerationCTA utilizationhardware-aware mergingtext-to-video generation
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.

Long video generation with diffusion transformers is bottlenecked by quadratic self-attention over huge spatio-temporal token sequences. Clustering-based sparse attention already groups similar tokens so kernels can skip work, but in practice two frictions remain: clustering itself is expensive, and the resulting blocks waste GPU thread capacity because their sizes do not match fixed tile shapes. This paper claims both problems can be solved without any retraining. It restricts centroid search to local 3D neighborhoods that match video structure, refreshes only the unstable tokens between denoising steps, and merges clusters specifically to minimize the number of padded CTA tiles. On large text-to-video models the method reports up to about 2.13× end-to-end speedup while matching or beating prior sparse baselines on fidelity to full attention. A sympathetic reader cares because the same recipe turns theoretical sparsity into wall-clock gains on production-scale video generators without changing the trained weights.

Core claim

HyperVAttention establishes a new quality–speed frontier among training-free sparse attention methods for video diffusion transformers: by jointly reducing clustering overhead and improving CTA utilization, it cuts end-to-end latency by up to 2.13× (2.23× in a faster variant) on HunyuanVideo text-to-video while improving PSNR/LPIPS fidelity over existing training-free sparse baselines.

What carries the argument

Three linked pieces: 3D local-window clustering (centroid search restricted to fixed spatio-temporal neighborhoods, run in a custom Triton kernel), hybrid clustering across denoising steps (full clustering only at anchors; intermediate steps update only a Lowe’s-ratio subset of ambiguous tokens), and hardware-aware cluster merging (parallel agglomerative merges that minimize CTA-aligned tiled cost and fill idle tile capacity).

Load-bearing premise

Video latent tokens stay coherent enough in local space–time, and their cluster labels stay stable enough across neighboring denoising steps, that local windows plus partial updates still give an attention mask close to full attention.

What would settle it

Measure end-to-end latency and full-attention reference metrics (PSNR/LPIPS) on the same long 720p clips under a controlled ablation that forces global reclustering every step and disables cluster merging; if speedup collapses toward baseline clustering methods while fidelity does not improve, or if high-motion/scene-cut videos show large quality drops relative to full attention under the default local-window hybrid schedule, the central joint claim fails.

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

If this is right

  • Longer token sequences (more frames or higher resolution) become practical without retraining the diffusion backbone, because clustering cost scales with a fixed local window rather than the full centroid set.
  • Theoretical token sparsity can translate into real GPU speedups once cluster blocks are aligned to CTA tiles, so density numbers alone are no longer the right optimization target.
  • Partial Lowe’s-ratio updates can amortize expensive clustering across many denoising steps whenever assignments are temporally stable.
  • Idle tile capacity can be used to admit slightly denser, more accurate attention at little or no extra hardware cost after merging.
  • The same training-free stack can be stacked with orthogonal accelerations such as caching, quantization, or distillation for further gains on long-video generation.

Where Pith is reading between the lines

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

  • The same local-window plus hybrid-update idea may transfer to other 3D or multi-frame transformers (video understanding, 4D generation) where tokens are spatio-temporally coherent.
  • Hardware-cost-aware merging is a general post-process for any clustering-based sparse attention whose blocks misalign with fixed kernel tiles, not only video diffusion.
  • If assignment stability weakens late in denoising or on highly dynamic content, adaptive anchor spacing (more frequent full clustering when Lowe’s ratios rise) would be a natural extension.
  • Gains will shrink on short clips or image-only DiTs where attention is no longer the dominant runtime share, matching the paper’s own limitation note.

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 / 6 minor

Summary. The paper proposes HyperVAttention (HVA), a training-free sparse attention method for Video Diffusion Transformers that targets two practical bottlenecks of clustering-based sparse attention: (i) clustering overhead and (ii) poor CTA/tile utilization from irregular cluster-induced blocks. It introduces 3D local-window clustering with a custom Triton kernel, a hybrid full/subset reclustering schedule across denoising steps guided by Lowe’s ratio, and hardware-aware parallel agglomerative cluster merging that minimizes a CTA-aligned cost C(S). On Wan2.2 (I2V/T2V) and HunyuanVideo (T2V) at 720p, HVA reports up to 2.13× end-to-end speedup (2.23× for HVA-Flash) versus full attention while improving PSNR/SSIM/LPIPS and LLM-judge scores over SpargeAttn, SVG, and SVG2, claiming a new quality–latency Pareto frontier for training-free sparse video attention.

Significance. If the empirical Pareto claim holds under fair controls, this is a solid systems contribution to long-video diffusion inference: it jointly attacks algorithmic clustering cost and GPU tile underutilization, ships a practical Triton kernel, and demonstrates multi-model wall-clock gains with better fidelity than strong training-free baselines. The hardware cost formulation C(i)/C(S) and parallel merge heuristic are concrete and useful for other cluster-based sparse attention pipelines. Strengths include head-to-head latency/fidelity tables, latency breakdowns, clustering-error-vs-latency curves, and theoretical-vs-CTA density comparisons. The work is complementary to quantization, caching, and distillation and is most relevant for long sequences where attention dominates runtime.

major comments (3)
  1. [§3–4.2, Fig. 3, Fig. 5b, Table 1] The fidelity half of the Pareto claim rests on the premise that 3D local-window clustering (Eq. 5; fixed windows such as (3,8,8)) plus hybrid anchor/subset updates (full clustering every two steps; top ~15% tokens by Lowe’s ratio ρ_ℓ in Eq. 6) preserves or improves attention quality relative to global k-means (SVG2). Section 3 and Fig. 3 only provide qualitative PCA maps, assignment-change rates, and Lowe-ratio distributions; Fig. 5b reports clustering error (a proxy) vs latency, not end-to-end PSNR/LPIPS, mask agreement, VBench, or LLM-judge under local-window vs global clustering with matched Top-p/Min-kc and density. Without that ablation, it remains unclear whether reported gains (e.g., Table 1 HunyuanVideo PSNR 33.97 vs SVG2 32.66) come from better semantic grouping or from other knobs (density after merging, Min-kc).
  2. [Table 1, Table 2, Fig. 5c, §5.1] Table 2 shows systematically different Min-kc (and resulting densities) for HVA vs SVG2 (e.g., HunyuanVideo Min-kc 0.15 vs 0.1; theoretical density 31.34% vs 26.27% in Table 1). Fig. 5c further shows merging raises theoretical density (39.70%→42.24%) while lowering CTA density. The paper attributes superior fidelity and wall-clock speed jointly to local clustering and merging, but does not report a controlled comparison with matched retained-key budget / matched theoretical density. Please add matched-budget runs (same Top-p and Min-kc or same token density) so quality and speedup can be attributed fairly rather than to free retention parameters.
  3. [§4.3, Fig. 5, §5.2] Component contributions to fidelity are not isolated end-to-end. Fig. 5a breaks down latency (clustering 83s→26s; merge +4s) and Fig. 5c shows density effects of merging, but there is no quality table for: (a) local-window only, (b) +hybrid subset updates, (c) +hardware-aware merging, each versus full attention and SVG2. Because merging unions key selections and can add tokens into idle tiles, it can improve approximation independently of clustering quality. A short factorial ablation on one model/setting is needed to support the joint-design narrative in the abstract and §7.
minor comments (6)
  1. [Abstract, §1, Table 1] Notation for the method name is inconsistent (HyperV Attention, HV A, HVA, Ours). Standardize to one acronym throughout abstract, body, tables, and figures.
  2. [Abstract, §1, Table 1] Abstract claims up to 2.13× while §1 and Table 1 also report 2.23× for HVA-Flash and ~1.8× on Wan2.2; align headline numbers with the main configuration vs Flash variant.
  3. [Fig. 3a] Fig. 3a cluster visualizations are hard to interpret without a clearer description of PCA→RGB mapping and whether colors are comparable across layers/heads; a short caption expansion would help.
  4. [Algorithm 1, §4.3, A.1] Algorithm 1’s bipartition merge is approximate; a brief note on sensitivity to R (fixed at 3) and random bipartitions (variance over seeds) would strengthen reproducibility claims.
  5. [§5, Appendix B.3] LLM-as-a-judge uses gemini-3-flash preview; report judge variance or multi-seed scoring if available, and fix minor prompt typos in B.3 ("Aesthestic", "MSUT").
  6. [§7] Limitation that gains are mainly for long videos is appropriate; consider one short quantitative note on shorter sequences or image DiT to bound the regime.

Circularity Check

0 steps flagged

No significant circularity: empirical systems paper with measured speedups and fidelity against external full-attention baselines, not definitional or fitted-as-prediction claims.

full rationale

HyperVAttention is a training-free systems paper whose central claims (end-to-end latency reduction up to 2.13× / 2.23× and improved PSNR/LPIPS/LLM-Judge vs SpargeAttn, SVG, SVG2) are established by wall-clock measurements and video-level metrics against a full-attention reference and external baselines (Table 1, Fig. 5a), not by redefining the target quantity. The algorithmic chain is constructive engineering, not a closed derivation: (i) 3D local-window clustering (Eq. 5) and hybrid Lowe’s-ratio subset updates (Eq. 6) are motivated by empirical observations (Fig. 3) and evaluated via clustering-error latency (Fig. 5b) and end-to-end quality; (ii) hardware cost C(i)/C(S) (Eqs. 7–9) is an explicit CTA-tile optimization objective for parallel agglomerative merging (Alg. 1), with density effects measured in Fig. 5c—not a proof that quality equals the cost objective by construction; (iii) cluster-level Top-p masking follows the SVG2 pipeline as a stated baseline, with quality still scored against full attention. Self-citations (e.g. Hyperattention) appear only in related work on LLM sparse attention and are not load-bearing uniqueness theorems for the video results. Hyperparameter choices (Table 2) are normal systems tuning, not fitted inputs renamed as first-principles predictions. No step reduces a claimed prediction to its own inputs by definition.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 3 invented entities

The central speed/quality claim rests on empirical domain regularities (locality and assignment stability), a hardware cost model tied to fixed CTA tiles, and many hand-chosen sparsity/clustering knobs. No new physical entities are postulated; the inventions are algorithmic. Free parameters dominate the ledger because Top-p, Min-kc, grids, windows, subset fraction, merge rounds, and warmups are set per model/task and directly shape reported density and fidelity.

free parameters (6)
  • Top-p cluster selection threshold
    Row-wise cumulative probability cutoff for selecting key clusters; fixed at 0.9 across settings and directly controls sparsity vs quality.
  • Min-kc minimum key-cluster retention ratio
    Per-setting floors (0.1–0.2) that force retention of key clusters even after Top-p; changes theoretical density and fidelity.
  • Q/K 3D grid and local window sizes
    Hand-chosen grids (e.g., (11,5,5)/(11,3,3) on Hunyuan) and windows (3,8,8) set centroid count and search volume; not derived from a uniqueness principle.
  • Hybrid clustering schedule and Lowe subset fraction
    Full clustering every two steps and caching top 15% highest Lowe’s-ratio tokens are chosen by hand; they control clustering overhead and assignment accuracy.
  • Parallel merge rounds R
    R=3 rounds of bipartition agglomerative merging; affects CTA density and extra tokens packed into idle tiles.
  • Time/layer full-attention warmup lengths
    Shared warmups (e.g., 10/50 timesteps; layer warmups) keep early steps dense; they are free design choices that improve quality at some speed cost.
axioms (6)
  • domain assumption Video diffusion latent tokens are locally coherent in space and time, so nearest centroids usually lie in a fixed 3D neighborhood.
    Motivated by PCA-colored cluster maps in Fig. 3a and used to justify O(n·window) local assignment instead of global k-means (§3, §4.1).
  • domain assumption Per-token cluster assignments are largely stable across adjacent denoising steps after early timesteps.
    Supported by change-rate heatmaps in Fig. 3b; enables hybrid full/partial reclustering (§3, §4.2).
  • domain assumption Lowe’s ratio (nearest/second-nearest centroid distance) identifies tokens most likely to change assignment.
    Fig. 3c and Eq. (6); used to select the intermediate-step update subset.
  • domain assumption FlashAttention-style kernels execute sparse attention in fixed CTA tiles (e.g., 128×96 on Hopper), so padded cluster blocks waste compute.
    Hardware premise for cost C(i)/C(S) and cluster merging (§3, §4.3, Eq. 7–9).
  • domain assumption Centroid-level scores with size-weighted softmax approximate cluster-pair importance well enough for Top-p masking.
    Inherited from SVG2-style clustering sparse attention (Eq. 2–4); quality claims depend on this proxy.
  • standard math Standard scaled-dot-product attention and DiT video generation setup (Wan2.2, HunyuanVideo) are valid evaluation substrates.
    Background ML machinery assumed throughout §2 and §5.
invented entities (3)
  • 3D local-window clustering (fixed grid + local centroid search) no independent evidence
    purpose: Replace global k-means assignment with O(n·wt·wh·ww) local search on the video token lattice.
    Algorithmic construct introduced in §4.1/Alg. 2; evidence is internal latency/error ablations, not an external physical prediction.
  • Hardware cost C(S) and parallel agglomerative cluster merging no independent evidence
    purpose: Minimize CTA-aligned execution cost of merged query groups while filling idle tile capacity.
    Defined in Eq. 7–10 and Alg. 1; validated only by density/latency ablations in this paper.
  • Hybrid anchor/subset reclustering schedule for denoising no independent evidence
    purpose: Avoid full reclustering every step by updating only high-Lowe’s-ratio tokens between anchors.
    Procedure in §4.2; depends on stability observations rather than external independent tests.

pith-pipeline@v1.1.0-grok45 · 23800 in / 4145 out tokens · 36089 ms · 2026-07-12T05:27:34.192180+00:00 · methodology

0 comments
read the original abstract

Video Diffusion Transformers (VDiTs) have demonstrated significant capabilities in high-fidelity video generation. However, their ability to produce long-duration videos is fundamentally constrained by the quadratic complexity of the self-attention mechanism. Recent clustering-based sparse attention methods improve the quality-speed trade-off by grouping semantically similar tokens, but their practical efficiency remains limited by two bottlenecks: substantial clustering overhead and low CTA utilization caused by irregular cluster-induced blocks. We propose HyperVAttention (HVA), a training-free sparse attention framework that addresses both bottlenecks jointly. To reduce clustering overhead, we introduce 3D local-window clustering, which exploits the spatio-temporal locality of video tokens to restrict centroid search to fixed local neighborhoods, and implement it with a custom Triton kernel for efficient execution. We further propose a hybrid clustering strategy that performs full clustering only at anchor steps and updates only subset tokens at intermediate steps, leveraging the temporal stability of cluster assignments across denoising steps. To improve CTA utilization, we present hardware-aware cluster merging that minimizes CTA-aligned execution cost through parallel agglomerative merging, improving block density and approximation fidelity by utilizing idle tile capacity. Together, these components reduce clustering overhead, avoid redundant updates, and better align sparse attention with the fixed tile structure of modern GPU kernels. Experiments on Text-to-Video generation show that HVA establishes a new Pareto frontier for training-free sparse attention in video diffusion, reducing end-to-end latency by up to $2.13\times$ while improving fidelity over existing training-free sparse attention baselines.

Figures

Figures reproduced from arXiv: 2607.03012 by Amir Zandieh, Dongyeun Lee, Insu Han, Junmo Kim, Vahab Mirrokni.

Figure 1
Figure 1. Figure 1: HyperVAttention is a training-free sparse attention framework for video diffusion that [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of how existing clustering-based sparse attention methods lead to thread [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: (a) Clustering results of video latent tokens. Frames are ordered from the top left to the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (a) Illustration of 3D Local-Window Clustering. (b) Latency comparison between PyTorch [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Efficiency analysis on diverse component of HyperVAttention. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of generated videos from Wan2.2 on Image-to-Video generation. [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of generated videos from Wan2.2 on Text-to-Video generation. [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of generated videos from HunyuanVideo on Text-to-Video generation. [PITH_FULL_IMAGE:figures/full_fig_p021_8.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

80 extracted references · 15 linked inside Pith

  1. [1]

    Slic superpixels compared to state-of-the-art superpixel methods.IEEE transactions on pattern analysis and machine intelligence, 34(11):2274–2282, 2012

    Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Süsstrunk. Slic superpixels compared to state-of-the-art superpixel methods.IEEE transactions on pattern analysis and machine intelligence, 34(11):2274–2282, 2012. 4

    2012

  2. [2]

    Monarchrt: Efficient attention for real-time video generation.arXiv preprint arXiv:2602.12271, 2026

    Krish Agarwal, Zhuoming Chen, Cheng Luo, Yongqi Chen, Haizhong Zheng, Xun Huang, Atri Rudra, and Beidi Chen. Monarchrt: Efficient attention for real-time video generation.arXiv preprint arXiv:2602.12271, 2026. 9

    arXiv 2026

  3. [3]

    Kandinsky 5.0: A family of foundation models for image and video generation.arXiv preprint arXiv:2511.14993, 2025

    Vladimir Arkhipkin, Vladimir Korviakov, Nikolai Gerasimenko, Denis Parkhomenko, Viach- eslav Vasilev, Alexey Letunovskiy, Nikolai Vaulin, Maria Kovaleva, Ivan Kirillov, Lev Novitskiy, et al. Kandinsky 5.0: A family of foundation models for image and video generation.arXiv preprint arXiv:2511.14993, 2025. 2

    Pith/arXiv arXiv 2025

  4. [4]

    Mixture of contexts for long video generation.arXiv preprint arXiv:2508.21058, 2025

    Shengqu Cai, Ceyuan Yang, Lvmin Zhang, Yuwei Guo, Junfei Xiao, Ziyan Yang, Yinghao Xu, Zhenheng Yang, Alan Yuille, Leonidas Guibas, et al. Mixture of contexts for long video generation.arXiv preprint arXiv:2508.21058, 2025. 2, 9

    arXiv 2025

  5. [5]

    Mllm-as-a-judge: Assessing multimodal llm-as- a-judge with vision-language benchmark

    Dongping Chen, Ruoxi Chen, Shilin Zhang, Yaochen Wang, Yinuo Liu, Huichi Zhou, Qihui Zhang, Yao Wan, Pan Zhou, and Lichao Sun. Mllm-as-a-judge: Assessing multimodal llm-as- a-judge with vision-language benchmark. InForty-first International Conference on Machine Learning, 2024. 7, 17

    2024

  6. [6]

    Sana-sprint: One-step diffusion with continuous-time consistency distillation

    Junsong Chen, Shuchen Xue, Yuyang Zhao, Jincheng Yu, Sayak Paul, Junyu Chen, Han Cai, Song Han, and Enze Xie. Sana-sprint: One-step diffusion with continuous-time consistency distillation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 16185–16195, 2025. 9

    2025

  7. [7]

    Sana-video: Efficient video generation with block linear diffusion transformer.arXiv preprint arXiv:2509.24695, 2025

    Junsong Chen, Yuyang Zhao, Jincheng Yu, Ruihang Chu, Junyu Chen, Shuai Yang, Xianbang Wang, Yicheng Pan, Daquan Zhou, Huan Ling, et al. Sana-video: Efficient video generation with block linear diffusion transformer.arXiv preprint arXiv:2509.24695, 2025. 9

    arXiv 2025

  8. [8]

    Deep compression autoencoder for efficient high-resolution diffusion models

    Junyu Chen, Han Cai, Junsong Chen, Enze Xie, Shang Yang, Haotian Tang, Muyang Li, Yao Lu, and Song Han. Deep compression autoencoder for efficient high-resolution diffusion models. arXiv preprint arXiv:2410.10733, 2024. 9

    Pith/arXiv arXiv 2024

  9. [9]

    Accelerating diffusion transformer via increment-calibrated caching with channel-aware singular value decomposition

    Zhiyuan Chen, Keyi Li, Yifan Jia, Le Ye, and Yufei Ma. Accelerating diffusion transformer via increment-calibrated caching with channel-aware singular value decomposition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18011–18020,

  10. [10]

    Flashattention: Fast and memory-efficient exact attention with io-awareness.Advances in neural information processing systems, 35:16344–16359, 2022

    Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness.Advances in neural information processing systems, 35:16344–16359, 2022. 2

    2022

  11. [11]

    Flex attention: A pro- gramming model for generating optimized attention kernels.arXiv preprint arXiv:2412.05496, 2(3):4, 2024

    Juechu Dong, Boyuan Feng, Driss Guessous, Yanbo Liang, and Horace He. Flex attention: A pro- gramming model for generating optimized attention kernels.arXiv preprint arXiv:2412.05496, 2(3):4, 2024. 2

    Pith/arXiv arXiv 2024

  12. [12]

    Veo 3.1: Text-to-video generation model

    Google DeepMind. Veo 3.1: Text-to-video generation model. https://gemini.google/ overview/video-generation/, 2026. 2

    2026

  13. [13]

    Hyperattention: Long-context attention in near-linear time

    Insu Han, Rajesh Jayaram, Amin Karbasi, Vahab Mirrokni, David Woodruff, and Amir Zandieh. Hyperattention: Long-context attention in near-linear time. InThe Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum? id=Eh0Od2BJIM. 2, 9

    2024

  14. [14]

    Efficientdm: Efficient quantization-aware fine-tuning of low-bit diffusion models

    Yefei He, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Efficientdm: Efficient quantization-aware fine-tuning of low-bit diffusion models. InInternational Conference on Learning Representations, 2024. 9 10

    2024

  15. [15]

    Tfmq-dm: Temporal feature maintenance quantization for diffusion models

    Yushi Huang, Ruihao Gong, Jing Liu, Tianlong Chen, and Xianglong Liu. Tfmq-dm: Temporal feature maintenance quantization for diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7362–7371, 2024. 9

    2024

  16. [16]

    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, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogniti...

    2024

  17. [17]

    Abdi, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, and Lili Qiu

    Huiqiang Jiang, Yucheng Li, Chengruidong Zhang, Qianhui Wu, Xufang Luo, Surin Ahn, Zhenhua Han, Amir H. Abdi, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. MInference 1.0: Accelerating pre-filling for long-context LLMs via dynamic sparse attention. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://o...

    2024

  18. [18]

    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, 7

    Pith/arXiv arXiv 2024

  19. [19]

    Flexprefill: A context- aware sparse attention mechanism for efficient long-sequence inference

    Xunhao Lai, Jianqiao Lu, Yao Luo, Yiyuan Ma, and Xun Zhou. Flexprefill: A context- aware sparse attention mechanism for efficient long-sequence inference. InThe Thirteenth International Conference on Learning Representations, 2025. URL https://openreview. net/forum?id=OfjIlbelrT. 9

    2025

  20. [20]

    Dmq: Dissecting outliers of diffusion models for post-training quantization

    Dongyeun Lee, Jiwan Hur, Hyounguk Shon, Jae Young Lee, and Junmo Kim. Dmq: Dissecting outliers of diffusion models for post-training quantization. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 18510–18520, 2025. 9

    2025

  21. [21]

    Svdquant: Absorbing outliers by low-rank components for 4-bit diffusion models.arXiv preprint arXiv:2411.05007, 2024

    Muyang Li, Yujun Lin, Zhekai Zhang, Tianle Cai, Xiuyu Li, Junxian Guo, Enze Xie, Chenlin Meng, Jun-Yan Zhu, and Song Han. Svdquant: Absorbing outliers by low-rank components for 4-bit diffusion models.arXiv preprint arXiv:2411.05007, 2024. 9

    arXiv 2024

  22. [22]

    Radial attention: O(nlogn) sparse attention for long video generation

    Xingyang Li, Muyang Li, Tianle Cai, Haocheng Xi, Shuo Yang, Yujun Lin, Lvmin Zhang, Songlin Yang, Jinbo Hu, Kelly Peng, Maneesh Agrawala, Ion Stoica, Kurt Keutzer, and Song Han. Radial attention: O(nlogn) sparse attention for long video generation. InThe Thirty- ninth Annual Conference on Neural Information Processing Systems, 2025. URL https: //openrevie...

    2025

  23. [23]

    Q-diffusion: Quantizing diffusion models

    Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535–17545, 2023. 9

    2023

  24. [24]

    MMIference: Accelerating pre-filling for long-context vlms via modality-aware permutation sparse at- tention

    Yucheng Li, Huiqiang Jiang, Chengruidong Zhang, Qianhui Wu, Xufang Luo, Surin Ahn, Amir H Abdi, Dongsheng Li, Jianfeng Gao, Yuqing Yang, and Lili Qiu. MMIference: Accelerating pre-filling for long-context vlms via modality-aware permutation sparse at- tention. InForty-second International Conference on Machine Learning, 2025. URL https://openreview.net/fo...

    2025

  25. [25]

    Vmonarch: Efficient video diffusion transformers with structured attention.arXiv preprint arXiv:2601.22275, 2026

    Cheng Liang, Haoxian Chen, Liang Hou, Qi Fan, Gangshan Wu, Xin Tao, and Limin Wang. Vmonarch: Efficient video diffusion transformers with structured attention.arXiv preprint arXiv:2601.22275, 2026. 9

    arXiv 2026

  26. [26]

    Ltx-2: Real-time video generation model

    Lightricks. Ltx-2: Real-time video generation model. https://www.lightricks.com/ltxv,

  27. [27]

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

    Feng Liu, Shiwei Zhang, Xiaofeng Wang, Yujie Wei, Haonan Qiu, Yuzhong Zhao, Yingya Zhang, Qixiang Ye, and Fang Wan. Timestep embedding tells: It’s time to cache for video diffusion model. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7353–7363, June 2025. 4, 9 11

    2025

  28. [28]

    Faster diffusion via temporal attention decomposition.Transactions on Machine Learning Research, 2025

    Haozhe Liu, Wentian Zhang, Jinheng Xie, Francesco Faccio, Mengmeng Xu, Tao Xiang, Mike Zheng Shou, Juan-Manuel Perez-Rua, and Jürgen Schmidhuber. Faster diffusion via temporal attention decomposition.Transactions on Machine Learning Research, 2025. URL https://openreview.net/forum?id=xXs2GKXPnH. 9

    2025

  29. [29]

    Distinctive image features from scale-invariant keypoints.International journal of computer vision, 60(2):91–110, 2004

    David G Lowe. Distinctive image features from scale-invariant keypoints.International journal of computer vision, 60(2):91–110, 2004. 4, 5

    2004

  30. [30]

    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. In The Thirteenth International Conference on Learning Representations, 2025. URL https: //openreview.net/forum?id=W49UjcpGxx. 4, 9

    2025

  31. [31]

    Model reveals what to cache: Profiling-based feature reuse for video diffusion models

    Xuran Ma, Yexin Liu, Yaofu Liu, Xianfeng Wu, Mingzhe Zheng, Zihao Wang, Ser-Nam Lim, and Harry Yang. Model reveals what to cache: Profiling-based feature reuse for video diffusion models. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 17150–17159, 2025. 9

    2025

  32. [32]

    Sora: Creating video from text.https://openai.com/sora, 2024

    OpenAI. Sora: Creating video from text.https://openai.com/sora, 2024. 2

    2024

  33. [33]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023. 2

    2023

  34. [34]

    Post-training quantization on diffusion models

    Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1972–1981, 2023. 9

    1972

  35. [35]

    Bitsfusion: 1.99 bits weight quantization of diffusion model

    Yang Sui, Yanyu Li, Anil Kag, Yerlan Idelbayev, Junli Cao, Ju Hu, Dhritiman Sagar, Bo Yuan, Sergey Tulyakov, and Jian Ren. Bitsfusion: 1.99 bits weight quantization of diffusion model. arXiv preprint arXiv:2406.04333, 2024. 9

    Pith/arXiv arXiv 2024

  36. [36]

    VORTA: Efficient video diffusion via routing sparse attention

    Wenhao Sun, Rong-Cheng Tu, Yifu Ding, Jingyi Liao, Zhao Jin, Shunyu Liu, and Dacheng Tao. VORTA: Efficient video diffusion via routing sparse attention. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. URL https://openreview. net/forum?id=gY9yOGYB48. 2, 9

    2025

  37. [37]

    Dsv: Exploiting dynamic sparsity to accelerate large-scale video dit training

    Xin Tan, Yuetao Chen, Yimin Jiang, Xing Chen, Kun Yan, Nan Duan, Yibo Zhu, Daxin Jiang, and Hong Xu. Dsv: Exploiting dynamic sparsity to accelerate large-scale video dit training. arXiv preprint arXiv:2502.07590, 2025. 9

    arXiv 2025

  38. [38]

    Quest: Query-aware sparsity for efficient long-context llm inference.arXiv preprint arXiv:2406.10774,

    Jiaming Tang, Yilong Zhao, Kan Zhu, Guangxuan Xiao, Baris Kasikci, and Song Han. Quest: Query-aware sparsity for efficient long-context llm inference.arXiv preprint arXiv:2406.10774,

    Pith/arXiv arXiv

  39. [39]

    Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805, 2023

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805, 2023. 7

    Pith/arXiv arXiv 2023

  40. [40]

    CUTLASS, January 2023

    Vijay Thakkar, Pradeep Ramani, Cris Cecka, Aniket Shivam, Honghao Lu, Ethan Yan, Jack Kosaian, Mark Hoemmen, Haicheng Wu, Andrew Kerr, Matt Nicely, Duane Merrill, Dustyn Blasig, Aditya Atluri, Fengqi Qiao, Piotr Majcher, Paul Springer, Markus Hohnerbach, Jin Wang, and Manish Gupta. CUTLASS, January 2023. URL https://github.com/NVIDIA/ cutlass. 5

    2023

  41. [41]

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

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025. 2, 7

    Pith/arXiv arXiv 2025

  42. [42]

    Cache me if you can: Accelerating diffusion models through block caching

    Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6211–6220, 2024. 4, 9 12

    2024

  43. [43]

    VMoBA: Mixture-of-block attention for video diffusion models

    Jianzong Wu, Liang Hou, Haotian Yang, Ye Tian, Pengfei Wan, Di ZHANG, and Yunhai Tong. VMoBA: Mixture-of-block attention for video diffusion models. InThe Fourteenth International Conference on Learning Representations, 2026. URL https://openreview. net/forum?id=oQaRElUdmh. 2, 9

    2026

  44. [44]

    Ptq4dit: Post-training quantization for diffusion transformers

    Junyi Wu, Haoxuan Wang, Yuzhang Shang, Mubarak Shah, and Yan Yan. Ptq4dit: Post-training quantization for diffusion transformers. InNeurIPS, 2024. 9

    2024

  45. [45]

    Sparse video-gen: Accelerating video diffusion transformers with spatial-temporal sparsity

    Haocheng Xi, Shuo Yang, Yilong Zhao, Chenfeng Xu, Muyang Li, Xiuyu Li, Yujun Lin, Han Cai, Jintao Zhang, Dacheng Li, et al. Sparse video-gen: Accelerating video diffusion transformers with spatial-temporal sparsity. InInternational Conference on Machine Learning, pages 68208–68224. PMLR, 2025. 2, 7, 9

    2025

  46. [46]

    Training-free and adaptive sparse attention for efficient long video generation

    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. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 15982–15993, 2025. 2, 9

    2025

  47. [47]

    Efficient streaming language models with attention sinks.arXiv, 2023

    Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks.arXiv, 2023. 9

    2023

  48. [48]

    Sana: Efficient high-resolution image synthesis with linear diffusion transformers.arXiv preprint arXiv:2410.10629, 2024

    Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution image synthesis with linear diffusion transformers.arXiv preprint arXiv:2410.10629, 2024. 9

    Pith/arXiv arXiv 2024

  49. [49]

    Xattention: Block sparse attention with antidiagonal scoring

    Ruyi Xu, Guangxuan Xiao, Haofeng Huang, Junxian Guo, and Song Han. Xattention: Block sparse attention with antidiagonal scoring. InProceedings of the 42nd International Conference on Machine Learning (ICML), 2025. 9

    2025

  50. [50]

    Sparse videogen2: Accelerate video generation with sparse attention via semantic-aware permutation.arXiv preprint arXiv:2505.18875, 2025

    Shuo Yang, Haocheng Xi, Yilong Zhao, Muyang Li, Jintao Zhang, Han Cai, Yujun Lin, Xiuyu Li, Chenfeng Xu, Kelly Peng, et al. Sparse videogen2: Accelerate video generation with sparse attention via semantic-aware permutation.arXiv preprint arXiv:2505.18875, 2025. 2, 7, 8, 9

    Pith/arXiv arXiv 2025

  51. [51]

    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, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024. 2

    Pith/arXiv arXiv 2024

  52. [52]

    Flashinfer: Efficient and customizable attention engine for llm inference serving.Proceedings of Machine Learning and Systems, 7, 2025

    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.Proceedings of Machine Learning and Systems, 7, 2025. 2, 6

    2025

  53. [53]

    Improved distribution matching distillation for fast image synthesis

    Tianwei Yin, Michaël Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. Advances in neural information processing systems, 37:47455–47487, 2024. 9

    2024

  54. [54]

    One-step diffusion with distribution matching distillation

    Tianwei Yin, Michaël Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6613–6623, 2024. 9

    2024

  55. [55]

    Sageatten- tion2: Efficient attention with thorough outlier smoothing and per-thread int4 quantization

    Jintao Zhang, Haofeng Huang, Pengle Zhang, Jia Wei, Jun Zhu, and Jianfei Chen. Sageatten- tion2: Efficient attention with thorough outlier smoothing and per-thread int4 quantization. In International Conference on Machine Learning (ICML), 2025. 9

    2025

  56. [56]

    Sla: Beyond sparsity in diffusion transformers via fine-tunable sparse-linear attention.arXiv preprint arXiv:2509.24006, 2025

    Jintao Zhang, Haoxu Wang, Kai Jiang, Shuo Yang, Kaiwen Zheng, Haocheng Xi, Ziteng Wang, Hongzhou Zhu, Min Zhao, Ion Stoica, et al. Sla: Beyond sparsity in diffusion transformers via fine-tunable sparse-linear attention.arXiv preprint arXiv:2509.24006, 2025. 9

    arXiv 2025

  57. [57]

    Sageattention3: Microscaling fp4 attention for inference and an exploration of 8-bit training.arXiv preprint arXiv:2505.11594, 2025

    Jintao Zhang, Jia Wei, Pengle Zhang, Xiaoming Xu, Haofeng Huang, Haoxu Wang, Kai Jiang, Jun Zhu, and Jianfei Chen. Sageattention3: Microscaling fp4 attention for inference and an exploration of 8-bit training.arXiv preprint arXiv:2505.11594, 2025. 9 13

    arXiv 2025

  58. [58]

    Sageattention: Accurate 8-bit attention for plug-and-play inference acceleration

    Jintao Zhang, Jia Wei, Pengle Zhang, Jun Zhu, and Jianfei Chen. Sageattention: Accurate 8-bit attention for plug-and-play inference acceleration. InInternational Conference on Learning Representations (ICLR), 2025. 9

    2025

  59. [59]

    Spargeattention: Accurate and training-free sparse attention accelerating any model inference

    Jintao Zhang, Chendong Xiang, Haofeng Huang, Jia Wei, Haocheng Xi, Jun Zhu, and Jianfei Chen. Spargeattention: Accurate and training-free sparse attention accelerating any model inference. InInternational Conference on Machine Learning, pages 76397–76413. PMLR,

  60. [60]

    Spargeattention2: Trainable sparse attention via hybrid top-k+ top-p masking and distillation fine-tuning.arXiv preprint arXiv:2602.13515, 2026

    Jintao Zhang, Kai Jiang, Chendong Xiang, Weiqi Feng, Yuezhou Hu, Haocheng Xi, Jianfei Chen, and Jun Zhu. Spargeattention2: Trainable sparse attention via hybrid top-k+ top-p masking and distillation fine-tuning.arXiv preprint arXiv:2602.13515, 2026. 9

    arXiv 2026

  61. [61]

    Sla2: Sparse-linear attention with learnable routing and qat

    Jintao Zhang, Haoxu Wang, Kai Jiang, Kaiwen Zheng, Youhe Jiang, Ion Stoica, Jianfei Chen, Jun Zhu, and Joseph E Gonzalez. Sla2: Sparse-linear attention with learnable routing and qat. arXiv preprint arXiv:2602.12675, 2026. 9

    arXiv 2026

  62. [62]

    Xing, and Hao Zhang

    Peiyuan Zhang, Yongqi Chen, Haofeng Huang, Will Lin, Zhengzhong Liu, Ion Stoica, Eric P. Xing, and Hao Zhang. Faster video diffusion with trainable sparse attention. InThe Thirty- ninth Annual Conference on Neural Information Processing Systems, 2025. URL https: //openreview.net/forum?id=VrYCLQ5inI. 9

    2025

  63. [63]

    Fast video generation with sliding tile attention

    Peiyuan Zhang, Yongqi Chen, Runlong Su, Hangliang Ding, Ion Stoica, Zhengzhong Liu, and Hao Zhang. Fast video generation with sliding tile attention. InForty-second Interna- tional Conference on Machine Learning, 2025. URL https://openreview.net/forum? id=U74MOXPEJd. 2, 9

    2025

  64. [64]

    The unreason- able effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreason- able effectiveness of deep features as a perceptual metric. InCVPR, 2018. 7

    2018

  65. [65]

    H2o: Heavy-hitter oracle for efficient generative inference of large language models.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 Ré, Clark Barrett, et al. H2o: Heavy-hitter oracle for efficient generative inference of large language models.Advances in Neural Information Processing Systems, 36:34661–34710, 2023. 9

    2023

  66. [66]

    Vidit-q: Efficient and accurate quantization of diffusion transformers for image and video generation.arXiv preprint arXiv:2406.02540, 2024

    Tianchen Zhao, Tongcheng Fang, Enshu Liu, Wan Rui, Widyadewi Soedarmadji, Shiyao Li, Zinan Lin, Guohao Dai, Shengen Yan, Huazhong Yang, et al. Vidit-q: Efficient and accurate quantization of diffusion transformers for image and video generation.arXiv preprint arXiv:2406.02540, 2024. 9

    Pith/arXiv arXiv 2024

  67. [67]

    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. InThe Thirteenth International Conference on Learning Representations,

  68. [68]

    URLhttps://openreview.net/forum?id=hDBrQ4DApF. 4, 9

  69. [69]

    Large scale diffusion distillation via score-regularized continuous-time consistency.arXiv preprint arXiv:2510.08431, 2025

    Kaiwen Zheng, Yuji Wang, Qianli Ma, Huayu Chen, Jintao Zhang, Yogesh Balaji, Jianfei Chen, Ming-Yu Liu, Jun Zhu, and Qinsheng Zhang. Large scale diffusion distillation via score-regularized continuous-time consistency.arXiv preprint arXiv:2510.08431, 2025. 9

    Pith/arXiv arXiv 2025

  70. [70]

    Judging llm-as-a-judge with mt-bench and chatbot arena.Advances in neural information processing systems, 36:46595–46623, 2023

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena.Advances in neural information processing systems, 36:46595–46623, 2023. 7, 17

    2023

  71. [71]

    Open-sora: Democratizing efficient video production for all

    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. arXiv preprint arXiv:2412.20404, 2024. 2

    Pith/arXiv arXiv 2024

  72. [72]

    Svg-ear: Parameter-free linear compen- sation for sparse video generation via error-aware routing.arXiv preprint arXiv:2603.08982,

    Xuanyi Zhou, Qiuyang Mang, Shuo Yang, Haocheng Xi, Jintao Zhang, Huanzhi Mao, Joseph E Gonzalez, Kurt Keutzer, Ion Stoica, and Alvin Cheung. Svg-ear: Parameter-free linear compen- sation for sparse video generation via error-aware routing.arXiv preprint arXiv:2603.08982,

    Pith/arXiv arXiv

  73. [73]

    Stereo for image-based rendering using image over-segmentation.International Journal of Computer Vision, 75(1):49–65, 2007

    C Lawrence Zitnick and Sing Bing Kang. Stereo for image-based rendering using image over-segmentation.International Journal of Computer Vision, 75(1):49–65, 2007. 4 14 A Implementation Details A.1 Hyperparameters For hybrid clustering, we perform full clustering every two denoising steps and cache the top15% most ambiguous tokens, ranked by Lowe’s ratio, ...

    2007

  74. [74]

    Your task is to evaluate an AI-generated video based on specific technical and artistic dimensions

    System Role - You are an expert Video Quality Assurance Engineer and Cinematographer. Your task is to evaluate an AI-generated video based on specific technical and artistic dimensions. You must provide objective, critical scores from 1 (Poor) to 10 (Excellent). ,→ ,→

  75. [75]

    - Assess the motion dynamics and temporal stability

    Evaluation Process - Analyze the video frame-by-frame to detect inconsistencies. - Assess the motion dynamics and temporal stability. - Score each dimension based on the strict criteria provided. - Calculate the final average

  76. [76]

    Strict Evaluation Criteria Dimension | Description & Scoring Guidelines

  77. [77]

    Subject Consistency | Does the main subject maintain its identity (features, clothing, shape) throughout the entire duration? Penalty for morphing.,→

  78. [78]

    cinematic

    Aesthestic Quality| General visual appeal, composition, lighting, and "cinematic" feel

  79. [79]

    flashing

    Temporal Flickering | Presence of high-frequency noise or "flashing" brightness/texture changes between frames.,→

  80. [80]

    overall_score

    Required Output Format You MUST ONLY output a single, valid JSON object. Do NOT include any introductory text, analysis, or conversational response outside of the JSON.,→ The JSON object MUST contain three keys: overall_score, and metrics. - overall_score: the simple average of all the metric scores - metrics: a dictionary containing the required metrics ...