Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

Sparse VideoGen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (2025)
DiTCtrl: Exploring Attention Control in Multi-Modal Diffusion Transformer for Tuning-Free Multi-Prompt Longer Video Generation (2024)
MSWA: Refining Local Attention with Multi-ScaleWindow Attention (2025)
UniCP: A Unified Caching and Pruning Framework for Efficient Video Generation (2025)
CascadeV: An Implementation of Wurstchen Architecture for Video Generation (2025)
AsymRnR: Video Diffusion Transformers Acceleration with Asymmetric Reduction and Restoration (2024)
Flash Window Attention: speedup the attention computation for Swin Transformer (2025)

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: \n\n@librarian-bot\n\t recommend

\n","updatedAt":"2025-02-11T01:35:53.873Z","author":{"_id":"63d3e0e8ff1384ce6c5dd17d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg","fullname":"Librarian Bot (Bot)","name":"librarian-bot","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":318,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.6835502982139587},"editors":["librarian-bot"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2502.04507","authors":[{"_id":"67a98cd1b8b21202c9004628","user":{"_id":"63565cc56d7fcf1bedb7d347","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63565cc56d7fcf1bedb7d347/XGcHP4VkO_oieA1gZ4IAX.jpeg","isPro":false,"fullname":"Zhang Peiyuan","user":"PY007","type":"user"},"name":"Peiyuan Zhang","status":"admin_assigned","statusLastChangedAt":"2025-02-10T16:07:27.309Z","hidden":false},{"_id":"67a98cd1b8b21202c9004629","user":{"_id":"65416817271d3bc4d70f6745","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65416817271d3bc4d70f6745/1YkW0MpuufejvxqksVMIx.jpeg","isPro":false,"fullname":"Yongqi Chen","user":"BrianChen1129","type":"user"},"name":"Yongqi Chen","status":"claimed_verified","statusLastChangedAt":"2025-02-10T09:49:48.410Z","hidden":false},{"_id":"67a98cd1b8b21202c900462a","user":{"_id":"65d7ed4823e83e1591beacc7","avatarUrl":"/avatars/2a6714a2a7bbd591f6b726a7330bafbc.svg","isPro":false,"fullname":"Su","user":"r3su9","type":"user"},"name":"Runlong Su","status":"claimed_verified","statusLastChangedAt":"2025-02-11T10:00:40.942Z","hidden":false},{"_id":"67a98cd1b8b21202c900462b","user":{"_id":"643a451ee2b979ae6141329d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/643a451ee2b979ae6141329d/HN3M5vyroanQoUEiXJFyB.jpeg","isPro":false,"fullname":"Hangliang Ding","user":"foreverpiano","type":"user"},"name":"Hangliang Ding","status":"claimed_verified","statusLastChangedAt":"2025-02-11T07:56:04.110Z","hidden":false},{"_id":"67a98cd1b8b21202c900462c","name":"Ion Stoica","hidden":false},{"_id":"67a98cd1b8b21202c900462d","name":"Zhenghong Liu","hidden":false},{"_id":"67a98cd1b8b21202c900462e","user":{"_id":"62d363143eebd640a4fa41fa","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/62d363143eebd640a4fa41fa/pvPwXlJ5OOb-UIfmffv4E.jpeg","isPro":false,"fullname":"Hao Zhang","user":"zhisbug","type":"user"},"name":"Hao Zhang","status":"claimed_verified","statusLastChangedAt":"2025-02-11T10:00:38.019Z","hidden":false}],"publishedAt":"2025-02-06T21:17:09.000Z","submittedOnDailyAt":"2025-02-10T02:52:26.568Z","title":"Fast Video Generation with Sliding Tile Attention","submittedOnDailyBy":{"_id":"63565cc56d7fcf1bedb7d347","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63565cc56d7fcf1bedb7d347/XGcHP4VkO_oieA1gZ4IAX.jpeg","isPro":false,"fullname":"Zhang Peiyuan","user":"PY007","type":"user"},"summary":"Diffusion Transformers (DiTs) with 3D full attention power state-of-the-art\nvideo generation, but suffer from prohibitive compute cost -- when generating\njust a 5-second 720P video, attention alone takes 800 out of 945 seconds of\ntotal inference time. This paper introduces sliding tile attention (STA) to\naddress this challenge. STA leverages the observation that attention scores in\npretrained video diffusion models predominantly concentrate within localized 3D\nwindows. By sliding and attending over the local spatial-temporal region, STA\neliminates redundancy from full attention. Unlike traditional token-wise\nsliding window attention (SWA), STA operates tile-by-tile with a novel\nhardware-aware sliding window design, preserving expressiveness while being\nhardware-efficient. With careful kernel-level optimizations, STA offers the\nfirst efficient 2D/3D sliding-window-like attention implementation, achieving\n58.79% MFU. Precisely, STA accelerates attention by 2.8-17x over\nFlashAttention-2 (FA2) and 1.6-10x over FlashAttention-3 (FA3). On the leading\nvideo DiT, HunyuanVideo, STA reduces end-to-end latency from 945s (FA3) to 685s\nwithout quality degradation, requiring no training. Enabling finetuning further\nlowers latency to 268s with only a 0.09% drop on VBench.","upvotes":51,"discussionId":"67a98cd7b8b21202c90047c5","ai_summary":"Sliding tile attention (STA) accelerates 3D attention in Diffusion Transformers (DiTs), reducing video generation latency without quality loss.","ai_keywords":["diffusion transformers","DiTs","3D full attention","sliding tile attention","STA","attention scores","hardware-aware sliding window","2D/3D sliding-window-like attention","FlashAttention-2","FlashAttention-3","HunyuanVideo","VBench"]},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"62ba66296501b0ff15ba1075","avatarUrl":"/avatars/de468727cb1240fd4b1f24a19fb237a6.svg","isPro":true,"fullname":"Yichao Fu","user":"Viol2000","type":"user"},{"_id":"62d363143eebd640a4fa41fa","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/62d363143eebd640a4fa41fa/pvPwXlJ5OOb-UIfmffv4E.jpeg","isPro":false,"fullname":"Hao Zhang","user":"zhisbug","type":"user"},{"_id":"64e5d04767a759dd7155a445","avatarUrl":"/avatars/70090182b71db4de31fb8c325d25e73b.svg","isPro":false,"fullname":"Yinmin Zhong","user":"PKUFlyingPig","type":"user"},{"_id":"6301d6455e305a35cb0846a7","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6301d6455e305a35cb0846a7/aT2AtzRMSY_T3y02MIUap.jpeg","isPro":true,"fullname":"Lanxiang Hu","user":"Snyhlxde","type":"user"},{"_id":"63565cc56d7fcf1bedb7d347","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63565cc56d7fcf1bedb7d347/XGcHP4VkO_oieA1gZ4IAX.jpeg","isPro":false,"fullname":"Zhang Peiyuan","user":"PY007","type":"user"},{"_id":"675893bc6f24ab221a777c1e","avatarUrl":"/avatars/ac8e04fe7e6e0032503053071410dae2.svg","isPro":false,"fullname":"Guangtao Zeng","user":"GGGGGuangtao","type":"user"},{"_id":"65621fd68631d43d2baf33b2","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/1JemNCnPkS1mE3SNygsE2.png","isPro":false,"fullname":"siqi zhu","user":"zsqzz","type":"user"},{"_id":"646e1ef5075bbcc48ddf21e8","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/_vJC0zeVOIvaNV2R6toqg.jpeg","isPro":false,"fullname":"Pu Fanyi","user":"pufanyi","type":"user"},{"_id":"643839d9581e6bf0fa9c835e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/643839d9581e6bf0fa9c835e/JxlgR-zQhms-rfF0sDxD8.jpeg","isPro":false,"fullname":"Junda Chen","user":"GindaChen","type":"user"},{"_id":"64bba541da140e461924dfed","avatarUrl":"/avatars/367993765b0ca3734b2b100db33ed787.svg","isPro":true,"fullname":"zhijie deng","user":"zhijie3","type":"user"},{"_id":"609115c79a8bcaa437b234a9","avatarUrl":"/avatars/1631a91030703d8397133363cf82c863.svg","isPro":false,"fullname":"Leng Sicong","user":"Sicong","type":"user"},{"_id":"64bb77e786e7fb5b8a317a43","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64bb77e786e7fb5b8a317a43/J0jOrlZJ9gazdYaeSH2Bo.png","isPro":false,"fullname":"kcz","user":"kcz358","type":"user"}],"acceptLanguages":["*"],"dailyPaperRank":0}">

Papers

arxiv:2502.04507

Fast Video Generation with Sliding Tile Attention

Published on Feb 6, 2025

· Submitted by

Zhang Peiyuan on Feb 10, 2025

Upvote

Authors:

Peiyuan Zhang ,

Yongqi Chen ,

Runlong Su ,

Hangliang Ding ,

Hao Zhang

Abstract

Sliding tile attention (STA) accelerates 3D attention in Diffusion Transformers (DiTs), reducing video generation latency without quality loss.

AI-generated summary

Diffusion Transformers (DiTs) with 3D full attention power state-of-the-art video generation, but suffer from prohibitive compute cost -- when generating just a 5-second 720P video, attention alone takes 800 out of 945 seconds of total inference time. This paper introduces sliding tile attention (STA) to address this challenge. STA leverages the observation that attention scores in pretrained video diffusion models predominantly concentrate within localized 3D windows. By sliding and attending over the local spatial-temporal region, STA eliminates redundancy from full attention. Unlike traditional token-wise sliding window attention (SWA), STA operates tile-by-tile with a novel hardware-aware sliding window design, preserving expressiveness while being hardware-efficient. With careful kernel-level optimizations, STA offers the first efficient 2D/3D sliding-window-like attention implementation, achieving 58.79% MFU. Precisely, STA accelerates attention by 2.8-17x over FlashAttention-2 (FA2) and 1.6-10x over FlashAttention-3 (FA3). On the leading video DiT, HunyuanVideo, STA reduces end-to-end latency from 945s (FA3) to 685s without quality degradation, requiring no training. Enabling finetuning further lowers latency to 268s with only a 0.09% drop on VBench.