PageSourceSearch

https://pardistaghavi.github.io/SparsePR-website/

html pardistaghavi.github.io collected 2026-10-03 09:44:22 UTC 39,923 bytes, 55 lines download raw bytes

1<!doctype html>
2<html lang="en">
3  <head>
4    <meta charset="UTF-8" />
5    <meta name="viewport" content="width=device-width, initial-scale=1.0" />
6    <meta name="description" content="SparsePR is training-free sparse attention for video generation and world models, evaluated across HunyuanVideo, Wan2.2, Cosmos-Predict2.5, and Cosmos3-Nano." />
7    <meta name="msvalidate.01" content="8149475937CBBE15CEA9FF9F5690C937" />
8    <link rel="preconnect" href="https://fonts.googleapis.com" />
9    <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin />
10    <link href="https://fonts.googleapis.com/css2?family=Google+Sans+Flex:opsz,[email protected],1..1000&display=swap" rel="stylesheet" />
11    <link rel="icon" href="/SparsePR-website/favicon.svg" type="image/svg+xml" />
12    <link rel="manifest" href="/SparsePR-website/site.webmanifest" />
13    <title>SparsePR: Training-Free Sparse Attention for Video Generation</title>
14    <meta name="robots" content="index,follow,max-image-preview:large,max-video-preview:-1,max-snippet:-1" />
15    <meta name="citation_title" content="Partition the Support, Reconstruct the Residual: Training-Free Sparse Attention for Video Generation and World Models" />
16    <meta name="citation_author" content="Pardis Taghavi" />
17    <meta name="citation_author" content="Reza Langari" />
18    <meta name="citation_author" content="Gaurav Pandey" />
19    <meta name="citation_publication_date" content="2026/08/19" />
20    <meta name="citation_arxiv_id" content="2608.18484" />
21    <meta name="citation_abstract_html_url" content="https://arxiv.org/abs/2608.18484" />
22    <meta name="citation_pdf_url" content="https://arxiv.org/pdf/2608.18484" />
23    <link rel="alternate" type="application/pdf" href="https://arxiv.org/pdf/2608.18484" />
24    <link rel="canonical" href="https://pardistaghavi.github.io/SparsePR-website/" />
25    <meta property="og:title" content="SparsePR: Training-Free Sparse Attention for Video Generation" />
26    <meta property="og:description" content="SparsePR is training-free sparse attention for video generation and world models, evaluated across HunyuanVideo, Wan2.2, Cosmos-Predict2.5, and Cosmos3-Nano." />
27    <meta property="og:type" content="website" />
28    <meta property="og:url" content="https://pardistaghavi.github.io/SparsePR-website/" />
29    <meta property="og:image" content="https://pardistaghavi.github.io/SparsePR-website/og.png" />
30    <meta property="og:image:width" content="1200" />
31    <meta property="og:image:height" content="630" />
32    <meta name="twitter:card" content="summary_large_image" />
33    <meta name="twitter:title" content="SparsePR: Training-Free Sparse Attention for Video Generation" />
34    <meta name="twitter:description" content="SparsePR is training-free sparse attention for video generation and world models, evaluated across HunyuanVideo, Wan2.2, Cosmos-Predict2.5, and Cosmos3-Nano." />
35    <meta name="twitter:image" content="https://pardistaghavi.github.io/SparsePR-website/og.png" />
36    
36<script type="application/ld+json">{"@context":"https://schema.org","@graph":[{"@type":"WebSite","@id":"https://pardistaghavi.github.io/SparsePR-website/#website","url":"https://pardistaghavi.github.io/SparsePR-website/","name":"SparsePR","description":"Training-free sparse attention for video generation and world models.","inLanguage":"en"},{"@type":"ScholarlyArticle","@id":"https://pardistaghavi.github.io/SparsePR-website/#paper","headline":"Partition the Support, Reconstruct the Residual: Training-Free Sparse Attention for Video Generation and World Models","author":[{"@type":"Person","name":"Pardis Taghavi","affiliation":{"@type":"CollegeOrUniversity","name":"Texas A&M University"}},{"@type":"Person","name":"Reza Langari","affiliation":{"@type":"CollegeOrUniversity","name":"Texas A&M University"}},{"@type":"Person","name":"Gaurav Pandey","affiliation":{"@type":"CollegeOrUniversity","name":"Texas A&M University"}}],"datePublished":"2026-08-19","identifier":"arXiv:2608.18484","image":"https://pardistaghavi.github.io/SparsePR-website/og.png","url":"https://arxiv.org/abs/2608.18484","sameAs":"https://arxiv.org/abs/2608.18484","encoding":{"@type":"MediaObject","contentUrl":"https://arxiv.org/pdf/2608.18484","encodingFormat":"application/pdf"},"isAccessibleForFree":true},{"@type":"SoftwareSourceCode","@id":"https://pardistaghavi.github.io/SparsePR-website/#code","name":"SparsePR","description":"Open-source implementation of training-free sparse attention for video generation and world models.","codeRepository":"https://github.com/PardisTaghavi/SparsePR","url":"https://github.com/PardisTaghavi/SparsePR","author":{"@type":"Person","name":"Pardis Taghavi"}},{"@type":"WebPage","@id":"https://pardistaghavi.github.io/SparsePR-website/#page","url":"https://pardistaghavi.github.io/SparsePR-website/","name":"SparsePR: Training-Free Sparse Attention for Video Generation","description":"SparsePR is training-free sparse attention for video generation and world models, evaluated across HunyuanVideo, Wan2.2, Cosmos-Predict2.5, and Cosmos3-Nano.","isPartOf":{"@id":"https://pardistaghavi.github.io/SparsePR-website/#website"},"about":{"@id":"https://pardistaghavi.github.io/SparsePR-website/#paper"},"primaryImageOfPage":{"@type":"ImageObject","url":"https://pardistaghavi.github.io/SparsePR-website/og.png"},"inLanguage":"en"}]}</script>
vendor: 68 bytes, lines 36-37
36
37    <script async src="https://www.googletagmanager.com/gtag/js?id=
37G-JTLS5XBWZM
vendor: 16 bytes, lines 37-38
37"></script>
38    
38<script>
39      
vendor: 145 bytes, lines 39-42
39window.dataLayer = window.dataLayer || [];
40      function gtag(){dataLayer.push(arguments);}
41      gtag('js', new Date());
42      gtag('config', '
42G-JTLS5XBWZM
vendor: 8 bytes, lines 42-43
42');
43    
43</script>
43
44    
44<script type="module" crossorigin src="/SparsePR-website/assets/index-C2CdUCLW.js"></script>
44
45    <link rel="stylesheet" crossorigin href="/SparsePR-website/assets/index-D0GvtT_a.css">
46  </head>
47  <body>
48    <div id="root"><link rel="preload" as="image" href="/SparsePR-website/media/figures/structural-observations-final.png"/><link rel="preload" as="image" href="/SparsePR-website/media/figures/results-final-pdf.png"/><main><nav class="topbar" aria-label="Primary navigation"><a class="brand" href="/SparsePR-website/">SparsePR</a><div class="navlinks"><a href="/SparsePR-website/videos/">Videos</a><a href="/SparsePR-website/method/">Method</a><a href="/SparsePR-website/results/">Results</a><a href="/SparsePR-website/#citation">Citation</a></div></nav><header id="top" class="hero content"><div class="eyebrow">Training-free sparse attention</div><h1><span>SparsePR</span></h1><p class="hero-kicker">Partition the Support. Reconstruct the Residual.</p><h2>Training-Free Sparse Attention for<br class="desktop-break"/> Video Generation and World Models</h2><p class="authors">Pardis Taghavi · Reza Langari · Gaurav Pandey</p><p class="affiliation">Texas A&amp;M University</p><div class="hero-actions"><a class="button primary" href="https://arxiv.org/abs/2608.18484" target="_blank" rel="noreferrer">Paper ↗</a><a class="button" href="https://github.com/PardisTaghavi/SparsePR" target="_blank" rel="noreferrer">Code ↗</a><a class="button" href="/SparsePR-website/results/">Results →</a><a class="button" href="#citation">BibTeX</a></div><div class="tldr"><strong>TL;DR</strong><span>SparsePR couples executable response-aware partitions with a small set of exact probe rows to recover the residual that sparse attention leaves behind.</span></div></header><section class="gallery-section" id="videos" aria-labelledby="teaser-heading"><div class="section-intro compact"><p class="section-number">See the result</p><h2 id="teaser-heading">SparsePR across four video models.</h2></div><div class="model-galleries"><section class="model-gallery" aria-label="Cosmos-Predict2.5-14B qualitative video gallery"><div class="gallery-heading"><div><h3>Cosmos-Predict2.5-14B</h3><p>Image-to-world · native 720p class</p></div></div><div class="gallery-stage"><button class="gallery-arrow previous" aria-label="Previous Cosmos-Predict2.5-14B videos">←</button><div class="gallery-grid"><article class="gallery-video" tabindex="0" aria-label="VBench video. Hover or focus to read the prompt."><video muted="" loop="" playsInline="" preload="metadata" poster="/SparsePR-website/media/thumbnails/cosmos25/0000_a_view_of_a_star_trail_in_the_night_sky.jpg"><source src="/SparsePR-website/media/gallery/cosmos25/0000_a_view_of_a_star_trail_in_the_night_sky.mp4" type="video/mp4"/></video><div class="prompt-overlay"><span>Prompt</span><p>A view of a star trail in the night sky.</p></div><div class="clip-meta"><span>VBench</span><strong>40.33 dB</strong></div></article><article class="gallery-video" tabindex="0" aria-label="VBench video. Hover or focus to read the prompt."><video muted="" loop="" playsInline="" preload="metadata" poster="/SparsePR-website/media/thumbnails/cosmos25/0012_an_aerial_view_of_a_rocky_beach_in_indonesia.jpg"><source src="/SparsePR-website/media/gallery/cosmos25/0012_an_aerial_view_of_a_rocky_beach_in_indonesia.mp4" type="video/mp4"/></video><div class="prompt-overlay"><span>Prompt</span><p>An aerial view of a rocky beach in Indonesia.</p></div><div class="clip-meta"><span>VBench</span><strong>26.19 dB</strong></div></article><article class="gallery-video" tabindex="0" aria-label="VBench video. Hover or focus to read the prompt."><video muted="" loop="" playsInline="" preload="metadata" poster="/SparsePR-website/media/thumbnails/cosmos25/0019_a_butterfly_sits_on_top_of_a_purple_flower.jpg"><source src="/SparsePR-website/media/gallery/cosmos25/0019_a_butterfly_sits_on_top_of_a_purple_flower.mp4" type="video/mp4"/></video><div class="prompt-overlay"><span>Prompt</span><p>A butterfly sits on top of a purple flower.</p></div><div class="clip-meta"><span>VBench</span><strong>19.82 dB</strong></div></article></div><button class="gallery-arrow next" aria-label="Next Cosmos-Predict2.5-14B videos">→</button></div>
48<div class="gallery-position" aria-live="polite">01<!-- --> / <!-- -->10</div></section><section class="model-gallery" aria-label="Cosmos3-Nano-16B qualitative video gallery"><div class="gallery-heading"><div><h3>Cosmos3-Nano-16B</h3><p>Image-to-world · 720p</p></div></div><div class="gallery-stage"><button class="gallery-arrow previous" aria-label="Previous Cosmos3-Nano-16B videos">←</button><div class="gallery-grid"><article class="gallery-video" tabindex="0" aria-label="VBench video. Hover or focus to read the prompt."><video muted="" loop="" playsInline="" preload="metadata" poster="/SparsePR-website/media/thumbnails/cosmos3/0117_a_highland_cow_with_long_horns_standing_in_a_field.jpg"><source src="/SparsePR-website/media/gallery/cosmos3/0117_a_highland_cow_with_long_horns_standing_in_a_field.mp4" type="video/mp4"/></video><div class="prompt-overlay"><span>Prompt</span><p>A highland cow with long horns standing in a field.</p></div><div class="clip-meta"><span>VBench</span><strong>26.07 dB</strong></div></article><article class="gallery-video" tabindex="0" aria-label="VBench video. Hover or focus to read the prompt."><video muted="" loop="" playsInline="" preload="metadata" poster="/SparsePR-website/media/thumbnails/cosmos3/0138_a_giraffe_walking_in_a_field.jpg"><source src="/SparsePR-website/media/gallery/cosmos3/0138_a_giraffe_walking_in_a_field.mp4" type="video/mp4"/></video><div class="prompt-overlay"><span>Prompt</span><p>A giraffe walking in a field.</p></div><div class="clip-meta"><span>VBench</span><strong>22.92 dB</strong></div></article><article class="gallery-video" tabindex="0" aria-label="VBench video. Hover or focus to read the prompt."><video muted="" loop="" playsInline="" preload="metadata" poster="/SparsePR-website/media/thumbnails/cosmos3/0141_a_warthog_is_walking_in_the_grass.jpg"><source src="/SparsePR-website/media/gallery/cosmos3/0141_a_warthog_is_walking_in_the_grass.mp4" type="video/mp4"/></video><div class="prompt-overlay"><span>Prompt</span><p>A warthog is walking in the grass.</p></div><div class="clip-meta"><span>VBench</span><strong>21.41 dB</strong></div></article></div><button class="gallery-arrow next" aria-label="Next Cosmos3-Nano-16B videos">→</button></div><div class="gallery-position" aria-live="polite">01<!-- --> / <!-- -->07</div></section><section class="model-gallery" aria-label="HunyuanVideo-13B qualitative video gallery"><div class="gallery-heading"><div><h3>HunyuanVideo-13B</h3><p>Text-to-video · 720p</p></div></div><div class="gallery-stage">
48<button class="gallery-arrow previous" aria-label="Previous HunyuanVideo-13B videos">←</button><div class="gallery-grid"><article class="gallery-video" tabindex="0" aria-label="VBench video. Hover or focus to read the prompt."><video muted="" loop="" playsInline="" preload="metadata" poster="/SparsePR-website/media/thumbnails/hunyuan/0025_hunyuan_t2v_025_3bc182d35d.jpg"><source src="/SparsePR-website/media/gallery/hunyuan/0025_hunyuan_t2v_025_3bc182d35d.mp4" type="video/mp4"/></video><div class="prompt-overlay"><span>Prompt</span><p>Deep in the dense forest, a lighthouse stands alone, its light at the top flickering on and off. The entire video presents a suspenseful atmosphere.</p></div><div class="clip-meta"><span>VBench</span><strong>33.99 dB</strong></div></article><article class="gallery-video" tabindex="0" aria-label="VBench video. Hover or focus to read the prompt."><video muted="" loop="" playsInline="" preload="metadata" poster="/SparsePR-website/media/thumbnails/hunyuan/0000_hunyuan_t2v_000_2a822491e2.jpg"><source src="/SparsePR-website/media/gallery/hunyuan/0000_hunyuan_t2v_000_2a822491e2.mp4" type="video/mp4"/></video><div class="prompt-overlay"><span>Prompt</span><p>Two dolphins are swimming in the blue sea.</p></div><div class="clip-meta"><span>VBench</span><strong>33.70 dB</strong></div></article><article class="gallery-video" tabindex="0" aria-label="VBench video. Hover or focus to read the prompt."><video muted="" loop="" playsInline="" preload="metadata" poster="/SparsePR-website/media/thumbnails/hunyuan/0020_hunyuan_t2v_020_d45099d39c.jpg"><source src="/SparsePR-website/media/gallery/hunyuan/0020_hunyuan_t2v_020_d45099d39c.mp4" type="video/mp4"/></video><div class="prompt-overlay"><span>Prompt</span><p>In the laboratory, Thomas Edison is using a pen to record his experimental data.</p></div><div class="clip-meta"><span>VBench</span><strong>32.91 dB</strong></div></article></div><button class="gallery-arrow next" aria-label="Next HunyuanVideo-13B videos">→</button></div><div class="gallery-position" aria-live="polite">01<!-- --> / <!-- -->12</div></section><section class="model-gallery" aria-label="Wan2.2-I2V-A14B qualitative video gallery"><div class="gallery-heading"><div><h3>Wan2.2-I2V-A14B</h3><p>Image-to-video · native aspect ratio</p></div></div><div class="gallery-stage"><button class="gallery-arrow previous" aria-label="Previous Wan2.2-I2V-A14B videos">←</button><div class="gallery-grid"><article class="gallery-video" tabindex="0" aria-label="VBench video. Hover or focus to read the prompt."><video muted="" loop="" playsInline="" preload="metadata" poster="/SparsePR-website/media/thumbnails/wan22/0000_wan22_i2v_000_a_blue_and_white_smoke_is_swirly_in_the_dark.jpg"><source src="/SparsePR-website/media/gallery/wan22/0000_wan22_i2v_000_a_blue_and_white_smoke_is_swirly_in_the_dark.mp4" type="video/mp4"/></video><div class="prompt-overlay"><span>Prompt</span><p>A blue and white smoke is swirly in the dark.</p></div><div class="clip-meta"><span>VBench</span><strong>28.73 dB</strong></div></article><article class="gallery-video" tabindex="0" aria-label="VBench video. Hover or focus to read the prompt."><video muted="" loop="" playsInline="" preload="metadata" poster="/SparsePR-website/media/thumbnails/wan22/0001_wan22_i2v_001_an_aerial_view_of_a_small_town_on_the_edge_of_the_ocean.jpg"><source src="/SparsePR-website/media/gallery/wan22/0001_wan22_i2v_001_an_aerial_view_of_a_small_town_on_the_edge_of_the_ocean.mp4" type="video/mp4"/></video><div class="prompt-overlay"><span>Prompt</span><p>An aerial view of a small town on the edge of the ocean.</p></div><div class="clip-meta"><span>VBench</span><strong>27.28 dB</strong></div></article><article class="gallery-video" tabindex="0" aria-label="VBench video. Hover or focus to read the prompt."><video muted="" loop="" playsInline="" preload="metadata" poster="/SparsePR-website/media/thumbnails/wan22/0002_wan22_i2v_002_Colorful_buildings_on_the_seaside_cliffs.jpg"><source src="/SparsePR-website/media/gallery/wan22/0002_wan22_i2v_002_Colorful_buildings_on_the_seaside_cliffs.mp4" type="video/mp4"/></video><div class="prompt-overlay"><span>Prompt</span><p>Colorful buildings on the seaside cliffs.</p></div><div class="clip-meta"><span>VBench</span><strong>28.89 dB</strong></div></article></div><button class="gallery-arrow next" aria-label="Next Wan2.2-I2V-A14B videos">→</button></div>
48<div class="gallery-position" aria-live="polite">01<!-- --> / <!-- -->10</div></section></div></section><section class="metrics-band" aria-label="Headline results"><div><strong>21.9–26.0%</strong><span>executed-pair density</span></div><div><strong>1.48–2.61×</strong><span>end-to-end speedup</span></div><div><strong>4</strong><span>models</span></div><div><strong>0</strong><span>training steps</span></div></section><section class="content prose-section" id="abstract"><p class="section-number">Abstract</p><h2>Executable sparsity needs more than concentrated attention.</h2><div class="abstract-grid"><p>Training-free block-sparse attention can accelerate video transformers, but attention concentration alone does not define an executable sparse operator. Shared query routes can expand support, and retained attention mass does not predict the output error from skipped interactions.</p><p><strong>SparsePR</strong> combines Response-Coupled Partitioning with Probe-Fitted Residual Reconstruction. It groups queries and paired K/V tokens by current-call responses, then uses a small set of exact query rows to correct the sparse output. Across four models, SparsePR preserves quality at 22% to 26% executed-pair density with 1.48× to 2.61× end-to-end speedups.</p></div></section><section class="wide-section insight-section" id="method"><div class="section-intro"><p class="section-number">Why partition geometry matters</p><h2>Why per-query sparsity is not executable sparsity.</h2><p><strong>Support density</strong> is the percentage of key tokens needed to retain 90% of attention mass. The pooled diagnostic measures their union across queries before block routing.</p></div><div class="support-lab"><div class="support-control"><label for="group-size"><span>Queries sharing one executable route</span><strong>8</strong></label><input id="group-size" type="range" min="1" max="8" step="1" value="8"/><div class="range-endpoints"><span>1 query · 6.2%</span><span>8 queries · 22.9%</span></div></div><div class="support-comparison"><article class="support-panel single-panel"><div class="support-panel-heading"><div><span>Single query</span><h3>Per-query support</h3></div><strong>6.2%</strong></div><p>One query needs a small key support.</p><div class="support-rows"><div class="support-row" aria-label="Q1: 3 highlighted K/V blocks"><span>Q1</span><div class="support-blocks" aria-hidden="true"><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i></div></div></div></article><article class="support-panel pooled-panel"><div class="support-panel-heading"><div><span>Shared route</span><h3>8<!-- --> <!-- -->queries<!-- --> pooled</h3></div><strong>22.9%</strong></div><p>Each query is sparse; a shared route must cover their union.</p><div class="support-rows"><div class="support-row" aria-label="Q1: 3 highlighted K/V blocks"><span>Q1</span><div class="support-blocks" aria-hidden="true"><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i></div></div><div class="support-row" aria-label="Q2: 3 highlighted K/V blocks"><span>Q2</span><div class="support-blocks" aria-hidden="true"><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i></div></div><div class="support-row" aria-label="Q3: 3 highlighted K/V blocks"><span>Q3</span><div class="support-blocks" aria-hidden="true"><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i></div></div><div class="support-row" aria-label="Q4: 3 highlighted K/V blocks"><span>Q4</span><div class="support-blocks" aria-hidden="true"><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i></div></div><div class="support-row" aria-label="Q5: 3 highlighted K/V blocks"><span>Q5</span><div class="support-blocks" aria-hidden="true"><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i></div></div><div class="support-row" aria-label="Q6: 3 highlighted K/V blocks"><span>Q6</span><div class="support-blocks" aria-hidden="true"><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i></div></div><div class="support-row" aria-label="Q7: 3 highlighted K/V blocks"><span>Q7</span><div class="support-blocks" aria-hidden="true"><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i></div></div><div class="support-row" aria-label="Q8: 3 highlighted K/V blocks"><span>Q8</span><div class="support-blocks" aria-hidden="true"><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i></div></div><div class="support-row union" aria-label="Route: 11 highlighted K/V blocks"><span>Route</span><div class="support-blocks" aria-hidden="true"><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class=""></i><i class=""></i><i class=""></i><i class="active"></i><i class=""></i><i class="active"></i><i class=""></i><i class="active"></i><i class="active"></i></div></div></div></article></div><div class="support-legend"><span><i></i>Omitted key support</span><span><i class="active"></i>Retained key support</span><span><i class="union"></i>Pooled support</span></div><aside class="method-callout"><strong>SparsePR</strong><p>groups queries with similar response patterns, increasing support overlap before constructing the shared route.</p></aside></div><figure class="paper-figure structural-figure"><img src="/SparsePR-website/media/figures/structural-observations-final.png" alt="Per-query sparsity versus pooled support density across four models and Cosmos3-Nano output error at matched retained attention mass"/><figcaption>Per-query concentration does not determine pooled support, and nearly identical retained attention mass can yield substantially different output errors.</figcaption></figure></section><section class="method-section"><div class="content"><p class="section-number light">Method</p><h2>One response geometry.<br/>Two coupled stages.</h2><p class="method-lead">SparsePR constructs the executable partition and reconstructs its error using features from the same current attention call.</p><div class="pipeline" aria-label="SparsePR method pipeline"><div class="pipe-card"><span>01</span><strong>Sample responses</strong><p>Probe a compact subset of queries against K/V.</p></div><div class="pipe-arrow">→</div><div class="pipe-card accent"><span>02</span><strong>Partition support</strong><p>Form paired K/V groups, then response-aligned query groups.</p></div><div class="pipe-arrow">→</div><div class="pipe-card"><span>03</span><strong>Execute sparse routes</strong><p>Select and evaluate hardware-ready query–K/V cells.</p></div><div class="pipe-arrow">→</div><div class="pipe-card accent2"><span>04</span><strong>Reconstruct residual</strong><p>Fit a call-specific correction from exact probe rows.</p></div></div><div class="method-columns"><article><span class="method-tag">RCP</span><h3>Response-Coupled Partitioning</h3><p>Sampled query responses define value-aware paired K/V groups. Their centroids become query-response coordinates, aligning queries that can efficie
48ntly share one route.</p></article><article><span class="method-tag">PFRR</span><h3>Probe-Fitted Residual Reconstruction</h3><p>A small stratified set of query rows is evaluated exactly. Their observed post-softmax residuals fit an affine correction in a low-rank probe-residual subspace.</p></article></div></div></section><section class="wide-section results-section" id="results"><div class="section-intro"><p class="section-number">Results</p><h2>Quality and efficiency across four video models.</h2><p>All quantitative results from the final draft. Values use matched hardware and sequence shapes where reproduced. Density includes routing and exact probe pairs.</p></div><article class="results-block"><div class="results-subhead"><p>Table 1</p><div><h3>Quality and efficiency</h3><span>Reference fidelity, task quality, executed-pair density, attention PFLOPs, and end-to-end speedup.</span></div></div><div class="results-table-wrap"><table class="results-data-table wide"><thead><tr><th>Model</th><th>Method</th><th>PSNR ↑</th><th>SSIM ↑</th><th>LPIPS ↓</th><th>ImgQual ↑</th><th>SubCons ↑</th><th>PBench ↑</th><th>Density ↓</th><th>PFLOPs ↓</th><th>E2E ↑</th></tr></thead><tbody><tr class="group-start"><td><strong>HunyuanVideo-13B</strong></td><td><strong>Dense</strong></td><td>–</td><td>–</td><td>–</td><td>0.850</td><td>0.976</td><td>–</td><td>100.0%</td><td>612.38</td><td>1.00×</td></tr><tr class=""><td><strong>HunyuanVideo-13B</strong></td><td><strong>SpargeAttn†</strong></td><td>25.31</td><td>0.832</td><td>0.217</td><td>0.763</td><td>0.928</td><td>–</td><td>40.19%</td><td>389.76</td><td>1.38×</td></tr><tr class=""><td><strong>HunyuanVideo-13B</strong></td><td><strong>SVG2†</strong></td><td>29.87</td><td>0.907</td><td>0.121</td><td>0.850</td><td>0.927</td><td>–</td><td>25.45%</td><td>299.02</td><td>2.30×</td></tr><tr class=""><td><strong>HunyuanVideo-13B</strong></td><td><strong>SVOO†</strong></td><td>24.87</td><td>0.843</td><td>0.224</td><td>0.679</td><td>0.976</td><td>–</td><td>33.26%</td><td>345.81</td><td>2.17×</td></tr><tr class=""><td><strong>HunyuanVideo-13B</strong></td><td><strong>SVG-EAR†</strong></td><td>30.54</td><td>0.918</td><td>0.098</td><td>0.845</td><td>0.903</td><td>–</td><td>22.17%</td><td>281.86</td><td>1.93×</td></tr><tr class="highlight"><td><strong>HunyuanVideo-13B</strong></td><td><strong>SparsePR</strong></td><td>31.84</td><td>0.932</td><td>0.087</td><td>0.850</td><td>0.976</td><td>–</td><td>21.92%</td><td>255.95</td><td>2.61×</td></tr><tr class="group-start"><td><strong>Wan2.2-I2V-A14B</strong></td><td><strong>Dense</strong></td><td>–</td><td>–</td><td>–</td><td>0.689</td><td>0.974</td><td>–</td><td>100.0%</td><td>658.46</td><td>
481.00×</td></tr><tr class=""><td><strong>Wan2.2-I2V-A14B</strong></td><td><strong>SpargeAttn†</strong></td><td>26.74</td><td>0.871</td><td>0.116</td><td>0.680</td><td>0.953</td><td>–</td><td>30.15%</td><td>396.83</td><td>1.58×</td></tr><tr class=""><td><strong>Wan2.2-I2V-A14B</strong></td><td><strong>SVG2†</strong></td><td>28.18</td><td>0.878</td><td>0.105</td><td>0.668</td><td>0.970</td><td>–</td><td>31.28%</td><td>393.95</td><td>1.59×</td></tr><tr class=""><td><strong>Wan2.2-I2V-A14B</strong></td><td><strong>SVOO†</strong></td><td>29.67</td><td>0.913</td><td>0.095</td><td>0.689</td><td>0.973</td><td>–</td><td>31.67%</td><td>389.62</td><td>1.61×</td></tr><tr class=""><td><strong>Wan2.2-I2V-A14B</strong></td><td><strong>SVG-EAR†</strong></td><td>29.75</td><td>0.921</td><td>0.086</td><td>0.680</td><td>0.970</td><td>–</td><td>23.64%</td><td>378.88</td><td>1.61×</td></tr><tr class="highlight"><td><strong>Wan2.2-I2V-A14B</strong></td><td><strong>SparsePR</strong></td><td>30.66</td><td>0.907</td><td>0.044</td><td>0.687</td><td>0.973</td><td>–</td><td>21.97%</td><td>336.55</td><td>1.80×</td></tr><tr class="group-start"><td><strong>Cosmos-Predict2.5-14B</strong></td><td><strong>Dense</strong></td><td>–</td><td>–</td><td>–</td><td>0.714</td><td>0.976</td><td>77.76</td><td>100.0%</td><td>526.87</td><td>1.00×</td></tr><tr class=""><td><strong>Cosmos-Predict2.5-14B</strong></td><td><strong>SVG2</strong></td><td>20.07</td><td>0.624</td><td>0.330</td><td>0.678</td><td>0.896</td><td>76.14</td><td>28.81%</td><td>286.51</td><td>1.24×</td></tr><tr class=""><td><strong>Cosmos-Predict2.5-14B</strong></td><td><strong>SVOO</strong></td><td>22.06</td><td>0.685</td><td>0.289</td><td>0.701</td><td>0.909</td><td>76.03</td><td>37.63%</td><td>315.38</td><td>1.03×</td></tr><tr class=""><td><strong>Cosmos-Predict2.5-14B</strong></td><td><strong>SVG-EAR</strong></td><td>25.54</td><td>0.908</td><td>0.062</td><td>0.710</td><td>0.976</td><td>77.76</td><td>29.75%</td><td>289.69</td><td>1.10×</td></tr><tr class="highlight"><td><strong>Cosmos-Predict2.5-14B</strong></td><td><strong>SparsePR</strong></td><td>26.33</td><td>0.942</td><td>0.068</td><td>0.714</td><td>0.976</td><td>77.75</td><td>22.14%</td><td>253.61</td><td>1.51×</td></tr><tr class="group-start"><td><strong>Cosmos3-Nano-16B</strong></td><td><strong>Dense</strong></td><td>–</td><td>–</td><td>–</td><td>0.700</td><td>0.950</td><td>77.31</td><td>100.0%</td><td>127.61</td><td>1.00×</td></tr><tr class=""><td><strong>Cosmos3-Nano-16B</strong></td><td><strong>SVG2</strong></td><td>22.45</td><td>0.735</td><td>0.216</td><td>0.677</td><td>0.915</td><td>75.03</td><td>37.29%</td><td>96.19</td><td>1.16×</td></tr><tr class=""><td><strong>Cosmos3-Nano-16B</strong></td><td><strong>SVOO</strong></td><td>16.64</td><td>0.573</td><td>0.381</td><td>0.707</td><td>0.962</td><td>77.59</td><td>67.32%</td><td>108.01</td><td>1.02×</td></tr><tr class=""><td><strong>Cosmos3-Nano-16B</strong></td><td><strong>SVG-EAR</strong></td><td>21.16</td><td>0.709</td><td>0.261</td><td>0.658</td><td>0.872</td><td>72.85</td><td>37.18%</td><td>96.14</td><td>1.10×</td></tr><tr class="highlight"><td><strong>Cosmos3-Nano-16B</strong></td><td><strong>SparsePR</strong></td><td>24.42</td><td>0.801</td><td>0.176</td><td>0.699</td><td>0.949</td><td>77.30</td><td>25.96%</td><td>81.79</td><td>1.48×</td></tr></tbody></table></div></article><figure class="paper-figure results-figure"><img src="/SparsePR-website/media/figures/results-final-pdf.png" alt="Error reduction from response-coupled partitioning across four models and Wan2.2 full-generation latency breakdown"/><figcaption>Response-coupled partitioning reduces mean error by 38.7% to 71.4% and p99 error by 33.1% to 60.6% at 22% density. On Wan2.2, SparsePR reaches 1.80× speedup and probe repair uses 1.1% of total latency.</figcaption></figure><article class="results-block"><div class="results-subhead"><p>Table 2</p><div><h3>Partitioning and reconstruction ablations</h3><span>Mean / p99 normalized attention output error at 22% total executed-pair density. Lower is better.</span></div></div><div class="results-table-wrap"><table class="results-data-table"><thead><tr><th>
48Configuration</th><th>HunyuanVideo</th><th>Wan2.2</th><th>Cosmos-Predict2.5</th><th>Cosmos3-Nano</th></tr></thead><tbody><tr class=""><td><strong>Semantic partition</strong></td><td>0.089 / 0.714</td><td>0.164 / 1.734</td><td>0.791 / 7.532</td><td>0.359 / 3.356</td></tr><tr class=""><td><strong>Key-response K/V partition</strong></td><td>0.085 / 0.812</td><td>0.156 / 1.659</td><td>0.769 / 7.270</td><td>0.341 / 3.174</td></tr><tr class=""><td><strong>Response-coupled partition</strong></td><td>0.073 / 0.696</td><td>0.108 / 0.647</td><td>0.761 / 7.227</td><td>0.331 / 3.150</td></tr><tr class=""><td><strong>Semantic + probe repair</strong></td><td>0.053 / 0.356</td><td>0.105 / 0.819</td><td>0.262 / 0.826</td><td>0.124 / 0.668</td></tr><tr class="highlight"><td><strong>SparsePR</strong></td><td>0.030 / 0.207</td><td>0.063 / 0.323</td><td>0.075 / 0.489</td><td>0.076 / 0.447</td></tr></tbody></table></div></article></section><section class="citation-section" id="citation"><div class="content citation-grid"><div><p class="section-number light">Citation</p><h2>Build on SparsePR.</h2></div><div class="bibtex"><button>Copy BibTeX</button><pre>@article{taghavi2026partition,
49  title={Partition the Support, Reconstruct the Residual: Training-Free Sparse Attention for Video Generation and World Models},
50  author={Taghavi, Pardis and Langari, Reza and Pandey, Gaurav},
51  journal={arXiv preprint arXiv:2608.18484},
52  year={2026}
53}</pre></div></div></section><footer class="content footer"><div><strong>SparsePR</strong><span>Partition the support. Reconstruct the residual.</span></div><p><a href="/SparsePR-website/method/">Method</a> · <a href="/SparsePR-website/results/">Results</a> · <a href="/SparsePR-website/videos/">Videos</a> · <a href="https://github.com/PardisTaghavi/SparsePR">Code</a><br/>Website structure adapted from <a href="https://nerfies.github.io/">Nerfies</a> under CC BY-SA 4.0. © 2026 Pardis Taghavi, Reza Langari, and Gaurav Pandey.</p></footer></main></div>
54  </body>
55</html>

Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.