[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-seo-328152-105":3,"doc-detail-328152-en":81,"detail-sidebar-cat-0-en-105":99},{"code":4,"msg":5,"data":6},0,"ok",{"site_id":7,"language":8,"slug":9,"title":10,"keywords":11,"description":12,"schema_data":13,"social_meta":74,"head_meta":76,"extra_data":78,"updated_unix":80},105,"en","dynamic-in-few-step-unifying-dynamic-computation-and-few-step-distillation-for-efficient-video-generation","Dynamic-in-Few-Step - Unifying Dynamic Computation and Few-Step Distillation for Efficient Video Generation","","Video Diffusion Models (VDMs) deliver high-quality generation but incur prohibitive inference costs. Existing few-step distillation accelerates inference yet typically keeps a static network architecture across denoising stages, overlooking that different noise levels require different computation. This work proposes a post-training acceleration framework that injects dynamic structured sparsification into distillation. It jointly optimizes denoising steps and step-specific structured Mixture-of-Models, stabilizing training via Progressive Training and Output Rollout, and uses a specialized inference engine for efficient deployment. ",{"@graph":14,"@context":73},[15,34,56],{"@type":16,"itemListElement":17},"BreadcrumbList",[18,23,27,31],{"item":19,"name":20,"@type":21,"position":22},"https://docshare.wps.com","Home","ListItem",1,{"item":24,"name":25,"@type":21,"position":26},"https://docshare.wps.com/document/","Document",2,{"item":28,"name":29,"@type":21,"position":30},"https://docshare.wps.com/document/research-report/","Research & Report",3,{"item":32,"name":10,"@type":21,"position":33},"https://docshare.wps.com/document/dynamic-in-few-step-unifying-dynamic-computation-and-few-step-distillation-for-efficient-video-generation/328152/",4,{"url":32,"name":10,"@type":35,"image":36,"author":41,"headline":10,"publisher":44,"fileFormat":47,"inLanguage":8,"description":12,"dateModified":48,"datePublished":49,"encodingFormat":47,"isAccessibleForFree":50,"interactionStatistic":51},"DigitalDocument",{"url":37,"@type":38,"width":39,"height":40},"https://docshare.wps.com/thumbnails/dynamic-in-few-step-unifying-dynamic-computation-and-few-step-distillation-for-efficient-video-generation/328152.png","ImageObject",300,407,{"name":42,"@type":43},"Theodora","Person",{"url":19,"name":45,"@type":46},"DocShare","Organization","application/pdf","2026-11-10","2026-09-21",true,{"@type":52,"interactionType":53,"userInteractionCount":55},"InteractionCounter",{"@type":54},"ViewAction",15,{"@type":57,"mainEntity":58},"FAQPage",[59,65,69],{"name":60,"@type":61,"acceptedAnswer":62},"What problem do existing few-step distillation methods fail to address in video diffusion models?","Question",{"text":63,"@type":64},"They usually enforce a static model architecture across all denoising stages, ignoring that computational needs vary across different noise levels during diffusion.","Answer",{"name":66,"@type":61,"acceptedAnswer":67},"How does the proposed method turn a pre-trained VDM into an efficient model for few-step inference?",{"text":68,"@type":64},"It integrates dynamic structural sparsification directly into the distillation process, jointly optimizing denoising steps and structured model sparsity to form a step-specific Mixture-of-Models (MoM).",{"name":70,"@type":61,"acceptedAnswer":71},"Why is training unstable when optimizing step reduction and structured sparsity jointly, and how is it addressed?",{"text":72,"@type":64},"Joint optimization can cause gradient conflicts and training instability. The paper introduces a Progressive Training Strategy with an Output Rollout mechanism to stabilize coherent learning across timesteps.","https://schema.org",{"og:url":32,"og:type":75,"og:title":10,"og:site_name":45,"og:description":12},"article",{"robots":77,"canonical":32},"index,follow",{"doc_id":79,"site_id":7},328152,1790466676,{"code":4,"msg":82,"data":83},"success",{"doc_id":79,"user_id":84,"nickname":42,"user_avatar":85,"doc_module":4,"category_id":86,"category_name":29,"doc_title":10,"doc_description":12,"doc_content":87,"file_id":88,"file_url":89,"file_type":90,"file_size":91,"view_count":55,"is_deleted":4,"is_public":22,"is_downloadable":22,"audit_status":22,"page_count":92,"language":93,"language_code":8,"site_id":7,"html_lang":8,"table_of_contents":94,"faqs":95,"seo_title":96,"seo_description":12,"update_tm":97,"read_time":98,"preview_page_count":4},687197207919,"https://ap-avatar.wpscdn.com/avatar/a000253d6f5f7c60be?x-image-process=image/resize,m_fixed,w_180,h_180&k=1779446848396160552",8,"arXiv :2607 .0663 1v 1 [ cs .CV] 7 Jul 2026  \nDynamic-in-Few-Step: Unifying Dynamic Computation and Few-Step Distillation for Efficient  \nVideo Generation  \nYu Cheng 1 ,2 ,3 Siyue Yao3 ,4 Zhongang Qi3 ,† Shanyan Guan3 Wei Li3 Fajie Yuan2 ,†  \n1Zhejiang University 2Westlake University 3BlueImage, vivo 4Xian Jiaotong-Liverpool University  \n[chengyu@westlake.edu.cn](chengyu@westlake.edu.cn) , [qizhongang@vivo.com](qizhongang@vivo.com) , [yuanfajie@westlake.edu.cn](yuanfajie@westlake.edu.cn)  \n†Corresponding authors.  \nAbstract  \nVideo Diffusion Models (VDMs) have demonstrated superior generation quality but suffer from prohibitive computational costs. While recent few-step distillation techniques significantly accelerate inference, they typically enforce a static model architecture across all denoising stages, ignoring the varying computational demands inherent to different noise levels. In this work, we propose a novel post-training acceleration framework that exploits this redundancy by integrating dynamic structural sparsification directly into the distillation process. Unlike conventional post-hoc compression applied to a fixed diffusion pipeline, our approach jointly optimizes the denoising steps and structured model sparsity, transforming a pre-trained VDM into a compact, step-specific Mixture-of-Models (MoM) . To address the training instability arising from this joint optimization, we introduce a Progressive Training Strategy coupled with an Output Rollout Mechanism, which ensures the coherent learning of structural decisions across timesteps. Furthermore, we develop a specialized inference engine to deploy the resulting MoM efficiently.  \nOur method is orthogonal to existing acceleration techniques and highly effective: On Wan-14B, it removes 24% of the per-step FLOPs on top of 4-step distillation, adding a 1.2× wall-clock gain and reaching a 30× speedup over the 50-step teacher while preserving competitive generation quality.  \nKeywords: Video Generation Speedup · Dynamic Network · Efficient Diffusion  \n1 Introduction  \nVideo Diffusion Models (VDMs) [30, 10, 25, 17] have revolutionized video creation, enabling breakthroughs in simulation, advertising, and film production. However, their practical deployment is severely constrained by high inference latency and substantial computational resource requirements. Recent advancements in post-training acceleration have been significant, focusing on step reduction [43, 27, 19], attention optimization [40, 37], quantization [15], token compression [2], and streamlined generation [36, 13] . Despite notable progress, these methods share a common limitation: they treat the denoising network as a static architecture, applying the same heavy computation uniformly across all timesteps. However, the diffusion process is inherently heterogeneous. Existing works [6, 41] observe that generation proceeds in a coarse-to-fine manner, where the model first constructs global structures and later refines high-frequency details. This suggests that distinct denoising stages require varying computational capacities, raising a pivotal question: Can we exploit this heterogeneity to dynamically allocate computation across timesteps, removing step-specific redundancy to accelerate VDMs?  \nPreprint.  \nWhile prior works [41, 42] have explored dynamic computation, they are typically trained under standard denoising objectives and often become unstable or ineffective when adapted to few-step distillation [41] . Given that few-step distillation has become the de facto paradigm for efficient VDMs, this incompatibility severely restricts their utility. Moreover, straightforward decoupled pipelines—either “prune-then-distill” or “distill-then-prune”—are insufficient to resolve this mismatch. Specifically, pruning before distillation limits the model capacity required to learn the complex few-step mapping, while pruning after distillation disrupts the carefully aligned generation trajectory, leading to severe ","cbCairAa3Ed29WlA","https://ap.wps.com/l/cbCairAa3Ed29WlA","pdf",6216299,19,"English","# Abstract\n# 1 Introduction\n## Video Diffusion Models and the heterogeneity of denoising stages\n## Limitation of prior acceleration and few-step distillation\n## Unified optimization for dynamic pruning and few-step distillation\n## Progressive Training and Output Rollout\n## Contributions","[{\"question\":\"What problem do existing few-step distillation methods fail to address in video diffusion models?\",\"answer\":\"They usually enforce a static model architecture across all denoising stages, ignoring that computational needs vary across different noise levels during diffusion.\"},{\"question\":\"How does the proposed method turn a pre-trained VDM into an efficient model for few-step inference?\",\"answer\":\"It integrates dynamic structural sparsification directly into the distillation process, jointly optimizing denoising steps and structured model sparsity to form a step-specific Mixture-of-Models (MoM).\"},{\"question\":\"Why is training unstable when optimizing step reduction and structured sparsity jointly, and how is it addressed?\",\"answer\":\"Joint optimization can cause gradient conflicts and training instability. The paper introduces a Progressive Training Strategy with an Output Rollout mechanism to stabilize coherent learning across timesteps.\"}]","Dynamic-in-Few-Step - Unifying Dynamic Computation and Few-Step Distillation for Efficient Video Generation | PDF",1789978745,48,{"code":4,"msg":82,"data":100},[101,105,109,113,118,123,128,131,136,139,143],{"id":22,"doc_module":4,"doc_module_name":25,"category_name":102,"show_sort_weight":103,"slug":104},"Story & Novel",90,"story-novel",{"id":26,"doc_module":4,"doc_module_name":25,"category_name":106,"show_sort_weight":107,"slug":108},"Literature",80,"literature",{"id":33,"doc_module":4,"doc_module_name":25,"category_name":110,"show_sort_weight":111,"slug":112},"Exam",70,"exam",{"id":114,"doc_module":4,"doc_module_name":25,"category_name":115,"show_sort_weight":116,"slug":117},5,"Comic",60,"comic",{"id":119,"doc_module":4,"doc_module_name":25,"category_name":120,"show_sort_weight":121,"slug":122},6,"Technology",50,"technology",{"id":124,"doc_module":4,"doc_module_name":25,"category_name":125,"show_sort_weight":126,"slug":127},7,"Healthcare",40,"healthcare",{"id":86,"doc_module":4,"doc_module_name":25,"category_name":29,"show_sort_weight":129,"slug":130},30,"research-report",{"id":132,"doc_module":4,"doc_module_name":25,"category_name":133,"show_sort_weight":134,"slug":135},9,"Religion & Spirituality",20,"religion-spirituality",{"id":134,"doc_module":4,"doc_module_name":25,"category_name":137,"show_sort_weight":134,"slug":138},"World Cup","world-cup",{"id":140,"doc_module":4,"doc_module_name":25,"category_name":141,"show_sort_weight":140,"slug":142},10,"Lifestyle","lifestyle",{"id":92,"doc_module":4,"doc_module_name":25,"category_name":144,"show_sort_weight":114,"slug":145},"General","general"]