[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-135651-105":59,"doc-detail-135651-en":123},{"code":4,"msg":5,"data":6},0,"success",[7,13,18,23,28,33,38,43,48,51,55],{"id":8,"doc_module":4,"doc_module_name":9,"category_name":10,"show_sort_weight":11,"slug":12},1,"Document","Story & Novel",90,"story-novel",{"id":14,"doc_module":4,"doc_module_name":9,"category_name":15,"show_sort_weight":16,"slug":17},2,"Literature",80,"literature",{"id":19,"doc_module":4,"doc_module_name":9,"category_name":20,"show_sort_weight":21,"slug":22},4,"Exam",70,"exam",{"id":24,"doc_module":4,"doc_module_name":9,"category_name":25,"show_sort_weight":26,"slug":27},5,"Comic",60,"comic",{"id":29,"doc_module":4,"doc_module_name":9,"category_name":30,"show_sort_weight":31,"slug":32},6,"Technology",50,"technology",{"id":34,"doc_module":4,"doc_module_name":9,"category_name":35,"show_sort_weight":36,"slug":37},7,"Healthcare",40,"healthcare",{"id":39,"doc_module":4,"doc_module_name":9,"category_name":40,"show_sort_weight":41,"slug":42},8,"Research & Report",30,"research-report",{"id":44,"doc_module":4,"doc_module_name":9,"category_name":45,"show_sort_weight":46,"slug":47},9,"Religion & Spirituality",20,"religion-spirituality",{"id":46,"doc_module":4,"doc_module_name":9,"category_name":49,"show_sort_weight":46,"slug":50},"World Cup","world-cup",{"id":52,"doc_module":4,"doc_module_name":9,"category_name":53,"show_sort_weight":52,"slug":54},10,"Lifestyle","lifestyle",{"id":56,"doc_module":4,"doc_module_name":9,"category_name":57,"show_sort_weight":24,"slug":58},19,"General","general",{"code":4,"msg":60,"data":61},"ok",{"site_id":62,"language":63,"slug":64,"title":65,"keywords":66,"description":67,"schema_data":68,"social_meta":116,"head_meta":118,"extra_data":120,"updated_unix":122},105,"en","phased-consistency-models-neurips-2024-paper","Phased Consistency Models - NeurIPS 2024 Paper","","Consistency Models accelerate diffusion-model generation, but their latent-space use for high-resolution, text-conditioned image synthesis remains insufficient. This paper diagnoses three design flaws in Latent Consistency Models (LCMs) affecting consistency, controllability, and multi-step efficiency. The work introduces Phased Consistency Models (PCMs), which generalize the design space by enforcing self-consistency separately on phased sub-trajectories. Evaluations show PCMs outperform LCMs across 1–16 steps, match prior 1-step quality, and extend to video generation via training a state-of-the-art few-step text-to-video generator.",{"@graph":69,"@context":115},[70,84,106],{"@type":71,"itemListElement":72},"BreadcrumbList",[73,77,79,82],{"item":74,"name":75,"@type":76,"position":8},"https://docshare.wps.com","Home","ListItem",{"item":78,"name":9,"@type":76,"position":14},"https://docshare.wps.com/document/",{"item":80,"name":40,"@type":76,"position":81},"https://docshare.wps.com/document/research-report/",3,{"item":83,"name":65,"@type":76,"position":19},"https://docshare.wps.com/document/phased-consistency-models-neurips-2024-paper/135651/",{"url":83,"name":65,"@type":85,"image":86,"author":91,"headline":65,"publisher":94,"fileFormat":97,"inLanguage":63,"description":67,"dateModified":98,"datePublished":99,"encodingFormat":97,"isAccessibleForFree":100,"interactionStatistic":101},"DigitalDocument",{"url":87,"@type":88,"width":89,"height":90},"https://docshare.wps.com/thumbnails/phased-consistency-models-neurips-2024-paper/135651.png","ImageObject",300,407,{"name":92,"@type":93},"Chloe Bennett","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-10-16","2026-08-21",true,{"@type":102,"interactionType":103,"userInteractionCount":105},"InteractionCounter",{"@type":104},"ViewAction",12,{"@type":107,"mainEntity":108},"FAQPage",[109],{"name":110,"@type":111,"acceptedAnswer":112},"What benefits do PCMs provide for controllability and few-step efficiency?","Question",{"text":113,"@type":114},"PCMs reduce controllability limitations caused by the CFG-augmented ODE solver used during consistency distillation, and they better preserve sample quality in fewer than four inference steps. The paper attributes LCM few-step weakness to insufficient supervision from traditional losses like L2 or Huber.","Answer","https://schema.org",{"og:url":83,"og:type":117,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":119,"canonical":83},"index,follow",{"doc_id":121,"site_id":62},135651,1787316673,{"code":4,"msg":5,"data":124},{"doc_id":121,"user_id":125,"nickname":92,"user_avatar":126,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":127,"file_id":128,"file_url":129,"file_type":130,"file_size":131,"view_count":105,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":132,"language":133,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":134,"faqs":135,"seo_title":136,"seo_description":67,"update_tm":122,"read_time":137},962084925782,"https://ap-avatar.wpscdn.com/davatar_9964176cb1d06d4a9deccf72a44ae3dc","Phased Consistency Models  \nFu-Yun Wang1 Zhaoyang Huang2 Alexander William Bergman3,6 Dazhong Shen4 Peng Gao4 Michael Lingelbach3,6 Keqiang Sun1 Weikang Bian1 Guanglu Song5 Yu Liu4 Xiaogang Wang1 Hongsheng Li1,4,7  \n1MMLab, CUHK 2Avolution AI 3Hedra  \n4 Shanghai AI Lab 5 Sensetime Research 6 Stanford University 7 CPII under InnoHK {fywang@link, xgwang@ee, [hsli@ee}.cuhk.edu.hk](hsli@ee}.cuhk.edu.hk)  \nAbstract  \nConsistency Models (CMs) have made significant progress in accelerating the generation of diffusion models. However, their application to high-resolution, text-conditioned image generation in the latent space remains unsatisfactory. In this paper, we identify three key flaws in the current design of Latent Consistency Models (LCMs) . We investigate the reasons behind these limitations and propose Phased Consistency Models (PCMs), which generalize the design space and address the identified limitations. Our evaluations demonstrate that PCMs outperform LCMs across 1–16 step generation settings. While PCMs are specifically designed for multi-step refinement, they achieve comparable 1-step generation results to previously state-of-the-art specifically designed 1-step methods. Furthermore, we show the methodology of PCMs is versatile and applicable to video generation, enabling us to train the state-of-the-art few-step text-to-video generator. Our code is available at [https://github.com/G-U-N/Phased-Consistency-Model](https://github.com/G-U-N/Phased-Consistency-Model).  \nFigure 1: PCMs: Towards stable and fast image and video generation.  \n38th Conference on Neural Information Processing Systems (NeurIPS 2024) .  \nFigure 2: Summative motivation. We observe and summarize three crucial limitations for (latent) consistency models, and generalize the design space, well tackling all these limitations.  \n1 Introduction  \nDiffusion models [20, 15, 57, 68] have emerged as the dominant methodology for image synthesis [45, 41, 7] and video synthesis [22, 53, 52, 63] . These models have shown the ability to generate highquality and diverse samples conditioned on varying signals. At their core, diffusion models rely on an iterative evaluation to generate new samples. This iterative evaluation trajectory models the probability flow ODE (PF-ODE) [56, 57] that transforms an initial normal distribution to a target real data distribution. However, the iterative nature of diffusion models makes the generation of new samples time-intensive and resource-consuming. To address this challenge, consistency models [56, 11, 63, 55] have emerged to reduce the number of iterative steps required to generate samples. These models work by training a model that enforces the self-consistency [56] property: any point along the same PF-ODE trajectory shall be mapped to the same solution point. These models have been extended to high-resolution text-to-image synthesis with latent consistency models (LCMs) [35] . Despite the improvements in efficiency and the ability to generate samples in a few steps, the sample quality of such models is still limited.  \nWe show that the current design of LCMs is flawed, causing inevitable drawbacks in controllability, consistency, and efficiency during image sampling. Fig. 2 illustrates our observations of LCMs. The limitations are listed as follows:  \n(1) Consistency. Due to the specific consistency property, CMs can only use the purely stochastic multi-step sampling algorithm, which assumes that the accumulated noise variable in each generative step is independent and causes varying degrees of stochasticity for different inferencestep settings. As a result, we can find inconsistency among the samples generated with the same seeds in different inference steps.  \n(2) Controllability. Even though diffusion models can adopt classifier-free guidance (CFG) [16] in a wide range of values (i.e. 2–15), equipped with weights of LCMs, they can only accept values of CFG within range of 1–2 . Larger values of CFG would cause the expos","cbCaimuGqTFYNwtw","https://ap.wps.com/l/cbCaimuGqTFYNwtw","pdf",47885545,59,"English","# Introduction\n## Background: diffusion and consistency models\n## Identified issues in LCMs\n## Proposed method: PCMs\n# Method Overview\n## Phased self-consistency on sub-trajectories\n## Two-sub-trajectory training example\n# Results and Evaluation\n## Step-wise performance comparison\n## Extension to video generation","[{\"question\":\"What benefits do PCMs provide for controllability and few-step efficiency?\",\"answer\":\"PCMs reduce controllability limitations caused by the CFG-augmented ODE solver used during consistency distillation, and they better preserve sample quality in fewer than four inference steps. The paper attributes LCM few-step weakness to insufficient supervision from traditional losses like L2 or Huber.\"}]","Phased Consistency Models - NeurIPS 2024 Paper | PDF",149]