[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-seo-328150-105":3,"doc-detail-328150-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","d2po-optimizing-diffusion-samplers-via-dynamic-preference","D2PO - Optimizing Diffusion Samplers via Dynamic Preference","","D2PO (Dynamic Direct Preference Optimization) proposes a principled framework for improving diffusion sampling policies by optimizing timestep schedules and classifier-free guidance (CFG) weights. It targets a key limitation of student-teacher regression: low-NFE students trained to mimic high-NFE teachers often preserve coarse structure while losing high-frequency texture fidelity, reducing perceptual quality under large acceleration gaps. D2PO reframes sampler learning as a preference-based alignment problem using DPO-style energy modeling and introduces dynamic preferences that self-improve as policies learn. Extensive experiments show stronger perceptual alignment and consistent gains over regression schedulers in low-NFE settings.",{"@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/d2po-optimizing-diffusion-samplers-via-dynamic-preference/328150/",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/d2po-optimizing-diffusion-samplers-via-dynamic-preference/328150.png","ImageObject",300,407,{"name":42,"@type":43},"Genevieve","Person",{"url":19,"name":45,"@type":46},"DocShare","Organization","application/pdf","2026-11-04","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 limitation of student-teacher regression does D2PO address?","Question",{"text":63,"@type":64},"D2PO addresses the limitation where low-NFE student samplers regress to high-NFE teachers, often sacrificing high-frequency texture fidelity and perceptual quality while maintaining coarse global structures.","Answer",{"name":66,"@type":61,"acceptedAnswer":67},"How does D2PO make preference optimization applicable to diffusion samplers?",{"text":68,"@type":64},"It models the sampling policy as an energy-based model (EBM), reformulating preference comparisons into tractable energy differences for diffusion sampling.",{"name":70,"@type":61,"acceptedAnswer":71},"What are dynamic preferences in D2PO?",{"text":72,"@type":64},"Dynamic preferences mean the preferred samples used for alignment improve progressively as the sampling policies are learned, replacing rigid static teacher supervision with iterative refinement.","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},328150,1790496107,{"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},1374391974585,"https://ap-avatar.wpscdn.com/davatar_276721f389ce27ea32af1340a28f341c",8,"arXiv :2607 .06609v 1 [ cs .LG] 7 Jul 2026  \nD2PO: Optimizing Diffusion Samplers via Dynamic Preference  \nJinkyu Kim∗1, Jinyoung Choi∗3, and Bohyung Han 1 ,2  \n1 ECE & 2 IPAI, Seoul National University, Korea  \n3 AIGS, Ulsan National Institute of Science and Technology, Korea  \nAbstract. We propose D2PO (Dynamic Direct Preference Optimization), a principled framework for optimizing diffusion sampling policies with respect to timestep schedules and classifier-free guidance (CFG) weights. Our work is motivated by a fundamental limitation of existing student-teacher regression frameworks; low-NFE student samplers are trained to mimic high-NFE teachers, often sacrificing high-frequency texture fidelity while preserving coarse global structures, thereby misaligning the sampler with perceptual quality. D2PO addresses this challenge by reformulating sampler optimization as a preference-based alignment problem, leveraging the Direct Preference Optimization (DPO) framework. To make DPO applicable to diffusion samplers, we model the sampling policy as an energy-based model (EBM), transforming preference comparisons into tractable energy differences. We further introduce a novel energy formulation derived directly from the pretrained score network, enabling preference evaluation in perturbed spaces that jointly capture structural consistency and fine-grained details. Moreover, we introduce dynamic preferences, where the preferred samples used for alignment progressively improve as the sampling policies are learned. This self-improving mechanism replaces rigid static teacher supervision with an iterative, preference-guided refinement process, providing progressively stronger alignment signals. Extensive experiments demonstrate that D2PO aligns diffusion samplers with perceptual quality more faithfully, unlocking the full potential of high-quality teachers and consistently outperforming conventional regression-based schedulers under low-NFE constraints.  \nKeywords: Diffusion models · Time step optimization · Direct preference optimization  \n1 Introduction  \nDiffusion Probabilistic Models (DPMs) [19, 46, 48] have achieved unprecedented fidelity in high-resolution image synthesis, text-to-image generation [10, 40], and video generation [21,45,54,72] . However, this performance comes at a substantial  \n∗ indicates equal contribution.  \n2 J. Kim et al.  \n“Two birds that are sitting in a marsh area.”  \n∆ = 1 ∆ = 2 ∆ = 3 ∆ = 4  \n“There is a small bus with several people standing next to it. \"  \n∆ = 1 ∆ = 2 ∆ = 3 ∆ = 4  \nFig. 1: Qualitative evidence of the performance bottleneck in LD3 [50] . All images are generated by the same model with NFE = 4 . The columns show the impact of increasing the NFE gap (∆ = T − S ) between the teacher (T) and the student (S ) . While a small gap (∆ = 1) yields high-quality outputs, larger gaps (up to ∆ = 4) lead to severe artifacts, demonstrating LD3’s inability to leverage high-fidelity teachers.  \ncomputational cost. DPMs are inherently iterative, requiring many function evaluations (NFE) during sampling, which makes high-quality generation expensive and limits practical deployment.  \nA broad range of approaches has been explored to mitigate this bottleneck, including accelerated numerical solvers [9,11,32,34,35,66,70], few-step knowledge distillation [25,42,43,47,63,64,71,73,74], architectural modifications [36,62], and training-time improvements [23,26,51,59] . More recently, directly optimizing the sampling policy parameters—such as timestep schedules [14, 28, 41, 50, 57, 60], classifier-free guidance weights [15], and high-order solver coefficients [14, 53, 67]—has emerged as a critical direction for acceleration.  \nIn optimizing these sampler parameters, prior works typically adopt either a distribution or an instance perspective. Specifically, one line of work optimizes distribution-level metrics, such as FID [28] or KID [56] over large sample sets, but such population-level objectives yield weak, high-va","cbCailVkPIGlQsXB","https://ap.wps.com/l/cbCailVkPIGlQsXB","pdf",13599301,31,"English","# Introduction\n## Diffusion probabilistic models and sampling cost\n## Prior approaches to sampler acceleration\n## Distribution vs instance optimization perspectives\n## Structural limitation of regression-based student-teacher training\n## Preference-based alignment via D2PO","[{\"question\":\"What limitation of student-teacher regression does D2PO address?\",\"answer\":\"D2PO addresses the limitation where low-NFE student samplers regress to high-NFE teachers, often sacrificing high-frequency texture fidelity and perceptual quality while maintaining coarse global structures.\"},{\"question\":\"How does D2PO make preference optimization applicable to diffusion samplers?\",\"answer\":\"It models the sampling policy as an energy-based model (EBM), reformulating preference comparisons into tractable energy differences for diffusion sampling.\"},{\"question\":\"What are dynamic preferences in D2PO?\",\"answer\":\"Dynamic preferences mean the preferred samples used for alignment improve progressively as the sampling policies are learned, replacing rigid static teacher supervision with iterative refinement.\"}]","D2PO - Optimizing Diffusion Samplers via Dynamic Preference | PDF",1789978744,78,{"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":144,"doc_module":4,"doc_module_name":25,"category_name":145,"show_sort_weight":114,"slug":146},19,"General","general"]