[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-seo-327782-105":3,"detail-sidebar-cat-0-en-105":81,"doc-detail-327782-en":130},{"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","self-improving-diffusion-classifiers-with-minority-preference-optimization","Self-Improving Diffusion Classifiers with Minority Preference Optimization","","Diffusion classifiers provide robust zero-shot classification, yet accuracy depends heavily on pretraining data distribution. Majority concepts in high-density manifold regions are recognized well, while minority concepts in low-density regions are substantially weaker. The study links minority sampling during generation to the classifier’s perception capability. It introduces MiPO, which fine-tunes a pretrained diffusion model using minority preference rewards from only arbitrary caption data via LoRA and Group Relative Policy Optimization, without extra images, foundation models, or external reward models. Experiments on five datasets show bias reduction and improved zero-shot recognition.",{"@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/self-improving-diffusion-classifiers-with-minority-preference-optimization/327782/",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/self-improving-diffusion-classifiers-with-minority-preference-optimization/327782.png","ImageObject",300,407,{"name":42,"@type":43},"Asher","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",19,{"@type":57,"mainEntity":58},"FAQPage",[59,65,69],{"name":60,"@type":61,"acceptedAnswer":62},"Why do diffusion classifiers perform worse on minority concepts?","Question",{"text":63,"@type":64},"Their recognition depends on generative coverage of the pretrained diffusion model. Noise-matching training is biased toward majority modes, so low-density minority regions are less well generated and therefore less recognized.","Answer",{"name":66,"@type":61,"acceptedAnswer":67},"What is MiPO and how does it improve minority coverage?",{"text":68,"@type":64},"MiPO fine-tunes a pretrained diffusion model using minority preference rewards. It generates candidates from arbitrary caption data, rewards those covering minority regions better, and updates the model to broaden coverage of underrepresented manifold areas.",{"name":70,"@type":61,"acceptedAnswer":71},"Does MiPO require extra image data or external reward models?",{"text":72,"@type":64},"No. MiPO uses only arbitrary caption data and fine-tunes with LoRA plus Group Relative Policy Optimization, without additional image data, external foundation models, or external reward models.","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},327782,1790216440,{"code":4,"msg":82,"data":83},"success",[84,88,92,96,101,106,111,115,120,123,127],{"id":22,"doc_module":4,"doc_module_name":25,"category_name":85,"show_sort_weight":86,"slug":87},"Story & Novel",90,"story-novel",{"id":26,"doc_module":4,"doc_module_name":25,"category_name":89,"show_sort_weight":90,"slug":91},"Literature",80,"literature",{"id":33,"doc_module":4,"doc_module_name":25,"category_name":93,"show_sort_weight":94,"slug":95},"Exam",70,"exam",{"id":97,"doc_module":4,"doc_module_name":25,"category_name":98,"show_sort_weight":99,"slug":100},5,"Comic",60,"comic",{"id":102,"doc_module":4,"doc_module_name":25,"category_name":103,"show_sort_weight":104,"slug":105},6,"Technology",50,"technology",{"id":107,"doc_module":4,"doc_module_name":25,"category_name":108,"show_sort_weight":109,"slug":110},7,"Healthcare",40,"healthcare",{"id":112,"doc_module":4,"doc_module_name":25,"category_name":29,"show_sort_weight":113,"slug":114},8,30,"research-report",{"id":116,"doc_module":4,"doc_module_name":25,"category_name":117,"show_sort_weight":118,"slug":119},9,"Religion & Spirituality",20,"religion-spirituality",{"id":118,"doc_module":4,"doc_module_name":25,"category_name":121,"show_sort_weight":118,"slug":122},"World Cup","world-cup",{"id":124,"doc_module":4,"doc_module_name":25,"category_name":125,"show_sort_weight":124,"slug":126},10,"Lifestyle","lifestyle",{"id":55,"doc_module":4,"doc_module_name":25,"category_name":128,"show_sort_weight":97,"slug":129},"General","general",{"code":4,"msg":82,"data":131},{"doc_id":79,"user_id":132,"nickname":42,"user_avatar":133,"doc_module":4,"category_id":112,"category_name":29,"doc_title":10,"doc_description":12,"doc_content":134,"file_id":135,"file_url":136,"file_type":137,"file_size":138,"view_count":55,"is_deleted":4,"is_public":22,"is_downloadable":22,"audit_status":22,"page_count":55,"language":139,"language_code":8,"site_id":7,"html_lang":8,"table_of_contents":140,"faqs":141,"seo_title":142,"seo_description":12,"update_tm":143,"read_time":144,"preview_page_count":4},687197207639,"https://ap-avatar.wpscdn.com/davatar_a8503ba1806abce46bf441b54a3ca4cd","arXiv :2607 .03770v 1 [ cs .CV] 4 Jul 2026  \nSelf-Improving Diffusion Classifiers with Minority Preference Optimization  \nHyunsoo Kim* 1, Jungmyung Wi* 1, Soobin Um2, Donghyun Kim†1, and  \nSuhyun Kim†3  \n1 Korea University, 2 Kook Min University, 3 Kyung Hee University  \n{climba, wjm333, [d_kim}@korea.ac.kr](d_kim}@korea.ac.kr)  \n[soobin.um@kookmin.ac.kr](soobin.um@kookmin.ac.kr)  \n[dr.suhyun_kim@gmail.com](dr.suhyun_kim@gmail.com)  \nAbstract. Prior studies have demonstrated that diffusion classifiers achieve robust zero-shot classification performance. However, their effectiveness is strongly tied to the pretraining data distribution: they perform well in majority, high-density regions of the data manifold, but are significantly less accurate in minority, low-density regions. Although prior works on minority sampling have focused on generating more minority-like images, what minority sampling fundamentally enables beyond generation remains underexplored. In this paper, we reveal a direct relationship between minority sampling in generation and the perception capability of diffusion classifiers. Specifically, we show that enhancing minority sampling broadens the coverage of underrepresented regions on the data manifold, thereby improving diffusion-based recognition. To exploit this connection, we propose Self-Improving Diffusion Classifiers with Minority Preference Optimization (MiPO), which fine-tunes a pretrained diffusion model using minority preference rewards. Using only arbitrary caption data, MiPO generates candidate samples, rewards those that better cover minority regions, and optimizes the model with LoRA and Group Relative Policy Optimization, without additional image data, external foundation models, or external reward models. This enables stable, prompt-adaptive minority sampling and translates low-density generative coverage into improved zero-shot diffusion classification. To sum up, we show that diffusion classifier perception is biased toward majority regions, demonstrate that this bias can be alleviated through minority preference optimization, and evaluate MiPO on five standard datasets.  \nKeywords: Diffusion Classifier · Minority Sampling · Reinforcement Learning for diffusion model  \n1 Introduction  \nBeyond the astonishing generative performance of diffusion [23, 39, 43] and flow models [16, 33, 35], a new wave of research is emerging that seeks to leverage  \n∗ These authors contributed equally to this work.†Co-corresponding author.  \n2 H. Kim et al.  \nMajority Samples Minority Samples  \nFig. 1: MiPO expands the perceptual coverage of diffusion classifiers. A standard diffusion classifier reliably recognizes majority (i.e., high-density) samples that are well covered by the pretrained generative distribution (e.g. , soap dispenser), but often fails on underrepresented (i.e., low-density) concepts such as lotion, duck, and sink. This illustrates that the recognition ability of diffusion classifiers is closely tied to the generative coverage of the underlying diffusion model. By optimizing minority preference rewards, MiPO encourages the model to better cover such low-density regions, thereby improving the classifier’s perception of minority visual categories without requiring additional downstream image data.  \nthese generative models to solve various understanding tasks. This trend mirrors the evolution in Natural Language Processing (NLP), where the advent and advancement of large language models (LLMs) resolved numerous downstream tasks such as machine translation [1] or sentiment analysis [37] . Similarly, recent works [30, 55] leverage large-scale vision generative model to solve vision perception tasks. These studies explore methods such as extracting useful features [45,46], learned during the generative training process, from the intermediate layers of the models to aid in segmentation or classification, or employing the generative models themselves to solve these tasks directly [30,55] .  \nA promin","cbCaiirJtXTtCTVf","https://ap.wps.com/l/cbCaiirJtXTtCTVf","pdf",2881155,"English","# Abstract\n# 1 Introduction\n## Motivation: generative coverage and minority failure\n## Proposed method: MiPO with minority preference rewards\n## Training approach and optimization\n## Evaluation on standard datasets","[{\"question\":\"Why do diffusion classifiers perform worse on minority concepts?\",\"answer\":\"Their recognition depends on generative coverage of the pretrained diffusion model. Noise-matching training is biased toward majority modes, so low-density minority regions are less well generated and therefore less recognized.\"},{\"question\":\"What is MiPO and how does it improve minority coverage?\",\"answer\":\"MiPO fine-tunes a pretrained diffusion model using minority preference rewards. It generates candidates from arbitrary caption data, rewards those covering minority regions better, and updates the model to broaden coverage of underrepresented manifold areas.\"},{\"question\":\"Does MiPO require extra image data or external reward models?\",\"answer\":\"No. MiPO uses only arbitrary caption data and fine-tunes with LoRA plus Group Relative Policy Optimization, without additional image data, external foundation models, or external reward models.\"}]","Self-Improving Diffusion Classifiers with Minority Preference Optimization | PDF",1789978028,48]