[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-156880-105":59,"doc-detail-156880-en":131},{"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":124,"head_meta":126,"extra_data":128,"updated_unix":130},105,"en","instance-wise-distribution-control-of-text-to-image-diffusion-models","Instance-wise distribution control of text-to-image diffusion models","","Text-to-image diffusion models are widely used to create synthetic datasets for downstream vision tasks, yet they often inherit attribute biases from large-scale web data, producing unbalanced attribute distributions in generated images. Existing distribution-control methods largely target single-object settings and fail to reliably control attributes across multiple object instances within the same scene. This work proposes instance-wise control by fine-tuning diffusion models with guidance from a pre-trained object detector and attribute classifier. The method aligns attribute proportions with a user-defined target distribution at the instance level, yielding high-quality multi-instance images. Experiments across various objects and attributes validate that generated images match the specified distribution, supporting scalable creation of distribution-aware synthetic datasets for in-the-wild vision applications.",{"@graph":69,"@context":123},[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/instance-wise-distribution-control-of-text-to-image-diffusion-models/156880/",{"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/instance-wise-distribution-control-of-text-to-image-diffusion-models/156880.png","ImageObject",300,407,{"name":92,"@type":93},"WPS_1786070896","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-11-03","2026-08-29",true,{"@type":102,"interactionType":103,"userInteractionCount":105},"InteractionCounter",{"@type":104},"ViewAction",14,{"@type":107,"mainEntity":108},"FAQPage",[109,115,119],{"name":110,"@type":111,"acceptedAnswer":112},"Why do text-to-image diffusion models produce unbalanced attribute distributions?","Question",{"text":113,"@type":114},"They inherit attribute biases from large-scale web training data, which typically lacks fairness constraints. As a result, generated images often amplify existing imbalances across attributes.","Answer",{"name":116,"@type":111,"acceptedAnswer":117},"What is the main limitation of prior distribution-control methods?",{"text":118,"@type":114},"Most methods assume single-object images and effectively control attributes only at the image level. They struggle to assign and balance attributes for multiple object instances within the same generated scene.",{"name":120,"@type":111,"acceptedAnswer":121},"How does the proposed instance-wise control method work?",{"text":122,"@type":114},"It fine-tunes diffusion models using guidance from a pre-trained object detector and an attribute classifer. This aligns the attribute distribution across object instances with a user-defined target distribution.","https://schema.org",{"og:url":83,"og:type":125,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":127,"canonical":83},"index,follow",{"doc_id":129,"site_id":62},156880,1787968327,{"code":4,"msg":5,"data":132},{"doc_id":129,"user_id":133,"nickname":92,"user_avatar":134,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":135,"file_id":136,"file_url":137,"file_type":138,"file_size":139,"view_count":105,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":140,"language":141,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":142,"faqs":143,"seo_title":144,"seo_description":67,"update_tm":130,"read_time":145,"preview_page_count":4},549768072016,"https://ap-avatar.wpscdn.com/davatar_155a257f0dc6eb9ab79c44ca47cae57d","| Tit le | Instance-wise d istr ibut ion contro l of text-toimage d iffusion mode ls |\n| --- | --- |\n| Author( s ) | Ian Chan, Weng; Santo, Hiroaki; Matsushita, Yasuyuki et a l. |\n| Citat ion | Pattern Recognit ion. 2025 , 172, p. 112614 |\n| Version Type | VoR |\n| URL | [https://hdl.handle.net/11094/103517](https://hdl.handle.net/11094/103517) |\n| rights | This art ic le is l icensed under a Creat ive\u003Cbr>Commons Attr ibut ion 4.0 Internat iona l License. |\n| Note |  |\n\nThe Un ivers ity of Osaka Inst itut iona l Know ledge Arch ive : OUKA  \n[https://ir. library.osaka-u.ac.jp/](https://ir. library.osaka-u.ac.jp/)  \nThe University of Osaka  \nPattern Recognition 172 (2026) 112614  \n| Instance-wise distribution control of text-to-image diffusion models Weng Ian Chan  ∗, Hiroaki Santo  , Yasuyuki Matsushita  , Fumio Okura  Graduate School of Information Science and Technology, The University of Osaka, Yamadaoka 1–5, Suita, 5650871, Osaka, Japan |  |  | |\n| --- | --- | --- | --- |\n| a r t i c l e i n f o |  | a b s t r a c t |  |\n| Keywords:\u003Cbr>Text-to-image generation Diffusion models Debiasing\u003Cbr>Multi-instance generation |  | Text-to-image diffusion models are increasingly used to generate synthetic datasets for downstream vision tasks. However, they often inherit biases from large-scale training data, which can result in unbalanced attribute distributions in the generated images. While prior efforts have attempted to mitigate these biases, most focus on single-object images and struggle to control attributes across object instances in multi-instance generations. To address this limitation, we propose an instance-wise control of the attribute distribution by fine-tuning diffusion models with guidance from a pre-trained object detector and an attribute classifier. Our approach aligns the attribute distribution over object instances in generated images with a user-defined distribution, which enables precise control over attribute proportions at the instance level. Experiments across various objects and attributes demonstrate that our method generates high-quality, multi-instance images that match the specified distribution, supporting the scalable creation of distribution-aware synthetic datasets for in-the-wild vision tasks. |  |\n\n1. Introduction  \nThe recent progress in text-to-image (T2I) diffusion models [1] has unlocked powerful capabilities for controllable image synthesis across diverse domains. Beyond artistic creation, T2I models have become a promising tool for generating synthetic datasets that support supervised learning tasks, especially when data collection from real-world environments is expensive or impractical [2]. In many visual domains, such as medical image segmentation [3] and scene understanding [4], collecting large-scale in-the-wild datasets is not only labor-intensive and time-consuming, but also prone to inherent attribute imbalance.  \nTo address these challenges, researchers increasingly rely on synthetic data generated by T2I models as a scalable alternative [2]. Nevertheless, current T2I models inherit the attribute biases present in their large-scale training datasets, which are typically crawled from the Internet without any fairness constraints [5]. As a result, generated images often reflect and amplify existing imbalances [6], leading to the underrepresentation of certain attributes, such as gender in stereotypical occupations or Asian and Indian faces, as illustrated in Fig. 1.  \nIn response to such biases, a growing body of work explores distributional control in T2I models, i.e., modifying generation to match a desired attribute distribution across outputs [7–9]. These approaches use inference-time guidance or model modification to balance representations across images. However, they are primarily developed and evaluated in single-object settings, where each image contains only one attribute-bearing instance. Thus, they implicitly assume that controlling  \nattributes at the image level is sufficient and do not","cbCaiceaqLwhofoz","https://ap.wps.com/l/cbCaiceaqLwhofoz","pdf",29114711,15,"English","# Introduction\n## Distributional bias in text-to-image diffusion models\n## Limitations of image-level control in multi-instance scenes\n## Consequences for downstream vision tasks\n## Related approaches and remaining challenges","[{\"question\":\"Why do text-to-image diffusion models produce unbalanced attribute distributions?\",\"answer\":\"They inherit attribute biases from large-scale web training data, which typically lacks fairness constraints. As a result, generated images often amplify existing imbalances across attributes.\"},{\"question\":\"What is the main limitation of prior distribution-control methods?\",\"answer\":\"Most methods assume single-object images and effectively control attributes only at the image level. They struggle to assign and balance attributes for multiple object instances within the same generated scene.\"},{\"question\":\"How does the proposed instance-wise control method work?\",\"answer\":\"It fine-tunes diffusion models using guidance from a pre-trained object detector and an attribute classifer. This aligns the attribute distribution across object instances with a user-defined target distribution.\"}]","Instance-wise distribution control of text-to-image diffusion models | PDF",38]