[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-seo-135643-105":3,"detail-sidebar-cat-0-en-105":81,"doc-detail-135643-en":131},{"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","metapoint-bridging-continuous-2d-spatial-control-in-agentic-visual-generation","MetaPoint - Bridging Continuous 2D Spatial Control in Agentic Visual Generation","","Generative visual models struggle with precise spatial control because they can read spatial descriptions in text but cannot directly map numeric 2D coordinates onto an image canvas. MetaPoint addresses this gap by encoding a continuous 2D coordinate as a single special token, leveraging the model’s existing positional encoding without adding new architecture. With one token or a two-token bounding box, it enables pixel-level object positioning and compositional primitives for planner-style agent decomposition, supporting more capable, interactive editing.",{"@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/metapoint-bridging-continuous-2d-spatial-control-in-agentic-visual-generation/135643/",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/metapoint-bridging-continuous-2d-spatial-control-in-agentic-visual-generation/135643.png","ImageObject",300,407,{"name":42,"@type":43},"Emma Mercer","Person",{"url":19,"name":45,"@type":46},"DocShare","Organization","application/pdf","2026-10-21","2026-08-21",true,{"@type":52,"interactionType":53,"userInteractionCount":55},"InteractionCounter",{"@type":54},"ViewAction",13,{"@type":57,"mainEntity":58},"FAQPage",[59,65,69],{"name":60,"@type":61,"acceptedAnswer":62},"Why do current unified multimodal visual models fail at precise coordinate-based placement?","Question",{"text":63,"@type":64},"They can interpret spatial descriptions in text but cannot accurately translate numeric 2D coordinates into grounded positions on the rendered canvas, leading to mismatches in object location.","Answer",{"name":66,"@type":61,"acceptedAnswer":67},"How does MetaPoint represent continuous 2D coordinates?",{"text":68,"@type":64},"MetaPoint encodes a continuous 2D coordinate as a single special token (\u003Cmp>) whose positional embedding is engineered to encode the target coordinates, while the token embedding conveys positional intent.",{"name":70,"@type":61,"acceptedAnswer":71},"What control granularity does MetaPoint provide without architectural changes?",{"text":72,"@type":64},"A single MetaPoint token controls an object’s point position at pixel level; two tokens can specify a bounding box, enabling precise editing and generation without adding new components or special masking.","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},135643,1787316522,{"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":128,"doc_module":4,"doc_module_name":25,"category_name":129,"show_sort_weight":97,"slug":130},19,"General","general",{"code":4,"msg":82,"data":132},{"doc_id":79,"user_id":133,"nickname":42,"user_avatar":134,"doc_module":4,"category_id":112,"category_name":29,"doc_title":10,"doc_description":12,"doc_content":135,"file_id":136,"file_url":137,"file_type":138,"file_size":139,"view_count":55,"is_deleted":4,"is_public":22,"is_downloadable":22,"audit_status":22,"page_count":140,"language":141,"language_code":8,"site_id":7,"html_lang":8,"table_of_contents":142,"faqs":143,"seo_title":144,"seo_description":12,"update_tm":80,"read_time":99},962084925502,"https://ap-avatar.wpscdn.com/davatar_6f874abed73319feea01a86fa6f0fab8","arXiv :2606 .0503 1v 1 [ cs .CV] 3 Jun 2026  \nMetaPoint: Unlocking Precise Spatial Control in Agentic  \nVisual Generation  \nDewei Zhou 1 ,2 ,∗ , Xinyu Huang2 ,∗ , Xun Wang2 ,∗ , Ji Xie 1 ,2 , Yabo Zhang2 , Liang Li2 ,  \nKunchang Li2 , Zongxin Yang3 , Yi Yang 1 ,†  \n1 Zhejiang University, 2 ByteDance Seed, 3 Harvard University  \n∗ Equal contribution, †Corresponding author, Xun Wang is the project leader  \nAbstract  \nGenerative visual models fundamentally struggle with precise spatial control. This arises from a core disconnect: models can process textual descriptions of space but cannot directly map numerical coordinates onto the 2D image canvas (as illustrated in Fig. 2) . We introduce MetaPoint, a method that bridges this gap by representing a continuous 2D coordinate as a single, special token. Crucially, MetaPoint requires no new architectural components; it directly leverages the model’s inherent positional encoding schemes to interpret these coordinates, treating our token asa virtual point on the canvas. This lightweight approach enables pixel-level control of an object’s position with one token or its bounding box with two, all without requiring architectural changes or bespoke attention masking. The MetaPoint tokens are designed to be compositional, serving as spatial primitives. This allows a planner agent to decompose a high-level user request into a structured sequence of primitives for the generator. By providing a simple, precise, and scalable building block for spatial control, MetaPoint unlocks more powerful compositional generative agents and enables intuitive, interactive editing systems.  \nDate: June 4, 2026  \nCorrespondence: Yi Yang at [yangyics@zju.edu.cn](yangyics@zju.edu.cn)  \n1 Introduction  \nUnified Multimodal Models (UMMs) [7, 14 , 40 , 45], which unify generation and understanding across text and image modalities, have achieved notable success in precisely rendering photorealistic imagery from complex user prompts. Earlier advances in Large Language Models (LLMs) showed that purely text-based generation–understanding models can easily parse and reason over numerical coordinates in prompts [11, 44] . This naturally motivates the question of whether UMMs, with their integrated visual generation capabilities, can not only understand spatial specifications in text but also translate that understanding into precise visual layouts. In practice, however, a notable gap emerges: even for simple instructions, current UMMs often fail to place or locate objects accurately at the specified coordinates on the canvas (see Fig. 2) . This gap between textual spatial specifications and the rendered visual outcome limits their reliability in applications that demand strict spatial accuracy.  \nResearchers have worked hard to improve position control, yet existing methods still have clear limitations (see Tab. 1) . For instance, attention-masking approaches [61 , 76] provide only coarse, patch-level control.  \nMulti-Instance Generation  \n(1-6): Golden, silver, mossy, fire, icy, glassy; (7- 12): Red, blue, pink, green, yellow, black; (13- 18): Cherry, strawberry, watermelon, peach, prunes, pear; (19-24): 1 to 6; (25-30): A to F.  \n13 watermelons as û 16 oranges as X  \nSmart Edit With Reasoning  \nMake men and women equal  \n5x6 egg-like objects  \nSmart Generation With Reasoning  \nAssemble into a toy  \nGeneration with Self-Reflection  \nPrompt: On an experimental table, plaster casts of cylinders, spheres, cones, cubes, and cuboids are placed, with the cylinder being the tallest and the sphere the shortest.  \nReflect  \n{  \nproblem: Cubic plaster not generated, box: [0.0, 0.6, 1.0, 1.0],  \ninstruction: Add a cubic plaster  \n},  \n{  \nproblem: The number and height of spheres do not meet the requirements, box: [0.1, 0.5, 0.3, 0.8],  \ninstruction: Remove the sphere  \n}  \nExecute  \nFigure 1 MetaPoint demonstrates a wide array of capabilities, including complex multi-instance generation, versatile object editing (move, insert, resiz","cbCaie5jLW2M0X99","https://ap.wps.com/l/cbCaie5jLW2M0X99","pdf",5287317,24,"English","# Introduction\n## Spatial gap in unified multimodal models\n## Related work and limitations\n## MetaPoint method\n## Compositional spatial primitives\n## Experimental illustrations","[{\"question\":\"Why do current unified multimodal visual models fail at precise coordinate-based placement?\",\"answer\":\"They can interpret spatial descriptions in text but cannot accurately translate numeric 2D coordinates into grounded positions on the rendered canvas, leading to mismatches in object location.\"},{\"question\":\"How does MetaPoint represent continuous 2D coordinates?\",\"answer\":\"MetaPoint encodes a continuous 2D coordinate as a single special token (\\u003cmp\\u003e) whose positional embedding is engineered to encode the target coordinates, while the token embedding conveys positional intent.\"},{\"question\":\"What control granularity does MetaPoint provide without architectural changes?\",\"answer\":\"A single MetaPoint token controls an object’s point position at pixel level; two tokens can specify a bounding box, enabling precise editing and generation without adding new components or special masking.\"}]","MetaPoint - Bridging Continuous 2D Spatial Control in Agentic Visual Generation | PDF"]