[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-82363-en":3,"doc-seo-82363-105":29,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},82363,687197207919,"Theodora","https://ap-avatar.wpscdn.com/avatar/a000253d6f5f7c60be?x-image-process=image/resize,m_fixed,w_180,h_180&k=1779446848396160552",8,"Research & Report","SIGLIP-HD by Fine-to-Coarse Supervision","High-quality visual representation is a long-standing goal in computer vision. For multimodal LLMs, higher-resolution images can generate more fine-grained visual tokens, but this increases computation and engineering complexity due to multiple forward passes and postprocessing. The paper asks whether medium-resolution inputs can already unlock comparable perception ability. It proposes SigLIP-HD, a simple fine-to-coarse supervision method: coarse mid-resolution features are trained to mimic fine-grained features from high-resolution images. Built on SigLIP 2, the model improves visual tokens under identical inference cost, with strong gains on MLLM benchmarks, especially OCR tasks.","arXiv :2607 .09488v1 [ cs .CV] 10 Jul 2026  \nSIGLIP-HD BY FINE-TO-COARSE SUPERVISION  \nLihe Yang1 Zhen Zhao2† Hengshuang Zhao1†  \n1The University of Hong Kong 2 Shanghai AI Laboratory  \n[https://github.com/LiheYoung/SigLIP-HD](https://github.com/LiheYoung/SigLIP-HD)  \nABSTRACT  \nHigh-quality visual representation is a long-standing pursuit in computer vision.  \nIn the context of multimodal LLMs (MLLMs), feeding higher-resolution images can produce more fine-grained visual tokens. However, it introduces additional computational and design complexity, due to multiple forward passes and postprocessing of increased tokens. Before simply adopting a higher resolution, have we truly unlocked the model’s full perception capability at a standard resolution? Therefore, we study an interesting problem: how to achieve fine visual perception under lower cost without larger images. We present SigLIP-HD in this work.  \nThe core is a highly simple fine-to-coarse supervision design. We enforce the coarse feature of a mid-resolution image to mimic the fine-grained feature of its high-resolution version. We build this framework on the advanced SigLIP 2 model.  \nOur final model produces better visual tokens at exactly the same inference budget.  \nIt is validated on extensive MLLM benchmarks and consistently delivers stronger results than our baseline model, especially on OCR-related tasks.  \n1 INTRODUCTION  \nFrom supervised pre-training (Deng et al., 2009), to vision-centric self-supervised learning (Wu et al., 2018 ; He et al., 2022), to vision-language contrastive learning (Radford et al., 2021), further to hybrid training paradigms (Maninis et al., 2025), the computer vision community keeps pursuing more transferable visual representations. High-quality image embeddings (Tschannen et al., 2025 ; Siméoni et al., 2025) have fundamentally advanced the development of a wide range of perception and generation tasks (Lin et al., 2014 ; Yu et al., 2025) . In recent years, witnessing the power of LLMs (Achiam et al., 2023), finding better visual tokens for multimodal LLMs (MLLMs) (Liu et al., 2023) is receiving growing attention (Tong et al., 2024b ;a) . The quality of these tokens is critical for MLLMs to accurately perceive and reason over visual signals.  \nThere are three mainstream approaches to improving the visual representations in MLLMs. The first is to directly pre-train a better vision encoder from scratch with better algorithms (Oquab et al., 2024 ; Tschannen et al., 2025), more data (Fan et al., 2025 ; Bolya et al., 2025), or larger models (Siméoni et al., 2025) . This line of work requires tremendous resources (e.g., million GPU hours, billion data), which are unaffordable for most researchers. The second approach is leveraging multiple well-trained encoders (Tong et al., 2024b ;a; Shi et al., 2025b) . Different encoders have their own strengths. CLIPs (Radford et al., 2021) are good at modeling vision-language correspondence, while vision-only models (Oquab et al., 2024) excel at capturing detailed visual correlations. Incorporating them may amplify their distinct advantages while suppressing their drawbacks. Despite being intuitively promising, the actual gain is limited or even negative (Tong et al., 2024a) .  \nThe last approach is simply to forward higher-resolution input (Liu et al., 2024a ;b ; Li et al., 2024) . Higher-resolution images can yield more fine-grained visual tokens, making MLLMs see more clearly. This principle is gradually strengthened by state-of-the-art MLLMs (Liu et al., 2025 ; Wu et al., 2024b ; Deitke et al., 2025), from resizing to a fixed larger size (Liu et al., 2024a), to fully preserving the native resolution (Li et al., 2025a ; Bai et al., 2025), steadily contributing to better OCR capability. Our work is inspired by this observation, but with fundamentally different roadmaps. We highlight that, although increased image size improves visual perception, it brings significant extra complexity in both compute a","cbCaiutCi1Mjp9cN","https://ap.wps.com/l/cbCaiutCi1Mjp9cN","pdf",2207970,1,15,"English","en",105,"# Introduction\n## Motivation: medium resolution perception\n## Existing approaches to improving visual representations\n## Direct high-resolution forwarding and its cost\n## Human visual system analogy and remaining challenge","[{\"question\":\"What challenge does SigLIP-HD target in multimodal LLMs?\",\"answer\":\"It targets the high compute and design complexity caused by using higher-resolution images to obtain fine-grained visual tokens, questioning whether full perception can be achieved at medium resolution instead.\"},{\"question\":\"How does the proposed fine-to-coarse supervision work?\",\"answer\":\"SigLIP-HD enforces that the coarse feature extracted from a mid-resolution image matches (mimics) the fine-grained feature produced by the corresponding high-resolution version.\"},{\"question\":\"What model and budget does SigLIP-HD build upon and preserve?\",\"answer\":\"The method is built on the SigLIP 2 model and is designed to produce better visual tokens while keeping the inference budget exactly the same.\"}]",1784179927,38,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":27},"siglip-hd-by-fine-to-coarse-supervision","",{"@graph":35,"@context":85},[36,53,68],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/siglip-hd-by-fine-to-coarse-supervision/82363/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-17","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"What challenge does SigLIP-HD target in multimodal LLMs?","Question",{"text":75,"@type":76},"It targets the high compute and design complexity caused by using higher-resolution images to obtain fine-grained visual tokens, questioning whether full perception can be achieved at medium resolution instead.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does the proposed fine-to-coarse supervision work?",{"text":80,"@type":76},"SigLIP-HD enforces that the coarse feature extracted from a mid-resolution image matches (mimics) the fine-grained feature produced by the corresponding high-resolution version.",{"name":82,"@type":73,"acceptedAnswer":83},"What model and budget does SigLIP-HD build upon and preserve?",{"text":84,"@type":76},"The method is built on the SigLIP 2 model and is designed to produce better visual tokens while keeping the inference budget exactly the same.","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":45,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":45,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":45,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":45,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":45,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":45,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":45,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]