[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-82229-en":3,"doc-seo-82229-105":29,"detail-sidebar-cat-0-en-105":90},{"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":4,"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},82229,1374391974468,"Eden","https://ap-avatar.wpscdn.com/davatar_29158cc5080c5b710cf443261637dec0",8,"Research & Report","What Pixels Are Enough? SEAMS: Sufficiency Saliency via MSE-Preservation Soft-Masks","Saliency maps identify image regions that preserve a model’s behaviour, enabling explanations that can be validated as meaningful visual evidence. SEAMS introduces a sufficiency-based saliency method that optimises a learnable soft mask with a preservation objective over a frozen differentiable model output, such as class probability, CLS embedding, or token representation. The framework uses a compact optimisation pipeline with a learnable budget and a three-way image composite, requiring no distractor dataset, architecture-specific attribution, or differentiable top-k relaxation. Experiments show compact, stable, competitive masks across ViT and ConvNeXt targets and reveal architecture-dependent sufficient evidence.","What Pixels Are Enough?  \nSEAMS: Sufficiency Saliency via MSE-Preservation Soft-Masks  \nMagdalena Tr˛edowicz  \nJagiellonian University [magdalena.tredowicz@doctoral.uj.edu.pl](magdalena.tredowicz@doctoral.uj.edu.pl)  \nŁukasz Struski Jagiellonian University  \nArkadiusz Lewicki University of Information Technology  \nand Management in Rzeszów  \nKarolina Pachota Jagiellonian University Medical College  \nAndrzej Grudzie´n Jagiellonian University Medical College  \narXiv :2607 .09164v1 [ cs .CV] 10 Jul 2026  \nMateusz Jagła Jagiellonian University Medical College  \nJacek Tabor Jagiellonian University  \nAbstract  \nSaliency maps are most useful when they identify the image regions that are sufficient to preserve a model’s behaviour. We introduce SEAMS, a sufficiency-based saliency method that directly optimises a soft mask using a preservation objective. Given a frozen differentiable model output, such as a class probability, CLS embedding, or token representation, SEAMS searches for a compact mask that preserves the selected output. The approach relies on a simple optimisation framework based on soft masks, a learnable budget, anda three-way image composite generated entirely from the query image. As a result, it requires no auxiliary distractor dataset, architecture-specific attribution mechanism, or differentiable top-k relaxation. Experiments with frozen ViT-S/16 and ConvNeXt models show that the same optimisation pipeline can generate objectlevel, class-conditioned, and token-level explanations by changing only the preserved target. The resulting masks are compact, interpretable, stable across random initialisations, and competitive on insertion and deletion benchmarks. Our results also indicate that different architectures often rely on different sufficient evidence while achieving similar preservation fidelity, highlighting the architecture-dependent nature of visual explanations.  \n1. Introduction  \nUnderstanding which image regions are responsible fora neural network prediction remains a central problem in  \nDINOv2 ViT-S/ 14 @ 5 18 ViT-S/16 @ 224  \nInput Mask Mask on Grey  \nFigure 1 . Sufficiency-based saliency across scales and architectures. SEAMS identifies compact image regions that are sufficient to preserve a selected model representation. Top: ImageNet image analysed with a supervised ViT-S/16 encoder. Bottom: DIV2K image analysed with a self-supervised DINOv2 ViT-S/14 encoder. From left to right: original image, optimised soft mask, and the corresponding saliency overlay. Despite differences in image resolution, training paradigm, and architecture, the same optimisation framework consistently identifies a small subset of pixels sufficient to preserve the target representation.  \nexplainable artificial intelligence. In computer vision, explanations should be spatially precise enough to distinguish relevant object parts from background content, allowing researchers to verify whether a model relies on meaningful visual evidence. This challenge has become particularly important with the growing adoption of deep  \nInput Class Probability CLS Embedding Patch-Token  \nViolin Lycaenid Ant  \nFigure 2 . One optimisation pipeline, multiple explanatory targets. Each row corresponds to a single ImageNet image, while each column preserves a different target output: a class probability, the CLS embedding, or the full patch-token representation. The optimisation procedure, regularisation, and image composite remain unchanged. Different targets produce different saliency patterns, showing that distinct components of a visual representation rely on different subsets of image pixels.  \nconvolutional networks and Vision Transformers, whose internal representations are difficult to interpret directly.  \nMost existing visual explanation methods are based on gradient attribution. Representative examples include vanilla saliency maps [19], SmoothGrad [20], DeepLIFT [18], Integrated Gradients [21], and Layer-wise Relevance Propagation [1] . More","cbCaipz9LnabO1DL","https://ap.wps.com/l/cbCaipz9LnabO1DL","pdf",24163378,1,22,"English","en",105,"# Abstract\n# Introduction\n## Motivation: sensitivity vs sufficiency\n## SEAMS overview and key idea\n## Visualisation and targets (class, CLS, token)","[{\"question\":\"What problem does SEAMS address in visual explanation methods?\",\"answer\":\"SEAMS targets the mismatch between sensitivity-based attribution and the question of which pixels are sufficient to preserve a model representation or prediction.\"},{\"question\":\"How does SEAMS generate explanations?\",\"answer\":\"SEAMS optimises a continuous soft mask using a preservation objective so that the selected frozen model output remains unchanged while only a compact region of the image is retained.\"},{\"question\":\"What are the main practical advantages claimed for SEAMS?\",\"answer\":\"SEAMS is fully post hoc and model agnostic, requiring no auxiliary distractor dataset, no architecture-specific attribution mechanism, and no differentiable top-k relaxation.\"}]",1784178991,55,{"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":85,"head_meta":87,"extra_data":89,"updated_unix":27},"what-pixels-are-enough-seams-sufficiency-saliency-via-mse-preservation-soft-masks","",{"@graph":35,"@context":84},[36,53,67],{"@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/what-pixels-are-enough-seams-sufficiency-saliency-via-mse-preservation-soft-masks/82229/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":61,"encodingFormat":60,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-16",true,{"@type":64,"interactionType":65,"userInteractionCount":4},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70,76,80],{"name":71,"@type":72,"acceptedAnswer":73},"What problem does SEAMS address in visual explanation methods?","Question",{"text":74,"@type":75},"SEAMS targets the mismatch between sensitivity-based attribution and the question of which pixels are sufficient to preserve a model representation or prediction.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"How does SEAMS generate explanations?",{"text":79,"@type":75},"SEAMS optimises a continuous soft mask using a preservation objective so that the selected frozen model output remains unchanged while only a compact region of the image is retained.",{"name":81,"@type":72,"acceptedAnswer":82},"What are the main practical advantages claimed for SEAMS?",{"text":83,"@type":75},"SEAMS is fully post hoc and model agnostic, requiring no auxiliary distractor dataset, no architecture-specific attribution mechanism, and no differentiable top-k relaxation.","https://schema.org",{"og:url":51,"og:type":86,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":88,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":91},[92,96,100,104,109,114,119,122,127,130,134],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":93,"show_sort_weight":94,"slug":95},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":97,"show_sort_weight":98,"slug":99},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":101,"show_sort_weight":102,"slug":103},"Exam",70,"exam",{"id":105,"doc_module":4,"doc_module_name":45,"category_name":106,"show_sort_weight":107,"slug":108},5,"Comic",60,"comic",{"id":110,"doc_module":4,"doc_module_name":45,"category_name":111,"show_sort_weight":112,"slug":113},6,"Technology",50,"technology",{"id":115,"doc_module":4,"doc_module_name":45,"category_name":116,"show_sort_weight":117,"slug":118},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":120,"slug":121},30,"research-report",{"id":123,"doc_module":4,"doc_module_name":45,"category_name":124,"show_sort_weight":125,"slug":126},9,"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":45,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":131,"doc_module":4,"doc_module_name":45,"category_name":132,"show_sort_weight":131,"slug":133},10,"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":45,"category_name":136,"show_sort_weight":105,"slug":137},19,"General","general"]