[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-84913-en":3,"doc-seo-84913-105":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},84913,1099514068035,"Ezra","https://ap-avatar.wpscdn.com/davatar_276721f389ce27ea32af1340a28f341c",8,"Research & Report","TILDE: TILt-based Distributional Erasure for Concept Unlearning","Concept unlearning in text-to-image diffusion models supports safe deployment by suppressing unwanted concepts after training, addressing privacy, copyright, trademark, and safety obligations. Effective unlearning must also preserve generation quality, diversity, and semantic coverage on benign prompts, ideally matching a retain-only model trained without the forbidden data. Existing erasure objectives leave the target post-unlearning distribution underspecified, making retention an implicit effect. TILDE formulates unlearning as distributional alignment with an energy-tilted, anchor-free conditional distribution under a forgetting constraint.","TILDE: TILt-based Distributional Erasure for Concept Unlearning  \nNaveen George∗  \nIndian Institute of Technology Hyderabad [ai23mtech12001@iith.ac.in](ai23mtech12001@iith.ac.in)  \nKonda Reddy Mopuri  \nIndian Institute of Technology Hyderabad  \nNaoki Murata  \nSony AI [naoki.murata@sony.com](naoki.murata@sony.com)  \nYuhta Takida  \nSony AI  \nYuki Mitsufuji  \nSony AI & Sony Group Corporation  \narXiv :2607 .06432v 1 [ cs .LG] 7 Jul 2026  \nAbstract  \nConcept unlearning in text-to-image diffusion models is critical for safe and practical deployment: with rising privacy concerns, copyright disputes, trademark constraints, and safety regulations, deployed systems must be able to suppress unwanted concepts after training. Existing methods often remove the target concept effectively, but practical unlearning also requires an equally fundamental property: the unlearned model should retain quality, diversity, and semantic coverage on benign generation. The gold standard is a retain-only model trained from scratch without the unwanted data. However, common erasure objectives do not specify which post-unlearning distribution should approximate this reference, leaving retention as an implicit consequence of the update rule. We propose TILDE, TILt-based Distributional Erasure, which formulates concept unlearning as a distributional alignment problem: the desired target is the minimum-deviation conditional distribution from the pretrained model under a forgetting constraint.  \nThis energy-tilted, anchor-free target suppresses concept-expressing images while preserving benign relative mass for each prompt. We instantiate this principle with residual ∇-GFlowNet training, which learns the score correction induced by the forget energy relative to the pretrained diffusion model. Across objects, artistic styles, and characters, TILDE achieves strong forgetting while improving retention and distributional fidelity over prior baselines.  \n1 Introduction  \nText-to-image diffusion models have rapidly become part of everyday creative and professional workflows, enabling high-fidelity image creation and editing across artistic styles, fictional characters, personal identities, products, and common scenes [32] . As these models become more widely used, the ability to edit their behavior after training is essential for legal compliance, privacy protection, and safety. Privacy regulations such as the GDPR recognize rights to erasure for personal data [8], while copyright, trademark, identity, and safety constraints can require suppressing specific concepts after training. Concept unlearning addresses this requirement by editing a pretrained text-to-image model so that it no longer generates images expressing an undesirable concept.  \nThe central challenge is that unlearning requires not only effective forgetting but equally effective preservation: the edited model should maintain image quality and prompt alignment on benign prompts, without degrading related or general concepts. An edited model that successfully forgets but degrades benign generation does not approximate unlearning; it solves the forgetting problem  \n∗Work done during an internship at Sony AI.  \nPreprint.  \n\n| Forget Prompt\u003Cbr>“A bakery counter\u003Cbr>in Van Gogh style.” |  | Related Prompt General Prompt Diversity\u003Cbr>“Starry night “Brown teddy “Starry Night\u003Cbr>mug on desk” bear on bed” mural on bedroom” |  |  |  | \u003Cbr>\u003Cbr>general benign\u003Cbr>related benign\u003Cbr>target concept |\n| --- | --- | --- | --- | --- | --- | --- |\n| Pretrained\u003Cbr>Direct\u003Cbr>suppression\u003Cbr>Reward-based\u003Cbr>optimization\u003Cbr>Distribution\u003Cbr>alignment\u003Cbr>(TILDE, proposed) |  |  |  |  |  |  |\n\nFigure 1: Concept unlearning failure modes and qualitative examples for Van Gogh style removal. Left: generations for forget, related, general, and diversity prompts under the pretrained model, direct suppression, reward-based optimization, and distribution alignment. Right: schematic postunlearning behavior over output regions. Direct suppres","cbCaidFmInBrpQIg","https://ap.wps.com/l/cbCaidFmInBrpQIg","pdf",2335894,1,37,"English","en",105,"# Introduction\n## Desiderata of Concept Unlearning\n## Existing Methods and Limitations\n## Proposed TILDE Approach","[{\"question\":\"Why is concept unlearning important for deployed text-to-image diffusion systems?\",\"answer\":\"Deployed systems must suppress undesirable concepts to satisfy privacy, copyright, trademark, and safety requirements after training.\"},{\"question\":\"What problem do retention failures reveal in concept unlearning?\",\"answer\":\"Strong erasure can harm semantically related benign concepts, while reward-style objectives may collapse diversity by concentrating outputs onto a narrow set.\"},{\"question\":\"How does TILDE define the goal of unlearning?\",\"answer\":\"TILDE treats concept unlearning as distributional alignment, targeting the minimum-deviation conditional distribution from the pretrained model subject to a forgetting constraint.\"}]",1784199315,93,{"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},"tilde-tilt-based-distributional-erasure-for-concept-unlearning","",{"@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/tilde-tilt-based-distributional-erasure-for-concept-unlearning/84913/",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},"Why is concept unlearning important for deployed text-to-image diffusion systems?","Question",{"text":75,"@type":76},"Deployed systems must suppress undesirable concepts to satisfy privacy, copyright, trademark, and safety requirements after training.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"What problem do retention failures reveal in concept unlearning?",{"text":80,"@type":76},"Strong erasure can harm semantically related benign concepts, while reward-style objectives may collapse diversity by concentrating outputs onto a narrow set.",{"name":82,"@type":73,"acceptedAnswer":83},"How does TILDE define the goal of unlearning?",{"text":84,"@type":76},"TILDE treats concept unlearning as distributional alignment, targeting the minimum-deviation conditional distribution from the pretrained model subject to a forgetting constraint.","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"]