[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-86195-en":3,"doc-seo-86195-105":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},86195,13056703019662,"Evangeline","https://ap-avatar.wpscdn.com/avatar/be000253a8e92610077?_k=1778726343310543188",8,"Research & Report","The Devil Is in the Leakage: A Disentangled Dual-Purification Framework for High-Fidelity Hairstyle Transfer","Hairstyle transfer synthesizes photorealistic portraits by transplanting a reference hairstyle onto a source subject while preserving the source’s identity. Large foundation models still fail at the zero-shot disentanglement required for precise local editing, because reference hair becomes entangled with facial identity and pose. A common pipeline uses a “bald” source image plus identity-agnostic hairstyle features, yet it produces identity inconsistency and incorrect hair geometry. The work identifies a twofold leakage problem and proposes DPF to mitigate it.","The Devil Is in the Leakage: A Disentangled Dual-Purification Framework for High-Fidelity Hairstyle Transfer  \nJijie Li  \n[lijijie2024@ia.ac.cn](lijijie2024@ia.ac.cn)[ ](lijijie2024@ia.ac.cn)SAI, UCAS; MAIS, CASIA Beijing, China  \nJiankuo Zhao  \n[zhaojiankuo2024@ia.ac.cn](zhaojiankuo2024@ia.ac.cn)[ ](zhaojiankuo2024@ia.ac.cn)MAIS, CASIA; SAI, UCAS Beijing, China  \nXiangyu Zhu∗ [xiangyu.zhu@ia.ac.cn](xiangyu.zhu@ia.ac.cn)[ ](xiangyu.zhu@ia.ac.cn)SAI, UCAS; MAIS, CASIA Beijing, China  \nZhen Lei  \n[zhen.lei@ia.ac.cn](zhen.lei@ia.ac.cn)[ ](zhen.lei@ia.ac.cn)MAIS, CASIA; SAI, UCAS Beijing, China CAIR, HKSIS, CAS Hong Kong, China SCSE, FIE, M.U.S.T Macau, China  \narXiv :2607 . 11281v1 [ cs .CV] 13 Jul 2026  \nSource Reference Result Source Reference Result  \nFigure 1: High-fidelity hairstyle transfer by our proposed Dual-Purification Framework (DPF). Our method successfully transfers complex hairstyles across subjects with significant variations in pose, gender, and appearance. DPF preserves the source identity with high fidelity and generates geometrically aligned, photorealistic results.  \nAbstract  \nHairstyle transfer aims to synthesize a photorealistic portrait by transplanting the hairstyle from a reference image onto a source subject, while preserving the source’s identity. While recent largescale foundation models exhibit remarkable generative capabilities, they struggle with the zero-shot disentanglement required for such precise local editing, inherently entangling the reference hairstyle with its original facial identity and pose. To address these limitations through structural decomposition, a standard pipeline for hairstyle transfer typically decouples the process by first generating a “bald”image from the source and extracting identity-agnostic hairstyle features from a reference, fusing them to produce the final result. However, this methodology is frequently prone to several types of artifacts, including identity inconsistency and mismatched hair geometry. In this paper, we demonstrate that these artifacts stem from a more fundamental issue, which we term the leakage problem. This leakage is twofold: First, Identity Leakage in Hairstyle occurs when hairstyle features remain entangled with the reference’s identity  \n∗ Corresponding author.  \nand pose. Second, Flaw Leakage in Bald arises when subtle geometric flaws left in the “bald” image are propagated into the synthesized hairstyle. To address these issues, we propose the Dual-Purification Framework (DPF), which integrates two complementary purification strategies. The Adversarial Hairstyle Purification (AHP) module explicitly purges identity information from hairstyle features by adversarially minimizing hairstyle–bald mutual information. Concurrently, the Contrastive Geometric Purification (CGP) module introduces a contrastive objective that penalizes the model’s reliance on these geometric flaws in the “bald” image, thereby suppressing the Flaw Leakage in Bald. By explicitly mitigating both components of leakage, DPF achieves state-of-the-art performance in high-fidelity, identity-preserving hairstyle transfer.  \nKeywords  \nHair Transfer, Diffusion Models, Hairstyle Editing, Latent Disentanglement  \n1 Introduction  \nIn real-world scenarios such as hair salons and wig stores, customers often wish to preview how they would look with a new  \n(a) Identity Leakage in Hairstyle  \nIdentity inconsistency  \n(b) Flaw Leakage in Bald  \nSource Bald Reference Failure Case  \n| Flaws |  pollute hair representation\u003Cbr> |\n| --- | --- |\n\nFigure 2: Illustration of the two factors of leakage in hairstyle transfer. (a): Identity Leakage in Hairstyle, where reference identity features distort the source person. (b): Flaw Leakage in Bald, where flaws in the “bald” image pollute the hair representation and misleads the generative model to produce a failed hair contour.  \nhairstyle by indicating a reference photo of a desired model. This motivates the task of hairstyle transfer, which aims","cbCaichCfrlqhaX1","https://ap.wps.com/l/cbCaichCfrlqhaX1","pdf",4570637,1,10,"English","en",105,"# Abstract\n# Keywords\n# 1 Introduction","[{\"question\":\"What is the main challenge in hairstyle transfer discussed in the paper?\",\"answer\":\"The paper emphasizes that achieving zero-shot disentanglement for precise local editing is difficult: generated hair features tend to entangle with the reference facial identity, pose, or lighting, preventing explicit control and leading to artifacts.\"},{\"question\":\"What does the paper mean by the “leakage problem”?\",\"answer\":\"The leakage problem is twofold: Identity Leakage in Hairstyle keeps hairstyle features entangled with the reference’s identity and pose, and Flaw Leakage in Bald propagates subtle geometric flaws from the generated bald image into the synthesized hair representation.\"},{\"question\":\"How does the Dual-Purification Framework (DPF) address leakage?\",\"answer\":\"DPF combines Adversarial Hairstyle Purification (AHP) to adversarially minimize mutual information between hairstyle and bald to remove identity content, and Contrastive Geometric Purification (CGP) to penalize reliance on bald-image geometric flaws, suppressing both leakage components for higher fidelity results.\"}]",1784209297,25,{"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},"the-devil-is-in-the-leakage-a-disentangled-dual-purification-framework-for-high-fidelity-hairstyle-transfer","",{"@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/the-devil-is-in-the-leakage-a-disentangled-dual-purification-framework-for-high-fidelity-hairstyle-transfer/86195/",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 is the main challenge in hairstyle transfer discussed in the paper?","Question",{"text":75,"@type":76},"The paper emphasizes that achieving zero-shot disentanglement for precise local editing is difficult: generated hair features tend to entangle with the reference facial identity, pose, or lighting, preventing explicit control and leading to artifacts.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"What does the paper mean by the “leakage problem”?",{"text":80,"@type":76},"The leakage problem is twofold: Identity Leakage in Hairstyle keeps hairstyle features entangled with the reference’s identity and pose, and Flaw Leakage in Bald propagates subtle geometric flaws from the generated bald image into the synthesized hair representation.",{"name":82,"@type":73,"acceptedAnswer":83},"How does the Dual-Purification Framework (DPF) address leakage?",{"text":84,"@type":76},"DPF combines Adversarial Hairstyle Purification (AHP) to adversarially minimize mutual information between hairstyle and bald to remove identity content, and Contrastive Geometric Purification (CGP) to penalize reliance on bald-image geometric flaws, suppressing both leakage components for higher fidelity results.","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},[93,97,101,105,110,115,120,123,128,131,134],{"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":21,"doc_module":4,"doc_module_name":45,"category_name":132,"show_sort_weight":21,"slug":133},"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":45,"category_name":136,"show_sort_weight":106,"slug":137},19,"General","general"]