[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-157808-105":59,"doc-detail-157808-en":130},{"code":4,"msg":5,"data":6},0,"success",[7,13,18,23,28,33,38,43,48,51,55],{"id":8,"doc_module":4,"doc_module_name":9,"category_name":10,"show_sort_weight":11,"slug":12},1,"Document","Story & Novel",90,"story-novel",{"id":14,"doc_module":4,"doc_module_name":9,"category_name":15,"show_sort_weight":16,"slug":17},2,"Literature",80,"literature",{"id":19,"doc_module":4,"doc_module_name":9,"category_name":20,"show_sort_weight":21,"slug":22},4,"Exam",70,"exam",{"id":24,"doc_module":4,"doc_module_name":9,"category_name":25,"show_sort_weight":26,"slug":27},5,"Comic",60,"comic",{"id":29,"doc_module":4,"doc_module_name":9,"category_name":30,"show_sort_weight":31,"slug":32},6,"Technology",50,"technology",{"id":34,"doc_module":4,"doc_module_name":9,"category_name":35,"show_sort_weight":36,"slug":37},7,"Healthcare",40,"healthcare",{"id":39,"doc_module":4,"doc_module_name":9,"category_name":40,"show_sort_weight":41,"slug":42},8,"Research & Report",30,"research-report",{"id":44,"doc_module":4,"doc_module_name":9,"category_name":45,"show_sort_weight":46,"slug":47},9,"Religion & Spirituality",20,"religion-spirituality",{"id":46,"doc_module":4,"doc_module_name":9,"category_name":49,"show_sort_weight":46,"slug":50},"World Cup","world-cup",{"id":52,"doc_module":4,"doc_module_name":9,"category_name":53,"show_sort_weight":52,"slug":54},10,"Lifestyle","lifestyle",{"id":56,"doc_module":4,"doc_module_name":9,"category_name":57,"show_sort_weight":24,"slug":58},19,"General","general",{"code":4,"msg":60,"data":61},"ok",{"site_id":62,"language":63,"slug":64,"title":65,"keywords":66,"description":67,"schema_data":68,"social_meta":123,"head_meta":125,"extra_data":127,"updated_unix":129},105,"en","hierarchical-diffusion-autoencoders-and-disentangled-image-manipulation-wacv-2024-paper","Hierarchical Diffusion Autoencoders and Disentangled Image Manipulation - WACV 2024 paper","","Diffusion models achieve high visual fidelity, yet their latent spaces remain difficult to probe and manipulate. Existing diffusion autoencoders encode semantics with a single latent code, overlooking low- and mid-level details and producing entangled representations. This paper introduces Hierarchical Diffusion Autoencoders (HDAE) that leverages a coarse-to-fine feature hierarchy to build richer latent representations and enable semantic-level disentanglement. HDAE converges faster and supports truncated-feature manipulation for more controllable attribute editing. Experiments cover reconstruction, style mixing, interpolation, editing, and multi-modal semantic synthesis, with code planned for release.",{"@graph":69,"@context":122},[70,84,105],{"@type":71,"itemListElement":72},"BreadcrumbList",[73,77,79,82],{"item":74,"name":75,"@type":76,"position":8},"https://docshare.wps.com","Home","ListItem",{"item":78,"name":9,"@type":76,"position":14},"https://docshare.wps.com/document/",{"item":80,"name":40,"@type":76,"position":81},"https://docshare.wps.com/document/research-report/",3,{"item":83,"name":65,"@type":76,"position":19},"https://docshare.wps.com/document/hierarchical-diffusion-autoencoders-and-disentangled-image-manipulation-wacv-2024-paper/157808/",{"url":83,"name":65,"@type":85,"image":86,"author":91,"headline":65,"publisher":94,"fileFormat":97,"inLanguage":63,"description":67,"dateModified":98,"datePublished":99,"encodingFormat":97,"isAccessibleForFree":100,"interactionStatistic":101},"DigitalDocument",{"url":87,"@type":88,"width":89,"height":90},"https://docshare.wps.com/thumbnails/hierarchical-diffusion-autoencoders-and-disentangled-image-manipulation-wacv-2024-paper/157808.png","ImageObject",300,407,{"name":92,"@type":93},"Terk","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-10-17","2026-08-29",true,{"@type":102,"interactionType":103,"userInteractionCount":39},"InteractionCounter",{"@type":104},"ViewAction",{"@type":106,"mainEntity":107},"FAQPage",[108,114,118],{"name":109,"@type":110,"acceptedAnswer":111},"Why do prior diffusion autoencoders struggle with controllable image editing?","Question",{"text":112,"@type":113},"They encode semantic information using a single latent code derived from the last semantic-layer features, which neglects low- and mid-level details and mixes different semantic levels, making attribute-specific directions hard to isolate.","Answer",{"name":115,"@type":110,"acceptedAnswer":116},"What is the core idea behind Hierarchical Diffusion Autoencoders (HDAE)?",{"text":117,"@type":113},"HDAE builds a hierarchical latent space using a coarse-to-fine feature hierarchy from the semantic encoder, mapping different feature scales to corresponding latent inputs for the diffusion-based decoder.",{"name":119,"@type":110,"acceptedAnswer":120},"How does truncated-feature manipulation improve disentanglement?",{"text":121,"@type":113},"The method manipulates using truncated features, motivated by the observation that most low-value feature channels are critical contributors to entanglement, enabling cleaner edits such as separating “old” from “wearing eyeglasses”.","https://schema.org",{"og:url":83,"og:type":124,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":126,"canonical":83},"index,follow",{"doc_id":128,"site_id":62},157808,1787989915,{"code":4,"msg":5,"data":131},{"doc_id":128,"user_id":132,"nickname":92,"user_avatar":133,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":134,"file_id":135,"file_url":136,"file_type":137,"file_size":138,"view_count":39,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":52,"language":139,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":140,"faqs":141,"seo_title":142,"seo_description":67,"update_tm":129,"read_time":143,"preview_page_count":4},1099525198933,"https://ap-avatar.wpscdn.com/davatar_155a257f0dc6eb9ab79c44ca47cae57d","This WACV paper is the Open Access version, provided by the Computer Vision Foundation.  \nExcept for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.  \nHierarchical Diffusion Autoencoders and Disentangled Image Manipulation  \nZeyu Lu 1 ,2 Chengyue Wu3 Xinyuan Chen2 Yaohui Wang2 Lei Bai2 Yu Qiao2 Xihui Liu3  \n1 Shanghai Jiao Tong University 2 Shanghai Artificial Intelligence Laboratory  \n3 The University of Hong Kong  \nReal Image DAE HDAE (ours)  \n(a) Image Reconstruction  \n(b) Disentangled Image Manipulation  \n+   \nReal Image B Coarse A + Fine B  \n(c) Style Mixing  \nSemantic Real Image B Layout A +  \nLayout A Detail B  \n(d) Multi-modal Semantic Image Synthesis  \nFigure 1 . Applications of Hierarchical Diffusion Autoencoders (HDAE). (a) Near-perfect image reconstruction. (b) Disentangled image manipulation. Our approach disentangles “arched eyebrows” with other related attributes such as “female” and “eyeglasses”. (c) Style mixing with different levels of features from different images. (d) Multi-modal semantic image synthesis with a layout image and a real  \nimage providing information on style and details.  \nAbstract  \nDiffusion models have attained impressive visual quality for image synthesis. However, how to probe and manipulate the latent space of diffusion models has not been extensively explored. Prior work diffusion autoencoders encode the semantic representations with a single latent code, neglecting the low-level details and leading to entangled representations. To mitigate those limitations, we propose Hierarchical Diffusion Autoencoders (HDAE) that exploits the coarse-tofine feature hierarchy for the latent space of diffusion models. Our HDAE converges 2+ times faster and encodes richer and more comprehensive coarse-to-fine representations of images. The hierarchical latent space inherently disentangles different semantic levels of features. Furthermore, we propose a truncated feature based approach for disentangled image manipulation. We demonstrate the effectiveness of our proposed HDAE with extensive experiments and applications on image reconstruction, style mixing, controllable interpolation, image editing, and multi-modal semantic image synthesis. The code will be released upon acceptance.  \n1. Introduction  \nDiffusion models [14, 47] have demonstrated impressive image generation quality and achieved remarkable success in various applications, such as text-to-image generation [30, 35], image editing [13, 32, 41], and inpainting [26, 40] .  \nA semantically meaningful, editable, and decodable latent space is of particular importance in interpreting generative models as well as applications such as image editing. There have been various works on designing and manipulating the latent space of GANs [12, 44, 45] . However, the latent space of diffusion models has been underexplored. Preechakul et al. proposed Diffusion Autoencoders (DAE) [32], which leverages a learnable encoder to discover high-level semantic representations and leverages a diffusion model to encode the stochastic variations and decode images from the semantic latent code and the stochastic latent code.  \nHowever, the semantic latent code of diffusion autoencoders is simply represented by a single feature vector predicted by the last layer features of the semantic encoder, ignoring the rich low-level and mid-level features. Such a latent space is insufficient to encode the rich information from images. In practice, we observe that the fine-grained  \nrepresentations (e.g., background and low-level details) are omitted by the semantic encoder and only encoded by the stochastic encoder, leading to difficulties in manipulating those fine attributes. Furthermore, different levels of representations are entangled within the single holistic semantic latent code, making it difficult to find a single direction in the latent space to manipulate a specific attribute without a","cbCaifWfubHe1xaE","https://ap.wps.com/l/cbCaifWfubHe1xaE","pdf",6213677,"English","# Abstract\n# 1. Introduction\n## Diffusion latent space and limitations\n## Proposed approach: Hierarchical Diffusion Autoencoders\n## Truncated-feature disentangled manipulation\n## Summary","[{\"question\":\"Why do prior diffusion autoencoders struggle with controllable image editing?\",\"answer\":\"They encode semantic information using a single latent code derived from the last semantic-layer features, which neglects low- and mid-level details and mixes different semantic levels, making attribute-specific directions hard to isolate.\"},{\"question\":\"What is the core idea behind Hierarchical Diffusion Autoencoders (HDAE)?\",\"answer\":\"HDAE builds a hierarchical latent space using a coarse-to-fine feature hierarchy from the semantic encoder, mapping different feature scales to corresponding latent inputs for the diffusion-based decoder.\"},{\"question\":\"How does truncated-feature manipulation improve disentanglement?\",\"answer\":\"The method manipulates using truncated features, motivated by the observation that most low-value feature channels are critical contributors to entanglement, enabling cleaner edits such as separating “old” from “wearing eyeglasses”.\"}]","Hierarchical Diffusion Autoencoders and Disentangled Image Manipulation - WACV 2024 paper | PDF",25]