[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-82901-en":3,"doc-seo-82901-105":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},82901,8796095461564,"Liam","https://ap-avatar.wpscdn.com/davatar_155a257f0dc6eb9ab79c44ca47cae57d",8,"Research & Report","RADIANCE: Relative Adaptive Denoising with IP-Adapter for Novel Concept Enhancement","Text-to-image diffusion models progress rapidly yet often fail on rare concepts that combine unusual attribute–object pairings, producing concept omission or semantic drift where a dominant entity overpowers the intended composition. RADIANCE introduces a training-free closed-loop inference framework that improves compositional balance during denoising. It adds a Compositional Similarity Monitor with CLIP feedback, a Bidirectional Scale Controller using positive/negative IP-Adapter scales, and a Feedback Guidance Scheduler coordinated across timesteps. Extensions include multi-object prompts via Delayed Adapter Activation and Layer-wise Alternating Guidance, keeping competitive latency while boosting success rate and throughput.","arXiv :2607 .05088v 1 [ cs .CV] 6 Jul 2026  \nRADIANCE: Relative Adaptive Denoising with  \nIP-Adapter for Novel Concept Enhancement  \nZi-Xiang Ni 1, Bo-Lun Huang 1, Teng-Fang Hsiao 1, Bo-Kai Ruan 1, and  \nHong-Han Shuai 1⋆  \nNational Yang Ming Chiao Tung University, Hsinchu, Taiwan {[zaq312.ee12](zaq312.ee12) , [kevin503.ee12](kevin503.ee12) , bluedyee.ee09, [bkruan.ee11](bkruan.ee11) ,  \n[hhshuai}@nycu.edu.tw](hhshuai}@nycu.edu.tw)  \nAbstract. Text-to-image (T2I) diffusion models have achieved striking progress but still struggle to synthesize rare concepts involving unusual attribute-object pairings, often resulting in concept omission or semantic drift where a dominant entity overwhelms the generation. Tracing these failures to a lack of compositional balance during the denoising trajectory, we propose RADIANCE, a training-free framework that treats inference as a closed-loop feedback process. RADIANCE augments pretrained backbones with three modular components: (1) a Compositional Similarity Monitor (CSM) that tracks the emergence of objects and attributes in intermediate latents via CLIP-based feedback; (2) a Bidirectional Scale Controller (BSC) that applies a reactive \"restoring force\" using positive and negative IP-Adapter scales to rebalance biased trajectories; and (3) a Feedback Guidance Scheduler (FGS) that coordinates these updates across timesteps without additional training. We further extend the framework to multi-object prompts via Delayed Adapter Activation (DAA) and Layer-wise Alternating Guidance (LAG) to prevent premature concept fusion. By overlapping monitoring and denoising through pipelined execution, RADIANCE maintains competitive latency while significantly enhancing the per-sample success rate and effective throughput. Experiments on RareBench and T2I-CompBench demonstrate that RADIANCE consistently enhances compositional alignment and perceptual quality over state-of-the-art baselines.  \n1 Introduction  \nText-to-image (T2I) diffusion models have rapidly become a foundation for visual content creation, enabling users to synthesize images directly from natural language prompts [4–7,10,32,35,37] . By combining with powerful text encoders such as CLIP [29], modern systems [26,28,34] can produce diverse and photorealistic  \nsamples that follow user intent. These capabilities have unlocked applications in digital art and design, advertising, and data augmentation for downstream ⋆ Corresponding author.  \n2 Z.-X. Ni et al.  \n\"A thorny dolphin\" \"A butterfly shaped bowl\" \"A zebra striped palm tree\" \"A lion dressed as Napoleon skiing in the desert\"  \nFig. 1: Generated samples from RADIANCE across different rare concept prompts (attributes highlighted in red and objects in blue. Prior methods typically focus on either generating objects or capturing attributes, but struggle to merge both faithfully. For example, as shown in the first column, existing approaches [8, 27] either fail to generate thorns or lose the dolphin semantics. In contrast, our trainingfree approach, RADIANCE, yields coherent compositions that correctly combine rare attributes with their corresponding objects.  \nvision tasks [39] . Recent MM-DiT-based architectures e.g ., SD 3.5 [8], further improve scalability and cross-modal reasoning, narrowing the gap between model outputs and human imagination.  \nHowever, even the strongest T2I diffusion models still struggle when prompts describe rare concepts, such as objects with unusual attributes, shapes, or textures (e.g ., “a thorny dolphin” or “a zebra striped palm tree” in Fig. 1) . In practice, training data follow a long-tailed distribution: frequent attribute-object combinations are well represented, while rare concepts appear only a handful of times [16, 27] . As a result, models tend to fall back to frequent co-occurrences, which leads to attribute omission, entangled attributes, or incorrect bindings. Prior work on compositional and attribute aware generation [2, 3, 11, 13, 16, 19, 21, 22, ","cbCaiqsRpNrYh0Sd","https://ap.wps.com/l/cbCaiqsRpNrYh0Sd","pdf",18733297,1,18,"English","en",105,"# Introduction\n## Problem: Rare concept compositional failures\n## Proposed method: RADIANCE framework\n## Monitoring and adaptive control modules\n## Extensions for multi-object prompts\n## Experimental evaluation and results","[{\"question\":\"What problem does RADIANCE address in text-to-image diffusion models?\",\"answer\":\"RADIANCE targets failures when prompts describe rare concepts with unusual attribute–object pairings, which can cause attribute omission or semantic drift during denoising. It aims to maintain a balanced composition so both the rare attributes and the corresponding objects are generated faithfully.\"},{\"question\":\"How does RADIANCE work at inference time without additional training?\",\"answer\":\"RADIANCE reframes inference as a closed-loop feedback process. It uses a CLIP-based Compositional Similarity Monitor to track object/attribute emergence, a Bidirectional Scale Controller to apply a reactive restoring force via positive and negative IP-Adapter scales, and a Feedback Guidance Scheduler to coordinate updates across timesteps.\"},{\"question\":\"What extensions does RADIANCE provide for multi-object prompts?\",\"answer\":\"For multi-object prompts, RADIANCE introduces Delayed Adapter Activation to avoid premature fusion and Layer-wise Alternating Guidance to better separate and coordinate concepts across layers during generation.\"}]",1784183816,45,{"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},"radiance-relative-adaptive-denoising-with-ip-adapter-for-novel-concept-enhancement","",{"@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/radiance-relative-adaptive-denoising-with-ip-adapter-for-novel-concept-enhancement/82901/",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 problem does RADIANCE address in text-to-image diffusion models?","Question",{"text":75,"@type":76},"RADIANCE targets failures when prompts describe rare concepts with unusual attribute–object pairings, which can cause attribute omission or semantic drift during denoising. It aims to maintain a balanced composition so both the rare attributes and the corresponding objects are generated faithfully.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does RADIANCE work at inference time without additional training?",{"text":80,"@type":76},"RADIANCE reframes inference as a closed-loop feedback process. It uses a CLIP-based Compositional Similarity Monitor to track object/attribute emergence, a Bidirectional Scale Controller to apply a reactive restoring force via positive and negative IP-Adapter scales, and a Feedback Guidance Scheduler to coordinate updates across timesteps.",{"name":82,"@type":73,"acceptedAnswer":83},"What extensions does RADIANCE provide for multi-object prompts?",{"text":84,"@type":76},"For multi-object prompts, RADIANCE introduces Delayed Adapter Activation to avoid premature fusion and Layer-wise Alternating Guidance to better separate and coordinate concepts across layers during generation.","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"]