[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-83117-en":3,"doc-seo-83117-105":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},83117,1099514067438,"River Wang","https://ap-avatar.wpscdn.com/avatar/100002539ee87300030?x-image-process=image/resize,m_fixed,w_180,h_180&k=1780474512215547542",8,"Research & Report","From RGB Generation to Dense Field Readout: Pixel-Space Dense Prediction with Text-to-Image Models","Large-scale text-to-image models are leveraged as backbones for dense prediction by shifting from generative target reconstruction to direct pixel-space readout. The approach notes that RGB synthesis pretraining learns semantic and geometric priors, while dense tasks require task-native fields on the same image plane. ReChannel keeps the DiT’s RGB-aligned input distribution via the VAE encoder, drops the target-side decoder, and adds a token-local linear head with task LoRA. Evaluated on FLUX-Klein across six tasks, it improves state of the art on trimap-free matting, KITTI depth, and referring segmentation while remaining efficient.","From RGB Generation to Dense Field Readout: Pixel-Space Dense Prediction with Text-to-Image Models  \nZanyi Wang1 Xin Lin1 Haodong Li1 Dengyang Jiang2 Yijiang Li1  \n1UCSD 2 HKUST  \n[https://github. com/xmz111/ReChannel](https://github. com/xmz111/ReChannel)  \narXiv :2607 .06553v2 [ cs .CV] 9 Jul 2026  \nFigure 1 . ReChannel: readout, not generation. A pretrained DiT organizes RGB inputs into a patch-aligned spatial token field, so dense prediction becomes reading out task-native quantities on the same image plane rather than reconstructing an RGB-style target. Each output patch is a spatial carrier in the DiT lattice; the readout reinterprets its channels from RGB appearance to task-native fields—depth, surface normals, matting, referring segmentation, pose, and saliency. Since these targets are evaluated as pixel-space fields, target-side VAE decoding is unnecessary.  \nAbstract  \nLarge-scale text-to-image models are attractive backbones for dense prediction because RGB generation pretraining learns rich semantic, structural, and geometric priors. Existing generative and editing approaches reuse these priors by casting dense prediction as target generation: annotations such as depth, normals, alpha mattes, masks, and heatmaps are encoded into an RGB-trained VAE latent space and decoded back as image-like targets. We argue this inherits more of the generative output interface than dense prediction requires: unlike RGB synthesis, dense prediction asks for pixel-correct, task-native fields on the same image plane, not new RGB content to be rendered. Our key observation is that a pretrained DiT already organizes RGB inputs through a patch→token→patch lattice on the image plane, so each token indexes a fixed output patch whose channels can carry task-native quantities instead of RGB appearance. We instantiate this as ReChannel: we keep the VAE encoder for the DiT’s input distribution but  \ndrop the target-side decoder, adapt the frozen DiT with task LoRA, and map each token to its 􀀾 × 􀀾 × 􀀠􀁂 pixel-space patch through a shared token-local linear head—about 33K parameters, no spatial mixing. Using FLUX-Klein, we evaluate on six dense prediction tasks and over a dozen benchmarks. This minimal interface sets new state-of-the-art on trimap-free matting, KITTI depth, and referring segmentation, and stays competitive on normals, saliency, and pose. In a matched 4B setting it is more accurate and 2.48× faster than an edit-plus-latent-decode counterpart—dense perception can benefit from generative pretraining without inheriting its output interface.  \n1. Introduction  \nRGB-based dense prediction begins with a simple constraint: the RGB image is the only visual observation from which pixel-aligned fields such as geometry, masks, mattes, and heatmaps are estimated. The dom-  \nFigure 2 . Participation ratio (PR) of the token field: the RGB input field (32.8) is high-dimensional, while task-adapted fields collapse to compact subspaces (1.1–4.2), motivating a token-local linear readout rather than a generative decoder. Per-task PR values are illustrative; the linear head’s sufficiency is shown empirically in Sec. 4.3.  \ninant discriminative paradigm addresses this problem by pairing pretrained visual encoders with taskspecific decoders, increasingly benefiting from foundation representations such as DINO, DINOv2, and SAM [3, 13, 26] . These models provide strong RGBconditioned visual features, but the dense target is still obtained through learned task decoders that translate the representation into pixel-space fields [5, 30, 44, 46] . Text-to-image models offer a different route at larger generative scale [2, 32, 36, 41] . Their RGB synthesis objective learns rich image-formation priors over appearance, semantics, structure, and layout, which recent generative dense predictors have reused across depth, surface normals, segmentation, and matting [8, 10, 11, 39] . These methods have rapidly evolved from iterative diffusion to deterministic one-step","cbCaipQbGKLI5ONk","https://ap.wps.com/l/cbCaipQbGKLI5ONk","pdf",1550964,1,10,"English","en",105,"# Introduction\n## Patch-to-token-to-patch interface for readout\n## Why target-side RGB decoding is unnecessary\n# Method: ReChannel\n## Token-local readout and task LoRA adaptation\n# Experiments\n## Evaluation on dense prediction benchmarks\n# Results and analysis\n## Participation ratio motivation","[{\"question\":\"What problem does ReChannel address in using text-to-image models for dense prediction?\",\"answer\":\"It addresses the mismatch between generative methods that reconstruct RGB-like or latent targets and dense prediction’s need for pixel-correct, task-native fields on the same image plane.\"},{\"question\":\"How does ReChannel differ from prior VAE-based generative or editing approaches?\",\"answer\":\"ReChannel keeps the VAE encoder for the DiT input distribution but drops the target-side decoder, using a token-local linear head to map tokens to pixel-space patches carrying task-native quantities.\"},{\"question\":\"Which dense prediction tasks show strong results for ReChannel?\",\"answer\":\"ReChannel sets new state of the art on trimap-free matting, KITTI depth, and referring segmentation, and remains competitive on normals, saliency, and pose.\"}]",1784185378,25,{"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},"from-rgb-generation-to-dense-field-readout-pixel-space-dense-prediction-with-text-to-image-models","",{"@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/from-rgb-generation-to-dense-field-readout-pixel-space-dense-prediction-with-text-to-image-models/83117/",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 ReChannel address in using text-to-image models for dense prediction?","Question",{"text":75,"@type":76},"It addresses the mismatch between generative methods that reconstruct RGB-like or latent targets and dense prediction’s need for pixel-correct, task-native fields on the same image plane.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does ReChannel differ from prior VAE-based generative or editing approaches?",{"text":80,"@type":76},"ReChannel keeps the VAE encoder for the DiT input distribution but drops the target-side decoder, using a token-local linear head to map tokens to pixel-space patches carrying task-native quantities.",{"name":82,"@type":73,"acceptedAnswer":83},"Which dense prediction tasks show strong results for ReChannel?",{"text":84,"@type":76},"ReChannel sets new state of the art on trimap-free matting, KITTI depth, and referring segmentation, and remains competitive on normals, saliency, and pose.","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"]