We are happy to release the project page, code, and model for <strong>Direct 3D-Aware Object Insertion via Decomposed Visual Proxies</strong>.</p>\n<p>Project page: <a href=\"https://gong1130.github.io/DIRECT/\" rel=\"nofollow\">https://gong1130.github.io/DIRECT/</a><br>Code: <a href=\"https://github.com/Gong1130/DIRECT\" rel=\"nofollow\">https://github.com/Gong1130/DIRECT</a><br>Model: <a href=\"https://huggingface.co/superGong/DIRECT\">https://huggingface.co/superGong/DIRECT</a></p>\n<p>The code repository also includes an <strong>interactive demo</strong>, which allows users to manipulate a 3D proxy and perform pose-controllable object insertion. We welcome the community to try it out and explore the results!</p>\n","updatedAt":"2026-06-08T07:16:44.436Z","author":{"_id":"657b1b6dff16eeb2ee1c2ddf","avatarUrl":"/avatars/94d220bb113dd45a71c046561a0ae7da.svg","fullname":"Jingbo Gong","name":"superGong","type":"user","isPro":true,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7818847894668579},"editors":["superGong"],"editorAvatarUrls":["/avatars/94d220bb113dd45a71c046561a0ae7da.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2606.06601","authors":[{"_id":"6a266193e4c258a0294921a9","name":"Jingbo Gong","hidden":false},{"_id":"6a266193e4c258a0294921aa","name":"Yikai Wang","hidden":false},{"_id":"6a266193e4c258a0294921ab","name":"Yushi Lan","hidden":false},{"_id":"6a266193e4c258a0294921ac","name":"Yuhao Wan","hidden":false},{"_id":"6a266193e4c258a0294921ad","name":"Ziheng Ouyang","hidden":false},{"_id":"6a266193e4c258a0294921ae","name":"Rui Zhao","hidden":false},{"_id":"6a266193e4c258a0294921af","name":"Ming-Ming Cheng","hidden":false},{"_id":"6a266193e4c258a0294921b0","name":"Qibin Hou","hidden":false},{"_id":"6a266193e4c258a0294921b1","name":"Chen Change Loy","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/657b1b6dff16eeb2ee1c2ddf/nTnlheSyzaBFr_gTggVNF.mp4"],"publishedAt":"2026-06-04T00:00:00.000Z","submittedOnDailyAt":"2026-06-08T00:00:00.000Z","title":"Direct 3D-Aware Object Insertion via Decomposed Visual Proxies","submittedOnDailyBy":{"_id":"657b1b6dff16eeb2ee1c2ddf","avatarUrl":"/avatars/94d220bb113dd45a71c046561a0ae7da.svg","isPro":true,"fullname":"Jingbo Gong","user":"superGong","type":"user","name":"superGong"},"summary":"Object insertion aims to seamlessly composite a reference object into a specified region of a background image. Recent diffusion-based methods achieve high visual quality but formulate insertion as a simple 2D inpainting task, providing no explicit control over the object's 3D pose and limiting their practical applicability. We propose DIRECT (Decomposed Injection for Reference Composition and Target-integration), a novel framework that integrates interactive pose manipulation with high-fidelity 2D image synthesis to enable pose-controllable object insertion. Our method decomposes the insertion conditions into three complementary components: appearance guidance capturing visual details from the reference object, geometry guidance derived from the user-adjusted 3D proxy, and context guidance from the target background. By injecting them through separate pathways, DIRECT avoids feature entanglement and simultaneously preserves reference appearance, follows the user-specified pose, and adapts the object to the target scene. We also introduce an automated data construction pipeline to improve the diversity and quality of training data. Experiments show that DIRECT outperforms previous methods in both geometric controllability and visual quality.","upvotes":19,"discussionId":"6a266193e4c258a0294921b2","projectPage":"https://gong1130.github.io/DIRECT/","githubRepo":"https://github.com/Gong1130/DIRECT","githubRepoAddedBy":"user","ai_summary":"DIFFUSION-BASED OBJECT INSERTION FRAMEWORK WITH POSE CONTROL THROUGH DECOMPOSED GUIDANCE COMPONENTS","ai_keywords":["diffusion-based methods","object insertion","3D pose","inpainting task","interactive pose manipulation","2D image synthesis","appearance guidance","geometry guidance","context guidance","3D proxy","feature entanglement","visual quality","geometric controllability"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct","githubStars":23},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"657b1b6dff16eeb2ee1c2ddf","avatarUrl":"/avatars/94d220bb113dd45a71c046561a0ae7da.svg","isPro":true,"fullname":"Jingbo Gong","user":"superGong","type":"user"},{"_id":"673c566071f89d461772aa36","avatarUrl":"/avatars/aff6378a57272e1c34bfcd827f39211a.svg","isPro":false,"fullname":"Ouyang Ziheng","user":"ziheng1234","type":"user"},{"_id":"637de917425b0c9abbacc9d8","avatarUrl":"/avatars/8f780c58bc0d4529474a438498823cfe.svg","isPro":true,"fullname":"Ze-Xin Yin","user":"JasonYinnnn","type":"user"},{"_id":"6530c9a5f4c60acf0cc9259a","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6530c9a5f4c60acf0cc9259a/B_8QPI39LYVcVN8mVdvtz.png","isPro":false,"fullname":"Ruiqi Wu","user":"wuruiqi0722","type":"user"},{"_id":"66a79b074fbb84115c75c37f","avatarUrl":"/avatars/d15e3dd6618aa756a11cd26b78378a29.svg","isPro":false,"fullname":"kai wang","user":"jiangmu","type":"user"},{"_id":"67738fdd9cc5d33bf6bc83c5","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/ui40QeRUXC0pCzdxKMyEo.png","isPro":false,"fullname":"Yu-Yan Xu","user":"YuriXu","type":"user"},{"_id":"64749a0d5aba8edfb2eeaba7","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64749a0d5aba8edfb2eeaba7/Tiy4DEdp3KQYh7Ij8Vmkn.png","isPro":false,"fullname":"Mutian Xu","user":"Minoday","type":"user"},{"_id":"6995ca0260d08f2f9f100eb0","avatarUrl":"/avatars/f9e927066144b80ac47dac30b0aa84ca.svg","isPro":false,"fullname":"zzz","user":"rozp","type":"user"},{"_id":"673c4448c8d16e1b1a20cfe1","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/MghRdFilmMK7tdgQhdvAV.png","isPro":false,"fullname":"ye_zhang99","user":"ylzHug","type":"user"},{"_id":"640f078206c3b5ca883ea522","avatarUrl":"/avatars/e1707f12ae833a956c264008bddf8508.svg","isPro":false,"fullname":"Quan-Sheng Zeng (SII)","user":"ashun989","type":"user"},{"_id":"678a17857bf05f8e1eb75cc6","avatarUrl":"/avatars/463551aa983ccdfe4243a2ecff59befb.svg","isPro":false,"fullname":"xbwang","user":"xbwang","type":"user"},{"_id":"69b518e39e9446423f06fe5e","avatarUrl":"/avatars/5d67686bd11a67b95ee63b3b633aa911.svg","isPro":false,"fullname":"Long Chen","user":"Clong00","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2606/2606.06601.md"}">
Direct 3D-Aware Object Insertion via Decomposed Visual Proxies
Abstract
DIFFUSION-BASED OBJECT INSERTION FRAMEWORK WITH POSE CONTROL THROUGH DECOMPOSED GUIDANCE COMPONENTS
Object insertion aims to seamlessly composite a reference object into a specified region of a background image. Recent diffusion-based methods achieve high visual quality but formulate insertion as a simple 2D inpainting task, providing no explicit control over the object's 3D pose and limiting their practical applicability. We propose DIRECT (Decomposed Injection for Reference Composition and Target-integration), a novel framework that integrates interactive pose manipulation with high-fidelity 2D image synthesis to enable pose-controllable object insertion. Our method decomposes the insertion conditions into three complementary components: appearance guidance capturing visual details from the reference object, geometry guidance derived from the user-adjusted 3D proxy, and context guidance from the target background. By injecting them through separate pathways, DIRECT avoids feature entanglement and simultaneously preserves reference appearance, follows the user-specified pose, and adapts the object to the target scene. We also introduce an automated data construction pipeline to improve the diversity and quality of training data. Experiments show that DIRECT outperforms previous methods in both geometric controllability and visual quality.
Community
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2606.06601 in a model README.md to link it from this page.
Cite arxiv.org/abs/2606.06601 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2606.06601 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.