Direct 3D-Aware Object Insertion via Decomposed Visual Proxies

Jingbo Gong, Yikai Wang, Yushi Lan, Yuhao Wan, Ziheng Ouyang, Rui Zhao, Ming-Ming Cheng, Qibin Hou, Chen Change Loy
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:35958-35975, 2026.

Abstract

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.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-gong26f, title = {Direct 3{D}-Aware Object Insertion via Decomposed Visual Proxies}, author = {Gong, Jingbo and Wang, Yikai and Lan, Yushi and Wan, Yuhao and Ouyang, Ziheng and Zhao, Rui and Cheng, Ming-Ming and Hou, Qibin and Loy, Chen Change}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {35958--35975}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/gong26f/gong26f.pdf}, url = {https://proceedings.mlr.press/v306/gong26f.html}, abstract = {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.} }
Endnote
%0 Conference Paper %T Direct 3D-Aware Object Insertion via Decomposed Visual Proxies %A Jingbo Gong %A Yikai Wang %A Yushi Lan %A Yuhao Wan %A Ziheng Ouyang %A Rui Zhao %A Ming-Ming Cheng %A Qibin Hou %A Chen Change Loy %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-gong26f %I PMLR %P 35958--35975 %U https://proceedings.mlr.press/v306/gong26f.html %V 306 %X 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.
APA
Gong, J., Wang, Y., Lan, Y., Wan, Y., Ouyang, Z., Zhao, R., Cheng, M., Hou, Q. & Loy, C.C.. (2026). Direct 3D-Aware Object Insertion via Decomposed Visual Proxies. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:35958-35975 Available from https://proceedings.mlr.press/v306/gong26f.html.

Related Material