arxiv:2606.06601

Direct 3D-Aware Object Insertion via Decomposed Visual Proxies

Published on Jun 4

· Submitted by

Jingbo Gong on Jun 8

Upvote

Authors:

Ziheng Ouyang ,

Abstract

DIFFUSION-BASED OBJECT INSERTION FRAMEWORK WITH POSE CONTROL THROUGH DECOMPOSED GUIDANCE COMPONENTS

Generated by Qwen/Qwen2.5-Coder-32B-Instruct

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.

View arXiv page View PDF Project page GitHub 31 Add to collection

Community

superGong

Paper submitter about 19 hours ago

We are happy to release the project page, code, and model for Direct 3D-Aware Object Insertion via Decomposed Visual Proxies.

Project page: https://gong1130.github.io/DIRECT/
Code: https://github.com/Gong1130/DIRECT
Model: https://huggingface.co/superGong/DIRECT

The code repository also includes an interactive demo, 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!

noahml

about 14 hours ago

Neat paper. Most inpainting-based insertion tools feel pretty limited because they treat everything as 2D pixels, so the idea of using decomposed proxies for actual pose control makes a lot of sense. It sounds like a much more practical way to handle object composition than just painting over a background.

I'm curious how the model handles lighting consistency when you rotate the 3D proxy, since the appearance guidance is coming from a fixed reference image?

I made a podcast on it with ResearchPod, it makes it easy to get the key concepts on the go:
https://researchpod.app/episode/74b41936-b0f5-4d51-97ea-9776c8824b55